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.0007592 | The evolutionary dynamics of DENV 4 genotype I over a 60-year period | Dengue virus serotype 4 (DENV 4) has had a relatively low prevalence worldwide for decades; however, likely due to data paucity, no study has investigated the epidemiology and evolutionary dynamics of DENV 4 genotype I (DENV 4-I). This study aims to understand the diversity, epidemiology and dynamics of DENV 4-I. We collected 404 full length DENV4-1 envelope (E) gene sequences from 14 countries using two sources: Yunnan Province in China (15 strains during 2013–2016) and GenBank (489 strains up to 2018-01-11). Conducting phylogenetic and phylogeographical analyses, we estimated the virus spread, population dynamics, and selection pressures using different statistical analysis methods (substitution saturation, likelihood mapping, Bayesian coalescent inference, and maximum likelihood estimation). Our results show that during the last 60 years (1956–2016), DENV 4-I was present in mainland and maritime Southeast Asia, the Indian subcontinent, the southern provinces of China, parts of Brazil and Australia. The recent spread of DENV 4-I likely originated in the Philippines and later spread to Thailand. From Thailand, it spread to adjacent countries and eventually the Indian subcontinent. Apparently diverging around years 1957, 1963, 1976 and 1990, the different Clades (Clade I-V) were defined. The mean overall evolution rate of DENV 4-I was 9.74 (95% HPD: 8.68–10.82) × 10−4 nucleotide substitutions/site/year. The most recent common ancestor for DENV 4-I traces back to 1956. While the demographic history of DENV 4-I fluctuated, peaks appeared around 1982 and 2006. While purifying selection dominated the majority of E-gene evolution of DENV 4-I, positive selection characterized Clade III (Vietnam). DENV 4-I evolved in situ in Southeast Asia and the Indian subcontinent. Thailand and Indian acted as the main and secondary virus distribution hubs globally and regionally. Our phylogenetic analysis highlights the need for strengthened regional cooperation on surveillance and sharing of sample sequences to improve global dengue control and cross-border transmission prevention efforts.
| Dengue virus (DENV) can be classified into four serotypes, DENV 1, 2, 3 and 4. Although DENV 4 is the first dengue serotype to diverge in phylogenetic analyses of the genus Flavivirus, this serotype occurs at a low prevalence worldwide and spreads the least rapidly. Similar to other serotypes, DENV 4 can also cause severe dengue (SD) disease manifestations, such as dengue haemorrhagic fever (DHF) and dengue shock syndrome (DSS). To date, no study has investigated the epidemiology and dynamics of DENV 4 genotype I comprehensively. In this study, we seek to address this gap. Our study shows that the distribution of DENV 4-I is mainly restricted to Southeast Asia and the Indian subcontinent. The most recent spread of DENV 4-I likely originated from Southeast Asia–initially circulating in the Philippines, then Thailand and later on the Indian subcontinent. Viruses evolved in situ in Southeast Asia and the Indian subcontinent, respectively. Although DENV 4-I occasionally spread elsewhere, this genotype did not become widely established. The overall evolution rate of DENV 4-I was comparable with that of DENV 2–4. The nucleotide sequences indicates that the demographic history of DENV 4-I fluctuated with peaks apparent during parts of the 1980s and 2000s. Although a weak positive selection existed in Clade III -predominately in Vietnam, purifying selection dominated the E-gene evolution of DENV 4-I.
| Dengue is a mosquito-borne viral infectious disease. Although the geographical origin of dengue is still under some debate, the recent global expansion has been attributed to environmental changes, unprecedented population growth, uncontrolled urbanization, spread of the mosquito vectors, and host population movements [1]. Currently, dengue is endemic in more than 100 countries in much of the globe’s tropical and subtropical areas, being reported predominantly in Southeast Asia, the Americas, and the Western Pacific, and less frequently in Africa and Eastern Mediterranean WHO regions [2]. The prevalence of dengue has increased 74.7% between 2006 and 2016 [3]. While dengue infections are most often asymptomatic, a recent study has estimated a global total of 58.4 (95% CI: 24–122) million symptomatic dengue cases occur annually costing some US$8.9 (95% CI: 3.7–19.7) billion [4].
The four antigenically distant serotypes comprising dengue virus are categorized as DENV 1, -2, -3, and -4. Each serotype is classified into different genotypes based on complete E gene sequences [5]. The four serotypes were first identified at different times and locations, and diffused globally at different rates. Although clearly circulating before isolation techniques enabled the viruses’ discovery and characterization, DENV 1 was first reported in 1943 in French Polynesia and Japan, DENV 2 in 1944 in Papua New Guinea and Indonesia, DENV 3 and DENV 4 both in 1953 in the Philippines and Thailand [6]. A study mapping the global spread of DENV 1–4 over the 70-year history (1943–2013), indicated DENV 1 was reported most often, followed by DENV 2, DENV 3, and DENV 4 [7]. Although, DENV 4 was the first dengue serotype to diverge in phylogenetic analyses of the genus Flavivirus [8], it spread the slowest geographically [7]. Similar to other serotypes, DENV 4 can also cause severe dengue including dengue haemorrhagic fever (DHF). An epidemiological study in Thailand from 1973 to 1999 revealed that despite the proportionately lower prevalence of DENV 4, it was responsible for 10% of all DHF cases in Children [9].
Among DENV 4, there are four genotypes (I, II, III and Sylvatic). The dengue cases from DENV 4 genotype I (DENV 4-I) have increased in recent years. In 2013, a large DENV outbreak occurred in central Vietnam with a total of 204,661 clinical cases reported, of which 48.9% were DENV 4-I cases [10]. The same year, DENV 1-I, DENV 2-I and DENV 4-I caused a large outbreak with 20,255 cases including 84 deaths in Myanmar [11]. In Sri Lanka, the dominant virus in the 2012 epidemic was DENV 1, but DENV 4-I infections were also commonly observed [12]. While all four serotypes have been detected, a 2015 study showed that dengue in China remains primarily an imported disease with DENV 1 most frequently found in samples [13]. However, since 2013 several strains of DENV 4-I have been detected in Yunnan Province, China. Among the recorded DENV 4-I strains, both imported and autochthonous cases were found.
The dengue viruses DENV 1–4 are typically prevalent in tropical and subtropical regions globally. However, the spatial distribution of different genotypes is not uniform, e.g., some genotypes exist only in specific parts of Asia and others are more cosmopolitan. While the distinct distribution patterns of different genotypes remain enigmatic, mapping of the genotypes’ distribution can generate hypotheses on their spatial pattern and support policies on dengue prevention and control effort. In the past, efforts have been made to infer the dispersal of DENV 1–4 and to elucidate the evolution of their diffusion patterns [14–17]. However, so far no studies investigated globally the spatial distribution of the single genotype, DENV 4-I, its diversity, and its temporal evolution. This may be due to the limited number of recorded cases and the distribution of DENV 4-I worldwide.
In this study, we used available GenBank data in addition to Chinese data sources to make the first attempt to more comprehensively understand the spatial and evolutionary patterns of DENV 4-I. Leveraging full envelope gene sequences in our analysis, we sought to investigate the origin and spreading routes of this less-studied, rare but deadly virus, in order to contribute important information for future dengue prevention and control efforts around the globe.
Ethical approval for the study was obtained from the Chinese Center for Disease Control and Prevention Ethical Committee (No. 201214).
Dengue viruses detected in Yunnan Province were recovered from serum samples of suspected dengue patients visiting hospital from 2013 to 2016. The envelope (E) genes of isolates were sequenced as described previously [18]. These have been assigned GenBank accession numbers (HM893690-HM893699, MG601754, KJ470764, KJ470765, KX262920, KX262923). The sequences of Yunnan were compared with published sequences by using the nucleotide blast program in the NCBI. All the sequences of human DENV 4-I with full length E-gene (1,485 nucleotides) were downloaded with the accession number, collection date and country/region (as of January 11th, 2018). All the sequences were aligned using MAFFT [19]. We chose only one sequence to represent sequences with 100% matching identity by location and time. Recombinants detected based on the analyses of RDP3 program [20] were also excluded. Ultimately, 404 sequences of DENV 4-I obtained from 14 countries were included for analyses in this study (S1 Table).
The phylogenetic signal of the aligned DENV 4-I was evaluated by plotting the observed number of transitions and transversions against genetic distance for the n (n-1)/2 pairwise comparisons in an alignment of n taxa using DAMBE [21]. It is expected that transitions and transversions increase linearly with the genetic distance, with transitions being more frequent than transversions. In the Supplementary Materials, S1 Fig shows that no substitution saturation was detected, indicating phylogeny reconstruction was appropriate. In likelihood mapping analysis, groups of four sequences (quartets) were evaluated using the maximum likelihood approach. For each quartet, the three possible unrooted tree topologies were weighted. The likelihood weights were then plotted into a triangular surface. The fully resolved tree topologies were plotted in the three corners, which indicated the presence of a tree-like phylogenetic signal, and the unresolved quartets, indicating a star-like signal were shown in the central region of the triangle [22]. Likelihood mapping was performed using the TREE-PUZZLE program [23], by analyzing 10,000 random quartets. S1 Fig showed tree-like area accounted for 96.4%, which further suggested that the data were reliable for phylogenetic inference.
Rates of nucleotide substitution per site per year and time to The Most Recent Common Ancestor (TMRCA) were estimated using Bayesian Markov Chain Monte Carlo (MCMC) and implemented using the BEAST v1.8.2 software package [24]. The best-fit model of nucleotide substitutions was determined using Bayesian Information Criteria (BIC) as implemented in jModelTest [25]. The calibration point was the “year” that each strain was isolated. Statistical simulations were performed under strict or relaxed (uncorrelated exponential and lognormal) clock model, with the Bayesian Skyride Coalescent Tree Prior [26]. To determine the best-fit combination, we have applied Posterior-simulation Akaike Information Criterion through MCMC (AICM) [27], Bayes Factor (BF) [28], Harmonic Mean (HM) [29], and Path Sampling (PS) and Stepping-stone Sampling (SS) [30] model selection methods. The results showed that the best fitting model was the combination of uncorrelated relaxed exponential clock model and the Bayesian Skyride Coalescent model (S2 Table). Statistical uncertainties in parameter values were given by the 95% Highest Probability Density (HPD) intervals. All chains were run sufficiently long to achieve convergence (the effective sample size of continuous parameters greater than 200) after burn-in, as checked using TRACER v1.5 (http://tree.bio.ed.ac.uk/software/tracer/). The programs TreeAnnotator v1.8.2 in the BEAST v1.8.2 software package and Figtree (http://tree.bio.ed.ac.uk/software/Figtree/) were used to summarize the posterior tree distribution and to visualize the annotated Maximum Clade Credibility (MCC) tree, respectively. Based on the MCC tree, we identified five Clades using visual judgement and comparison among all the countries/regions that reported DENV 4-I. Using the definition of a minimum of three sequences of monophyletic origin, DENV 4-I were labelled with Clade I to V (the largest five) where every Clade included as many strains as possible.
The spatial diffusion of DENV 4-I was estimated using the Bayesian Markov chain Monte Carlo (MCMC) statistical framework implemented in the BEAST v1.8.2 package. The phylogeographical diffusion process was identified using the Bayesian Stochastic Variable Search Selection (BSVSS). Effective population size dynamics were estimated using the Bayesian Skyride Coalescent statistical approach.
Open source data from http://tapiquen-sig.jimdo.com (Carlos Efraín Porto Tapiquén. Orogénesis Soluciones Geográficas. Porlamar, Venezuela, 2015) were used in this study for the results shown in Figs 1 & 4 with help of ArcGIS 10.2 and Adobe Illustrator.
We used a variety of computational methods to explore the selection pressures. A Maximum Likelihood (ML) method was used to examine selection pressures [31]. In the analysis, the non-synonymous to synonymous rate ratio (ω = dN/dS) was determined codon-by-codon using various models of codon substitution. These models differ in how ω ratios are allowed to vary along the sequence. Four models of codon substitution were conducted in the study: M1a (ω < = 1; nearly neutral), M2a (ω < = 1 and ω > 1; positive selection), M7 (beta; a discrete distribution with 10 site classes to model values of ω between 0 and 1) and M8 (beta and ω > 1). M1a is nested with M2a, and M7 is nested with M8. Models that are nested are compared statistically using a Likelihood Ratio Test (LRT). Positive selection can be inferred when a group of codons with a ω ratio > 1 is identified and the likelihood of the codon substitution model in question is significantly higher (p < 0.05) than that of a nested model that does not take positive selection into account. Lastly, using Bayes Empirical Bayes (BEB) methods, posterior probabilities were calculated to identify sites under positive selection (posterior Bayesian probability (Pp) > 95%). All the analyses were performed by using CODEML from the PAML v4.9 package [32].
Evolution rate, effective population size dynamics, divergence time and selection pressure were estimated based on two different types of datasets: (i) all sequences of DENV 4-I and (ii) those from different Clades. The spatial diffusion was estimated based on all sequences of DENV 4-I. In order to minimize oversampling of Thailand and Vietnam during the spatial diffusion analysis, we down-sampled dataset for sensitive analysis. The down-sampled dataset included 50 sequences at random, from each Thailand and Vietnam and all the available sequenced strains from other countries/regions, therefore making the sample size 207 sequences.
Fig 1 shows that DENV 4-I were detected in Mainland Southeast Asia and the adjacent provinces of China, Maritime Southeast Asia, the Indian subcontinent, Brazil and Australia. Specifically, the recorded samples revealed that DEVN 4-I was mainly observed in Mainland Southeast Asia, especially Thailand and Vietnam. Collection of DENV 4-I covered a period of 60 years. The first strain of DENV 4-I was detected in 1956 in the Philippines, where it transmitted exclusively for some 20 years, according to known reporting and sequencing records. Over the two decades following 1976, most detected strains of DENV 4-I were found to be circulating in Thailand, while a few strains were discovered in other four countries. Detected in a total of 14 countries, DENV 4-I continued to diffuse to more areas around the globe between 1996 and 2016.
Fig 2 shows the evolution and spread of DENV 4-I over time. During the last 60 years, great geographical and genetic diversity has occurred. This is especially prominent during the last two decades after DENV 4-I became more prevalent as shown in the genetic record.
Fig 3 shows the Maximum Clade Credibility (MCC) tree. It indicates that the recent spread of DENV 4-I most likely originated from the Philippines with 0.98 posterior location probability. Viruses evolved in the Philippines and then spread across the sea to Thailand, Cambodia, Australia and China at different times. Thailand played the dominate role in spreading the viruses, gradually spreading virus to the Indian subcontinent, Myanmar, Cambodia, Singapore, Indonesia and China. These viruses diverged around 1957, 1963, 1976 and 1990, and shaped different Clades (Clade I to V). Since the introduction from Thailand, DENV 4-I has evolved in the Indian subcontinent (Clade IV), Myanmar (Clade II) and Cambodia (Clade I and III), respectively. Strains obtained in Vietnam correspond to Clade III, which evolved notably in situ for three decades after introduction from Cambodia. In the Indian subcontinent (Clade IV), dengue virus first apparently arrived in Sri Lanka in the 1960s from Thailand and then spread onto India in the early 1970s. India then became an epicenter for transmission and spread virus to Pakistan and back to Sri Lanka in the 2000s.
Fig 4 shows the detailed spatial diffusion of DENV 4-I as summarized from the MCC tree. The result of the down-sampled dataset showed that the phylogenetic topology and spatial spreading patterns were equivalent with those from full dataset (S2 Fig).
Fig 5A illustrates the demographic history of DENV 4-I. A fluctuation was observed over the 60 years with an approximately “M” shape. The two highest plateaus were around 1982 and 2006 with a width about 6-years, while the lowest point was around 1996. Analysis of the Clade I dataset revealed, as Fig 5B exhibits, that the effective population size increased approximately linearly from 1992 to 2004, and then decreased slowly from 2004 to 2013. The effective population size of Clade II, as Fig 5C conveys, increased slowly for the first two decades and then much more sharply (2007–2013), before a rapid recent decrease (2013–2016). Data from the Clade III in Fig 5D shows that there were two peaks (~2000 and ~2011) over nearly four decades of demographic analysis. The effective population size of Clade IV over the same period decreased slowly and then stayed constant with the inflexion lying around 1990, as Fig 5E depicts. In Fig 5F, one can see the long demographic history pattern of Clade V (1956–2013) is mirrored by Clade IV, but with the inflexion point much earlier (~1965).
Table 1 shows the evolution rate and the divergence time of the different Clades’ datasets. The overall evolution rate was 9.74 × 10−4 (95% HPD: 8.68 × 10−4–10.82 × 10−4) nucleotide substitutions/site/year and TMRCA of DENV 4-I was 1956 (95% HPD: 1955–1956) (year). The mean evolution rate of different Clades was comparable, with the smallest being 8.66 × 10−4 nucleotide substitutions/site/year in Clade V and the largest being 11.3 × 10−4 nucleotide substitutions/site/year in Clade II. TMRCA of Clade I to V was 1991, 1990, 1977, 1971 and 1962, respectively.
In the Supplementary Material, S3 Table shows the summary of positive selection analysis performed on different datasets. M1a vs M2a test and M7 vs M8 test indicated consistently that there was no positive selection across overall DENV 4-I, Clade I, II, IV and V datasets. In Clade III dataset, no positive selection was indicated by M1a vs M2a test, however, M7 vs M8 test inferred a weak positive selection (5.6% of codons with ω = 1.193) with amino acid site 95 Pp > 0.95.
In this study, using the most comprehensive and largest dataset(s) available we mapped the spatial distribution and 60-years evolution of one genotype from one of the four serotypes of dengue virus, DENV 4-I. We further separated this large dataset into different Clades and analyzed the evolutionary dynamics of Clades separately. Our study shows that the spatial distribution of DENV 4-I is mainly restricted to Southeast Asia and the Indian subcontinent. The recent spread of DENV 4-I likely originated from Southeast Asia–namely the Philippines, from where it then spread all the way to the Indian subcontinent, Australia and Latin America. DENV 4-I evolved in situ in Southeast Asia and on the Indian subcontinent. Although DENV 4-I cases occasionally were found elsewhere, this genotype didn’t undergo in situ evolution and largely failed to establish. We found that mainland Southeast Asia, specifically Thailand, was at the center of the global spread of the viruses, which in time contributed to the observed diversity.
Although Thailand’s strains dominated the records, the Philippines’ strains were at the basal location of the phylogeny. The analyzed diverge time (95% HPD: 1955–1956) is consistent with the timing of DENV 4-I virus being first detected in the Philippines. Therefore, we propose that global spread of DENV 4-I originated very likely from the Philippines. From the Philippines, DENV 4-I spread to Thailand and then on to nearby countries including Sri Lanka. On this point our findings clash with one previous study’s results indicating that DENV 4-I originated from Thailand, from where it spread to the neighboring countries of the Philippines and Sri Lanka [17].
We found that the strains of DENV 4-I—monophyletic Clade IV, detected in the Indian subcontinent including Sri Lanka, India and Pakistan, probably originated in Thailand, then evolved in situ, but as of 2016, had not yet spread outside of the Indian subcontinent. Dengue is endemic in the Indian subcontinent and four serotypes have been co-circulating, with DENV 2 and DENV 3 dominating [33]. However, the distribution and dissemination patterns of the different genotypes are not uniform. A study showed that the DENV 3 genotype III (DENV 3-III) was spread from the Indian subcontinent to East Africa first and then from there to Latin America [34]. Another study on DENV 2 showed that the cosmopolitan genotype was spread from the Indian subcontinent directly to Latin America [14]. In addition to the Indian subcontinent, dengue is also endemic in Brazil. Even in two separate regions of the world where dengue is endemic, the specific evolutionary drivers of a given serotype and genotype may differ.
DENV4-I was introduced to Brazil in 2011. Our study indicated that the strain was probably imported from mainland Southeast Asia, which was consistent with the analysis conducted in a previous study [35]. Curiously, DENV 4-I apparently disappeared quickly from Brazil after the introduction. This is different from the expansion/establishment patterns of DENV 4-II, which has become established in the Caribbean since early 1980s after being introduced from Southeast Asia [36]. Human movement is known to play a significant role in contributing to the virus distribution at large spatial scales (e.g., national, international) [37], because of the limited range of mosquitos’ flying distance. However, human movement alone cannot explain the different expansion pathways and establishment of new genotypes, for example, in the DENV 4-I dynamics on the Indian subcontinent and Brazil even after controlling for the climate suitability and mosquito abundance. These studies show that the spread of dengue virus is not merely shaped by human movement. Human movement is a necessary but not sufficient condition. Different pathways of genotypes spread may be associated with the strain virulence, human population immunity, previous exposure to other dengue viruses, in addition to the volume and timing of human population movements, the local environment, and numerous other factors.
Our study on evolution rate showed that the substitution rate of DENV 4-I is 9.74 (95% HPD: 8.68–10.82) × 10−4 nucleotide substitutions/site/year. This is comparable with the value found by Klungthong et al. [38] (10.72 (95% HPD: 8.41–13.11) × 10−4 nucleotide substitutions/site/year), but is approximately twice that of the one estimated by Twiddy et al. (5.42 × 10−4 substitutions/site/year) using an admittedly much smaller dataset [39]. The range of mean evolution rate from DENV 2–4 is found to be (8.3–10.7) × 10−4 nucleotide substitutions/site/year [16], which included Asian I and Asian-American genotype of DENV 2, genotype I-III of DENV 3, DENV 4-II. Although DENV 4-I was shown to evolve with a relatively large substitution rate comparing to some other genotypes [16], there is no significant rate differences among the different dengue serotypes. Among the five Clades of DENV 4-I, we found that the highest evolution rate was among Clade II strains (Table 1). The MCC tree (Fig 3) indicates that the strains of Clade II spread rapidly in mainland Southeast Asia. The rapid spread and the subsequent replication in a large susceptible population of hosts could account for the higher rate of evolution, when compared to the other Clades in DENV 4-I having relatively smaller pools of susceptible hosts. We found no evidence of positive selection in most of the Clades of DENV 4-I, although there was evidence for weak positive selection in Clade III. Some studies have suggested purifying selection [38], where majority of amino acid changes within infected hosts, are deleterious in the long run and are eventually removed from the population [40]. This purifying selection induces the distinct ladder-like phylogeny. Our study confirms this purifying selection process as the predominant evolutionary force acting on DENV 4-I.
Our study also showed that DENV 4-I experienced a fluctuating demographic pattern although occurring in a low prevalence compared to other serotypes. The fluctuation of effective infected population size might be shaped by human population immunity/susceptibility. DENV 4-I was first detected in the Philippines and spread to Thailand and Sri Lanka before 1980, where humans were lacking previous exposure and thus immunity. The effective infected population size decreased once immunity was built up as result of previous exposure. DENV 4-I viruses spread onto Vietnam around 1992 (95% HPD: 1986–1996) and evolved in situ, which might have induced the secondary increase. This observation agrees with the study by Villabona-Arenas et al. using the dataset from 1956 to 2008. They indicated that the effective population size was estimated to be the largest at two time periods around 1982 and 2005 [17], which is similar to our result shown in Fig 5A using a larger dataset covering a longer period.
Dengue is endemic in Vietnam. Most of the DENV 4-I strains detected in Vietnam shaped the Clade III, as the in situ evolution occurred since DENV 4-I started circulating in Vietnam in 1998. It did not dominate during dengue epidemics until 2013 when a large outbreak was occurred having 204,661 clinical cases with nearly 50% of them having DENV 4-I in central Vietnam [10]. This occurrence coincided with the demographic history of Clade III. Lacking immunity to DENV 4-I among the Vietnamese human population could be a reason to cause second infection and therefore severe dengue which resulted in being more likely reported. Amino acid 95 under positive selection could also account for this large outbreak. The E protein ectodomain can be divided into three structural domains designated domain I-III. Domain II contains the fusion loop (residue 98–111), which interacts with the host endosomal membrane, leading to virus-mediated membrane fusion and allowing the newly infecting virus to initiate the cellular replication cycle. As residue 95 is located three residues downstream of the fusion loop, it is likely to indirectly affect the process of membrane fusion. The Vietnam strains experiencing positive selection pressure did not spread outside of Vietnam based on the available data and our findings. Enhanced surveillance to these strains could be very helpful to aid in understanding and controlling this potentially devastating virus strain.
Although our study represents the most comprehensive study on the evolutionary dynamics of DENV 4-I using the largest dataset available to date, the results should be interpreted cautiously given various limitations. For example, itself, reporting of DENV, especially in the genetic record, is a source of many types of potential bias, of particular concern in locations with limited resources for virologic diagnostic and reporting capacity. While our dataset represents an opportunistic, but highly useful sample of occurrence, it should be noted that we worked with a limited subset of data, rather than a complete record of global DENV 4-I transmission dynamics. To account for sampling bias, in this study we constructed and analyzed a down-sampled dataset for sensitive analysis, finding an equivalent result to that which was obtained with the full dataset. Nevertheless, the quality of this type of study will be increased greatly with an enhanced global dengue surveillance, greater access to next generation diagnostic and sequencing tools, and improved data sharing systems. Our current study can be described as exploratory research, as this is the first time that the geographic spread and evolutionary dynamics of the DENV 4-I was mapped out and analyzed based on a large dataset. Given that other dengue research in this area is sparse and rarely conducts genotype-specific analysis, our targeted focus on just one genotype of DENV 4 makes comparison difficult. We hope that our study can stimulate further research studies in this area so that in the future researchers can compare different genotypes and understand better the similarities and differences among them. Our mapping of the genotypes’ distribution pattern may help to generate hypotheses on the specific mechanisms mediating the spread of DENV 4-I. This understanding is potentially of great utility in the generation of health policies and practices on dengue prevention and control.
In this study, we have investigated the global patterns of DENV 4-I dissemination—its spatial and temporal distribution. This is the most extensive molecular epidemiological study of DENV 4 genotype I to date to our knowledge. Our results indicate that recent spread of DENV 4-I originated in maritime Southeast Asia, probably from the Philippines, from where it spread to mainland Southeast Asia, and then on to the Indian subcontinent. Thailand acted as a distribution hub for spreading the virus regionally and globally. Within the India subcontinent, India was the distribution center for spreading the virus. We found that there is no uniform spreading pattern among genotypes. In addition, purifying selection was still the dominant acting force on E gene to shape the evolution, but weak positive selection existed in dengue viruses detected in Vietnam.
This work is a first step towards increased understanding of the underlying mechanisms governing the spread of DENV 4-I virus. Our study suggests that surveillance could be enhanced to better leverage next generation sequencing for informing dengue control practices. Regional cooperation should be strengthened to determine and communicate information on the genotype-specific spreading pathways, to explore the related underlying mechanisms, and ultimately to better coordinate dengue control efforts globally.
|
10.1371/journal.pntd.0006021 | Schistosoma japonicum transmission risk maps at present and under climate change in mainland China | The South-to-North Water Diversion (SNWD) project is designed to channel fresh water from the Yangtze River north to more industrialized parts of China. An important question is whether future climate change and dispersal via the SNWD may synergistically favor a northward expansion of species involved in hosting and transmitting schistosomiasis in China, specifically the intermediate host, Oncomelania hupensis.
In this study, climate spaces occupied by the four subspecies of O. hupensis (O. h. hupensis, O. h. robertsoni, O. h. guangxiensis and O. h. tangi) were estimated, and niche conservatism tested among each pair of subspecies. Fine-tuned Maxent (fMaxent) and ensemble models were used to anticipate potential distributions of O. hupensis under future climate change scenarios. We were largely unable to reject the null hypothesis that climatic niches are conserved among the four subspecies, so factors other than climate appear to account for the divergence of O. hupensis populations across mainland China. Both model approaches indicated increased suitability and range expansion in O. h. hupensis in the future; an eastward and northward shift in O. h. robertsioni and O. h. guangxiensis, respectively; and relative distributional stability in O. h. gangi.
The southern parts of the Central Route of SNWD will coincide with suitable areas for O. h. hupensis in 2050–2060; its suitable areas will also expand northward along the southern parts of the Eastern Route by 2080–2090. Our results call for rigorous monitoring and surveillance of schistosomiasis along the southern Central Route and Eastern Route of the SNWD in a future, warmer China.
| The South-to-North Water Diversion (SNWD) project is designed to channel fresh water from the Yangtze River north to more industrialized parts of China. An important question is whether future climate change and dispersal via the SNWD may synergistically favor northward expansion of schistosomiasis in China. Our models indicated increased suitability and range expansion in Oncomelania h. hupensis in the future; an eastward and northward shift in O. h. robertsioni and O. h. guangxiensis, respectively; and relative stability in O. h. gangi. The southern Central Route of SNWD will coincide with suitable areas for O. h. hupensis in 2050–2060; its suitable areas will also expand northward along the southern Eastern Route in 2080–2090. Our results call for rigorous monitoring and surveillance of schistosomiasis along the southern Central Route and Eastern Route of the SNWD in a future, warmer China.
| Schistosomiasis is a neglected tropical disease that is known to have affected people in China for more than 2100 years, with presently ~800,000 infected and ~65 million people at risk of infection [1]. The challenge of combatting this disease lies in the wide distribution of its snail hosts and the broad range of domestic and wild mammals that act as reservoirs for human infections [2]. Chinese schistosomiasis is caused by the digenetic blood trematode Schistosoma japonicum, a parasitic flatworm that completes its life cycle through one intermediate (i.e. the snail Oncomelania hupensis) and diverse definitive (i.e. mammals) hosts. Over the past five decades, China has made remarkable progress in reducing S. japonicum infections in humans through a combination of chemotherapy and snail control, but schistosomiasis has re-emerged in recent years owing to changes in ecological and socio-economic factors, together with termination of the World Bank Loan Project on schistosomiasis control in 2001 [3]. Given that schistosomiasis is unlikely to be eliminated, considering whether and how future climates are likely to impact its transmission becomes increasingly important.
Based on the environmental variables that associated with species’ occurrence records, ecological niche modeling (ENM) seeks to characterize environmental conditions suitable (i.e. realized niche) for a particular species and then identify where suitable environmental habitats are distributed in the space [4], it is a powerful tool in studies of effects of global climate change on the geography of disease transmission [5]. Assumptions under which ENMs work best include equilibrium between species’ distributions and their ecological requirements, and conservatism of ecological niche [4]. Among them, niche conservatism providing support for using ENMs has been widely noticed, the degree to which plants and animals retain their ancestral ecological traits and environmental distributions ('niche conservatism') is hotly debated, in part because of its relevance to the fate of modern species facing climate change [6].
ENM tools, however, are also subject to issues including the need to balance goodness-of-fit against model complexity [7], and the importance of considering uncertainty in model predictions [8]. These issues are particularly critical in studies involving transfer of models across space or time (e.g. climate change effects). Recent efforts have developed methods to reduce model complexity and characterize uncertainty, and thereby improve model transferability in forecasting climate change effects [9–12]. These steps include species-specific tuning of settings (rather than default setting) to improve model performance [9,10], evaluation using spatially independent training and testing data sets [12], and integrating multiple predictions via ensemble approaches [11,12].
Oncomelania hupensis is the sole intermediate snail host of S. japonicum in China, which thus depends entirely on this snail species for transmission [13]. However, the taxonomy of O. hupensis in mainland China has been debated in view of marked morphological variation. Liu et al. recognized 5 subspecies [14], whereas Davis et al. treated only 3 subspecies based on shell form, allozyme data, and biogeography [15]. However, Zhou et al. separated O. h. guangxiensis from O. h. hupensis based on molecular characters, and recognized 4 subspecies in mainland China [16], which was later verified by Li et al. based on internal transcribed spacer (ITS) and 16S fragments [17,18]. Here, we consider the four subspecies [14,16,18,19]; at present, O. h. hupensis and O. h. robertsoni dominate transmission of S. japonicum, as control measures have reduced O. h. guangxiensis and O. h. tangi considerably [13]. These four subspecies differ in shell size and structure, breeding environment, growth rates, population genetics, and potential for infection by S. japonicum [17].
Previous attempts to predict spatial dimensions of transmission risk of schistosomiasis have characterized transmission environments of S. japonicum [20–22] or ecological requirements of O. hupensi [23,24]; these studies were generally conducted at local geographic scales and with limited temporal coverage. Several environmental correlates of S. japonicum transmission have been identified, including distance to snail habitat and wetlands, seasonal land surface temperature, and seasonal variation of vegetation indices [21,22]. Climate conditions explain much variation in transmission of schistosomiasis, especially at regional and continental scales [25,26]. Understanding ecological dimensions and potential distribution of O. hupensis is thus crucial [20], and yet has not seen detailed analysis.
The South-to-North Water Diversion (SNWD) project is a multi-decade mega-project in China. It is the biggest inter-basin transfer scheme in the world, aiming to channel 25 × 109 m3 fresh water annually from the Yangtze River in southern China to the more arid and industrialized north via two routes (i.e. the Central Route and Eastern Route, Fig 1). In the context of climate change, in which the geographic potential of O. hupensis may change, the relationship of such changes to planned SNWD corridors remains unknown. Surveillance sites were established during 2002–2010 across mainland China (Fig 1); however, most sites were located along the Yangtze River at low elevations, focused on transmission by O. h. hupensis and O. h. robertsoni. The questions of whether future climate change and the SNWD project may synergistically favor expansion of some population of O. hupensis, and whether the existing surveillance sites are sufficient, necessitate the present study.
In this study, we used a unique dataset of O. hupensis presences from more than 5 thousand villages to explore ecological dimensions and potential distributions of O. hupensis in mainland China. The aims of this study were to (1) compare climate spaces occupied by the four subspecies of O. hupensis, to (2) test whether climate niches were conserved during the four subspecies’ divergence (i.e. climate niche conservatism evaluation), to (3) predict their potential distributions using state-of-the-art modelling techniques, to (4) investigate the potential impacts of future climate change and the SNWD project on O. hupensis. The overall purpose was to predict the S. japonicum transmission risk at present and under climate change in mainland China.
Occurrence data for subspecies of O. hupensis were assembled from Qian [27]. This national surveillance effort of schistosomiasis was carried out at the village level between the 1950s and 1980s across 12 Chinese provinces. In all, 5029 towns and villages reported presence of O. hupensis [27]. Rather than using centroids of infested counties, which reduces precision, we georeferenced individual villages using Google Maps. These points varied in terms of clumping, so we subsampled them to reduce sampling bias and spatial autocorrelation [28], as follows. First, we arranged infested provinces according to sample density (i.e. number of occurrence points divided by area of the province). The median served as the standard sampling effort, and all provinces presenting densities above that value were subsampled randomly to a lower density. In the end, we had 1996 occurrence points: 1402 O. h. hupensis, 470 O. h. robertsoni, 64 O. h. guangxiensis, and 60 O. h. tangi (S1 File).
Several approaches have been used to select environmental datasets for ecological niche modeling; the best environmental datasets would be ecological relevant to species in question [29]. At regional and continental scales, climatic factors have excellent predictive power in determining risk associated with disease transmission (e.g. schistosomiasis [25,26], West Nile virus [30]). Hence, we used subsets of the 19 bioclimatic variables developed by Hijmans et al. [31], chosen as follow. First, variables that combined temperature and precipitation (i.e. mean temperature of wettest quarter, mean temperature of driest quarter, precipitation of warmest quarter, precipitation of coldest quarter) were excluded because they display artificial discontinuities between adjacent grid cells in some areas [32]. The importance of each of the remaining 15 variables was assessed by a jackknife analysis of variable importance in Maxent ([33], see below for Maxent detail), and unimportant variables were discarded. Highly correlated variables were then removed in SDMtoolbox, a python-based GIS toolkit for spatial analysis [34]. Eight variables (S1 Table) that showed ecological relevance (regularized training gain >0.14) and low correlation with other variables (Pearson correlation <0.9) were chosen in the end. All variables were analyzed at a spatial resolution of 2.5 minute.
Climatic spaces occupied by the four subspecies were first compared along each environmental dimensions using violin plots, which combine the functions of boxplot and kernel density, providing a better indication of the shape of the data distribution. We used NicheA, a toolkit to create and visualized ecological niches in environmental spaces [35], to visualize climate niches occupied by each subspecies in reduced multiple environmental spaces: we displayed the first three principal components derived from the 8 bioclimatic layers, and plotted minimum volume ellipsoids (MVEs) around occupied conditions. We quantified niche overlap between pairs of subspecies using Schoener’s D [36]; this metric ranges from 0 (no overlap) to 1 (complete overlap), and was used to test niche identity and niche similarity between subspecies. Niche identity and similarity tests were performed to determine whether climate spaces occupied by the two subspecies were identical or exhibited significant difference, and whether these differences were caused by the environmental feature spaces [37]. Niche identity was tested by randomly allocating occurrence records within each pair 500 times, according to observed numbers of records, and comparing observed and simulated Schoener’s D estimates. In contrast, niche similarity was tested by shifting the centroid of the observed occurrence densities to a random location within the available environmental space 500 times, and comparing observed with the null distribution of simulated estimates of Schoener’s D [37]; climate variables measured at locations across the available backgrounds of subspecies were combined and projected onto the first two principal components using PCA_env package [37]. Smoothed densities of occurrences and available environments in each grid cell were calculated and compared among the four subspecies [37].
Background environments for climate niche comparisons and niche model calibration should include only areas that have been accessible to the populations under study [38]. We delimited this area by buffering a convex hull around known occurrences by 200 km (Fig 1) in SDMtoolbox [34]. This approach reflects a compromise between including all environments that have been accessible to the species, and still covering a broad-enough extent to minimize extrapolation and detect climatic differences between presence and background records [39].
To forecast climate change effects, we used fMaxent (fine-tuned Maxent, see below) and ensemble approaches [11,12] to calibrate models under present conditions, which were then transferred onto climate conditions for 2050 and 2080. Maxent is the most commonly used method in ENM, and it can fit arbitrarily complex models to explain relationships between environmental variables and occurrence data (version 3.3.3k; [40]). However, because an excessively complex model will be extremely specific to input data and perform poorly when extrapolating, Warren and Seifert proposed using a sample-size-adjusted Akaike information criterion (AICc) as criteria with which to address overfitting; this approach does not control model fit directly, but rather uses AICc to choose appropriate settings [7]. We used the “ENMeval” package [9] to fine-tune Maxent models by seeking the minimum value of AICc among candidate models. ENMeval provides an automated way to execute Maxent models across a user-specified range of regularization multiplier (RM) values and features combinations (FC). We set the RM range to 0.5–6.0 with increments of 0.5, and used 6 FCs, to cover a broad range of model settings. The block method was used to partition occurrence data into four bins, 3 of which were used for training and the remaining one for testing (bin combination, BC), which is desirable for studies involving model transferring [9]. In all, 2160 models (12 RMs × 6 FCs× 6 BCs × 5 occurrence groups) were generated for the four subspecies and for O. hupensis as a whole.
Ensemble models are used commonly in forecasting climate change effects, seeking to generate a consensus estimate that reduces individual model uncertainty by reflecting the central tendency of multiple models [11,12,41]. Here, outputs from six modelling algorithms, including generalized additive models (GAM), generalized boosted models (GBM), generalized linear models (GLM), random forests (RF), genetic algorithms (GARP), and the fMaxent model described above were included in ensembles. Individual GAM, GBM, GLM, and RF models were developed using BIOMOD2 [42], as implemented in R [43]; GARP models were developed in desktopGARP [44]. Details of implementation of each algorithm are provided in the supporting information (S2 Table). Model ensembles typically use a weighted averaging approach, in which models are weighted according to their interpolative performance (e.g. [30]). However, a recent assessment pointed to the challenge of balancing model interpolative accuracy against transferability [29,45]. Therefore, rather than using weighted averages, we used the PCA (median) method to identify the “central tendency” of individual model predictions [8,11]. The PCA measures, for each model, its ability to follow the general trend of predictions of the six models. This method calculates the median of the four individual models that had higher factor values among the six models [8,11,12].
The occurrence data used to fit the niche models were split randomly into two datasets, for calibration (70% of points) and interpolation evaluation (30% of points). Performance of individual and consensus models was evaluated via a partial ROC (receiver operating characteristic) approach [46]. Comparing to traditional AUC (area under the ROC curve), which was criticized because present data are more reliable than absence data in model evaluation [46], the partial ROC approach takes the quality of occurrence points into account and weights more on omission error [46]). Here, AUC calculations were limited to ROC spaces over which models actually made predictions, and only omission errors <5% were considered (i.e. E = 5%; [46]).
Final model runs incorporating all point data were used for visualizations and risk assessments. A modified least training presence threshold based on E = 5% was applied to fMaxent model predictions for O. hupensis and O. h. hupensis to generate binary predictions. We did not generate threshold predictions in ensemble future projection because such predictions are not applicable and hard to interpret (i.e. individual models for generating consensus models were different in present and future predictions).
Future climate variables were downloaded from WorldClim [31], the Consultative Group on International Agricultural Research (CGIAR), and the research program on Climate Change, Agriculture and Food Security (CCAFS). To reduce uncertainty regarding future climate conditions (S1 Fig), rather than using the 13 original global climate models (GCMs, S3 Table) from the IPCC 5th Assessment, the PCA (median) protocol was also used to generate consensus “climate models” among the 13 GCMs for each climate dimensions for 2050–2060 and 2080–2090 (S4 Table).
The fMaxent and ensemble models based on present predictions were applied to these future conditions. Future climate models applied to the intermediate scenario of representative concentration pathways of 4.5 (i.e. “RCP45”; [47]) in which future anthropogenic greenhouse gas emissions were estimated to peak around 2040. This scenario was chosen because it represents the middle range of available four scenarios, and as such is considered more realistic than models based on extremely high or extremely conservative scenarios [47]. Climatic similarities between present and future in 2050 and 2080 were assessed using mobility-oriented parity (MOP) metrics, a correction and simplification of multivariate environmental similarity surfaces [39].
Different degrees of overlap were observed in the eight climate dimensions among the four subspecies (S2 Fig). Oncomelania hupensis hupensis and O. h. robertsoni occupied similar temperature and precipitation dimensions in terms of annual mean temperature (bio1), mean diurnal temperature range (bio2), and annual precipitation (bio12), but not isothermality (bio3), temperature seasonality (bio4), mean temperature of warmest quarter (bio10), or precipitation of driest month (bio14); O. h. guangxiensis and O. h. tangi occupied similar temperature and precipitation regimes in terms of annual mean temperature (bio1), temperature seasonality (bio4), mean temperature of warmest quarter (bio10), annual precipitation (bio12), and precipitation of driest month (bio14), but not mean diurnal temperature range (bio2) or isothermality (bio3) (S2 Fig). The four subspecies showed diverse responses to precipitation seasonality (bio15).
Minimum volume ellipsoids occupied by the subspecies overlapped broadly (Fig 2). The size of the MVEs corresponded roughly to the geographic range extent of each subspecies (Figs 1 and 2), with O. h. hupensis and O. h. robertsoni occupying larger volumes than O. h. tangi and O. h. guangxiensis. Niche overlaps between pairs of subspecies also corresponded roughly to their genetic distances estimated by 16S sequence (Fig 2 and S3 Fig): i.e. the close relationship between O. h. hupensis and O. h. tangi coincided with the highest climatic niche overlap (D = 0.215) among all pairs (S3 Fig). Similar patterns were observed between O. h. tangi and O. h. guangxiensis (D = 0.147), but to a lesser extent (Fig 2 and S3 Fig). The null hypothesis of niche identity was rejected in all pairwise comparisons (S3 Fig). However, in analyses of niche similarity, the null hypothesis could not be rejected, except for O. h. robertsoni versus O. h. tangi (S3 Fig). Results of niche identity and similarity thus suggest that, although the four subspecies occupy unique climate spaces, the nonequivalence of niche spaces derives from a background effect, and not from biological differences.
Individual model performances varied across model algorithms and subspecies in interpolation validations (Fig 3). The machine learning methods (i.e. fMaxent, GBM, RF) generally showed better discriminant ability than regression models (i.e. GAM, GLM); GARP showed unstable performance (Fig 3). Similar to the machine learning models, consensus models showed good discriminant ability for the individual subspecies and for O. hupensis as a whole (Fig 3).
Using all of the occurrence data, parameters of AICc-selected models (i.e. fMaxent) differed from default settings (S4 Fig). Based on block partitions of occurrence data, mean AUCtest values of fMaxent models were 0.79, 0.79, 0.80, 0.86, and 0.94 for O. hupensis (as a whole), O. h. guangxiensis, O. h. hupensis, O. h. robertsoni and O. h. tangi, respectively, with fMaxent models of O. h. hupensis (AUCdiff = 0.09) and O. h. tangi (AUCdiff = 0.02) showing less overfitting than the other three (S4 Fig). In species-wide consensus models, the first principal component explained 46.7–72.2% of individual model variation (Table 1). Consensus models were discriminated by the first axis of the PCA, and each individual model was selected in consensus model processing (Table 1).
Variation among individual model predictions spatially was observed in both present and future (2050 and 2080; S5 Fig). Some areas identified as suitable nonetheless corresponded to environments beyond the climate envelope of the calibration area at present, thus involving non-analog climate conditions (S6 Fig). For example, fMaxent identified disjunct suitable areas around Beijing in northern China (Fig 4), but these areas involved model transfer into novel climate conditions (S6 Fig), making their interpretation uncertain and unwise. Within the distribution of each subspecies (Fig 1), projection of present ENMs onto future climate datasets generally involved little extrapolation (MOP metrics; S6 Fig). Transferring present-day models onto future climate scenarios, fMaxent models were more conservative than ensemble models (Figs 4 and 5): the western part of the predicted distribution based on consensus models was cleaner than predictions based on fMaxent, and the consensus method did not make the isolated predictions in the fMaxent model (Fig 4). Both fMaxent and consensus approaches identified a pattern of range expansion and suitability increase in O. h. hupensis (Figs 4 and 5). In O. h. robertsoni, both models identified an eastward shift, whereas in O. h. guangxiensis, a northward shift was indicated (Figs 4 and 5). In O. h. tangi, the two models showed contrasting predictions (Figs 4 and 5).
Binary predictions were based on fMaxent outputs, as thresholding future predictions from ensembles is difficult. Overlapping the Central Route and Eastern Route of SNWD with the binary future predictions for O. hupensis and O. h. hupensis, the southern Central Route coincides with suitable areas for O. h. hupensis in 2050–2060, and its suitable areas will expand northward along the southern Eastern Route by 2080–2090 (Fig 6). All of these areas are beyond the reach of present surveillance sites for schistosomiasis monitoring. Because a northward expansion of O. hupensis may occur considering future climate warming, these potential expansion areas need to be better covered by future surveillance efforts. Future surveillance efforts should also consider potential re-emergence of O. h. guangxiensis and O. h. tangi, as some areas of increasing suitability were noted for these two subspecies as well (Fig 5), although present intervention efforts have brought the snails to near extinction.
Limitations on materials and methodologies employed in this study need to be addressed here. While this paper focused on climate drivers, these factors occur in a complex milieu of other non-climatic drivers of snail distribution and parasite endemicity [21,22], although the non-climatic factors usually functioned at a small scale. Although we adopted ensemble forecasting approach to minimize the uncertainty of individual model predictions, the uncertainty exists in consensus models [29]. Ecological niche conservatism is of increasing importance given the complex impacts of ongoing climate change on biodiversity [4,6]. Many studies have evaluated niche conservatism across diverse evolutionary time spans [4,6]. Future projections for species involved in disease transmission and likely to respond to climate change are usually fraught with uncertainties and complexities; however, these assessments are crucial in identifying appropriate mitigation strategies [26]. Here, we tested climatic niche conservatism among the four subspecies of O. hupensis across mainland China, and integrated state-of-the-art modelling techniques (fMaxent and ensemble models) to forecast climate change effects. Our results have important implications regarding genetic divergence of O. hupensis and likely climate change effects on schistosomiasis transmission in mainland China.
The ecological niches of the four subspecies of O. hupensis were not identical, but we were unable to reject the null hypothesis that climatic niches are similar (except O. h. robertsoni versus O. h. tangi). Although failure to reject the null hypothesis does not assure that the climatic niche has been conserved, no evidence indicates that they have not been conserved, and broad climate spaces overlapped among the four subspecies (Fig 2 and S2 Fig). The relationship between niche overlap and phylogenetic relationships of the four subspecies further supports the idea that climate niches have been conserved (Fig 2 and S3 Fig). The signal of climate niche conservatism suggests that factors other than climate likely account for the genetic divergence of O. hupensis populations. Li et al. suggested that genetic differentiation of O. hupensis in mainland China is ultimately structured by landscape ecology [18], with populations falling into four different ecological settings (Fig 1): swamps and lakes in the Yangtze River Basin (O. h. hupensis); the mountainous region of Sichuan and Yunnan Provinces (O. h. robertsoni); the hilly, littoral part of Fujian province (O. h. tangi); and the karst landscape of Guangxi Autonomous Region (O. h. guangxiensis). This landscape-level segmentation of the four subspecies is generally consistent with the foundational work of Liu et al. [14]: indeed, clear geographic barriers separate the four subspecies (Fig 1; [14,16]). Climate niche divergence between O. h. robertsoni and O. h. tangi might relate to the long geographic distance separating them.
Previous studies have found that long-term climate warming tends to favor geographic expansion of S. japonicum in mainland China, but most such risk assessments have relied solely on mechanistic approaches (e.g. [23,25,48]). Although mechanistic models may be more desirable in that they estimate dimensions of the fundamental niche and in that they avoid problems with extrapolation [49], correlative ENMs have practical advantages in terms simplicity and flexibility, particularly as regards parameterization [50]. Comparing with mechanistic models, which predict a broad northward and westward expansion of S. japonica [23,25,48], correlative ENMs suggest a similar pattern, but with more detailed spatial predictions. Increased suitability and range expansion were observed consistently in O. h. hupensis, eastward and northward shifts in O. h. robertsoni and O. h. guangxiensis, and relatively stability status in O. h. gangi were observed in all our future model predictions (Figs 4 and 5).
Most current surveillance sites are distributed along the Yangtze River, designated to monitor transmission by O. h. hupensis and O. h. robertsoni. However, in a climate change context, both of these subspecies are expected to expand or shift distributionally (Fig 5). Surveillance sites distribution will have to broaden in coverage to be able to detect these shifts. In addition, the potential of O. h. guangxiensis and O. h. tangi to re-remerge should also be considered, as sites presenting increased suitability were identified (Fig 5). The southern parts of the Central Route of South-to-North Water Diversion (SNWD) project will become suitable for O. h. hupensis in 2050–2060, and suitable areas will expand northward along the southern parts of the Eastern Route of SNWD by 2080–2090: these areas are not covered by present surveillance efforts (Fig 6). Our results call for more rigorous monitoring and surveillance of schistosomiasis in the northern of potential expansion areas, although schistosomiasis currently has not been detected along either the southern Central Route or the Eastern Route; nonetheless, range expansion may open potential for emergence [48,51].
|
10.1371/journal.pgen.1000719 | Gene Dosage, Expression, and Ontology Analysis Identifies Driver Genes in the Carcinogenesis and Chemoradioresistance of Cervical Cancer | Integrative analysis of gene dosage, expression, and ontology (GO) data was performed to discover driver genes in the carcinogenesis and chemoradioresistance of cervical cancers. Gene dosage and expression profiles of 102 locally advanced cervical cancers were generated by microarray techniques. Fifty-two of these patients were also analyzed with the Illumina expression method to confirm the gene expression results. An independent cohort of 41 patients was used for validation of gene expressions associated with clinical outcome. Statistical analysis identified 29 recurrent gains and losses and 3 losses (on 3p, 13q, 21q) associated with poor outcome after chemoradiotherapy. The intratumor heterogeneity, assessed from the gene dosage profiles, was low for these alterations, showing that they had emerged prior to many other alterations and probably were early events in carcinogenesis. Integration of the alterations with gene expression and GO data identified genes that were regulated by the alterations and revealed five biological processes that were significantly overrepresented among the affected genes: apoptosis, metabolism, macromolecule localization, translation, and transcription. Four genes on 3p (RYBP, GBE1) and 13q (FAM48A, MED4) correlated with outcome at both the gene dosage and expression level and were satisfactorily validated in the independent cohort. These integrated analyses yielded 57 candidate drivers of 24 genetic events, including novel loci responsible for chemoradioresistance. Further mapping of the connections among genetic events, drivers, and biological processes suggested that each individual event stimulates specific processes in carcinogenesis through the coordinated control of multiple genes. The present results may provide novel therapeutic opportunities of both early and advanced stage cervical cancers.
| Genetic gains and losses, i.e. changes in gene dosages, are common abnormalities of human cancers. Discovering these defects and understanding the biological meaning can lead to improved therapeutic opportunities. This paper reports a large scale screening of gene dosage alterations in cervical cancer and gives a broader exploration of the expression and function of genes with gains or losses. We have focused on the most frequent gene dosage alterations and the alterations associated with survival after chemoradiotherapy, since these defects are likely to be of major importance for developing disease. The most notable finding was the discovery of a set of biological processes that are known hallmarks of cancer and were associated with gains and losses of specific genes. Moreover, novel loci associated with chemoradioresistance independent of existing clinical markers were found, and the genes involved were depicted. Our results indicated that gene dosage alterations play a causative role in the carcinogenesis and chemoradioresistance of cervical cancer and pinpointed candidate biomarkers of the disease.
| Cervical cancer is one of the most common malignancies affecting women worldwide and a major cause of cancer death for women globally [1]. Radiotherapy combined with cisplatin is the treatment of choice at the locally advanced stages [2]. Improved therapy is needed, since more than 30% of the patients show progressive disease within 5 years after diagnosis and treatment related side effects to organs within the pelvis are frequent. Tumor stage, size, and lymph node involvement are the most powerful markers of aggressive disease, but do not fully account for the observed variability in outcome and are not biologically founded. A better handling of the disease may be provided by the discovery of efficient biomarkers for therapeutic planning and intervention, but requires more insight into the mechanisms underlying cervical carcinogenesis and treatment relapse.
During carcinogenesis, genetic and epigenetic alterations drive the evolution of tumor towards increased malignancy and treatment resistance. The changes enable tumor cells to overcome microenvironmental constraints, sustain proliferation, and invade adjacent tissues and distinct organs [3]–[5]. Gene dosage alterations like gains and losses regulate the expression of genes and are motive forces for this evolution [6],[7]. Tumor cells bearing an increasing number of gains and losses successively emerge and are selected for based on the growth advantage caused by the genetic changes. Discovery and functional assessment of gene dosage alterations involved in carcinogenesis are therefore essential for understanding the biology of the disease.
At the locally advanced stages of cervical cancer, numerous gene dosage alterations and severe aneuploidy are frequently seen [8]–[10]. Moreover, pronounced intratumor heterogeneity in the gains and losses exists within the tumors, reflecting a high genetic instability [9]. The consequences of these alterations for the tumor phenotype are difficult to predict, since large chromosomal regions involving multiple genes are generally affected and some aberrations may be random events without biological significance [11]. Genome wide screening of DNA copy numbers in a decent number of patients enables identification of recurrent gene dosage alterations; i.e., alterations characteristic of the disease, and alterations associated with the clinical outcome [12], which are likely to be important in carcinogenesis and treatment resistance. Combining the data with expression profiles of the same tumors reveals the genes that are regulated primarily by the genetic events. The potential of this integrative strategy was recently demonstrated in a study on 15 early stage cervical cancers, where genes affected by aberrations on 1q, 3q, 11q, and 20q were reported [13]. Genetic events promoting tumor evolution and treatment resistance have, however, not been explored on a genome wide scale, and their biological meaning has not been addressed.
The present work was conducted to discover candidate driver genes and assess their function in the carcinogenesis and chemoradioresistance of cervical cancers. Genome wide screening of DNA copy numbers and expressions was performed in 102 patients with locally advanced disease. Of these, pairwise data were available for 95 patients. Reliable comparison of gains and losses across the patients was ensured by using the tumor ploidy, as determined by flow cytometry, and the GeneCount method to correct for the normal cell content of the samples and extract the absolute copy numbers and thereby the gene dosages [14]. The use of GeneCount also enabled mapping of the intratumor heterogeneity in the gene dosage alterations, providing information of the chronological order in which they had occurred during tumor evolution [14]. The recurrent gene dosage alterations, the alterations associated with outcome after chemoradiotherapy, and the genes that were regulated by these alterations were identified. Further analysis of gene ontology (GO) categories [15] was performed to identify biological processes that were overrepresented among the affected genes and therefore probably regulated by the gene dosage alterations. Such large scale and combined genomic, transcriptional, and functional analysis is powerful in detection of driver genes and their biological meaning, but has not been presented before. We demonstrate the potential of this approach by the identification of five biological processes in carcinogenesis that were associated with recurrent and predictive gains and losses of a set of genes. The set included four genes within the predictive losses for which repressed expression was related to poor outcome in the patient group and in an independent cohort of 41 patients. The genes are candidate drivers of the genetic events and novel biomarkers of cervical cancers.
Cervical cancer patients subjected to curative chemoradiotherapy were included in the study (Table 1). Most cases were squamous cell carcinoma and human papillomavirus (HPV) positive. Aneuploidy was seen in about half of the tumors, including some of the adenosquamous carcinomas and HPV negative cases (Figure S1A, S1B). Based on 97 patients, we generated an absolute gene dosage profile of the cancer genome by the use of array comparative genomic hybridization (aCGH) and the GeneCount analysis tool (Figure 1A). All chromosomes were affected with gains and losses, however, some regions were more frequently found to be aberrant than others (Figure 1B). Clustering of the patients based on gene dosages revealed no clear groups with characteristic aberrations.
The recurrent gains and losses were identified by considering both the amplitude and frequency of each alteration in Figure 1B [16]. Hence, a larger weight was given to high-amplitude events that are less likely to be random aberrations without biological significance. The recurrent alterations comprised more than 42% of the genome, and consisted of 14 regions (528 Mb) with gain and 15 (734 Mb) with loss (Figure 1C). Most of these alterations were also seen in the adenosquamous carcinomas and the HPV negative tumors (Figure S1C, S1D). The most common alterations were gain on 1q, 3q, 5p, 20q, and Xq and loss on 2q, 3p, 4p, 11q, and 13q, each involving 44–76% of the patients (Figure 1C, Table 2). High level amplification (seven regions) and homozygote deletion (six regions) helped to depict the peak of five recurrent gains and two recurrent losses (Table 2, Table S1). The frequency of the homozygote deletions was low (1–3%, Table S1), and none of the tumors had more than one of them. Homozygote alteration is therefore probably not a common mechanism of gene regulation in cervical cancers, in contrast to the highly frequent heterozygote deletion. The highest gene dosage of 36 was found in a diploid tumor with a copy number of 72 on 11q22.1-2 (Table 2).
Gene dosage alterations responsible for poor clinical outcome may not be as common as the recurrent ones. All alterations in Figure 1B were therefore included in the survival analysis. The LASSO method identified three regions with loss, 3p11.2-p14.1, 13q13.1-q21.1, and 21q22.2-3, which jointly showed the strongest association to progression free survival (Table 2). The 3p11.2-p14.1 and 13q13.1-q21.1 regions overlapped with the recurrent 3p12.3-p14.2 and 13q12.2-q21.32 losses, whereas the predictive loss of 21q22.2-3 was distal of the recurrent loss of 21q21.1-3. The predictive losses were not correlated and were related to poor outcome also when analyzed separately (Figure 2A–2C). The intratumor heterogeneity of the losses was low and similar to that of the recurrent losses (Figure 1D).
Most patients had more than one of the predictive 3p, 13q, and 21q losses. We therefore investigated whether there was an increased risk of relapse in cases of two or three losses. Kaplan-Meier plots for patients with different combinations of the predictive losses revealed three major groups with different outcome (Figure S3). Patients without any of the losses had a low risk of relapse and a survival probability of 91% (Figure 2D). Patients with 3p and/or 13q loss, without 21q loss, had an intermediate survival probability of 68%, whereas those with 21q loss had the lowest survival probability of 44%. The risk of relapse therefore seemed to be particularly high when loss of 21q22.2-3 was involved.
The predictive impact of the 3p, 13q, and 21q losses were assessed by multivariate analysis together with tumor size, stage, and lymph node status. Histological type, HPV status, and heterogeneity status showed no correlation to outcome in univariate analysis and were therefore not included. The losses and tumor size had independent predictive value (Table 3), showing that the gene data contained information of the progression free survival that was not covered by tumor size. Since tumor size is a strong predictor (Figure 3A), we also investigated the predictive impact of the three losses for small and large tumors separately. About 20% of the patients with tumor size less than the median had relapse and all of them had one or more of the losses (Figure 3B). In the cases of tumors larger than the median, about 47% of the patients progressed and all except two of them had one or more of the losses (Figure 3C). None of the patients with loss involving 21q were disease free after 28 months, suggesting a particularly high risk of relapse in cases of a large tumor bearing loss of 21q22.2-3. There was no difference in tumor size for patients with and without loss in Figure 3B or in Figure 3C (data not shown). The gene data therefore enabled identification of high and low risk patients both in cases of a small and a large tumor.
To find genes regulated by the recurrent and predictive gene dosage alterations, we used cDNA microarrays and generated a cancer gene expression profile. The profile was based on 100 patients, including 95 of those analyzed with aCGH. Expression data were available for 1357 of the about 4000 known genes within the altered regions, and a significant correlation to gene dosage was found for 191 of them (Table 2). Several correlating genes were identified for each region, except for 8q24.13-22, 10q23.31, and 11p12, where no genes were found. Typical examples of correlation plots are shown in Figure S4. The results were confirmed with the Illumina gene expression assay on 52 patients. Although the Illumina analysis was based on a lower number of patients, an excellent correlation between the Illumina and cDNA data and between the Illumina and gene dosage data was found for almost all of the genes, as demonstrated in Table S2. We also performed a second cDNA analysis, including only tumors with more than 70% tumor cells in hematoxylin and eosin (HE) stained sections. Totally 179 of the genes (94%) were identified, suggesting few false positive results due to normal cells in the samples. The observations supported our conclusion that the genes in Table 2 were gene dosage regulated. The latter analysis identified 26 genes that were not depicted when all patients were considered. These genes were not considered further, since the results were based on only half of the data set.
Expression of known oncogenes and tumor suppressor genes within the depicted regions, like MYC (8q24.21), BRCA2 (13q13.1), RB1 (13q14.2), and TP53 (17p13.1), was not significantly correlated to gene dosage. These genes are therefore probably not regulated primarily by gains and losses. The TP53 and RB1 results were consistent with the high frequency of HPV positive tumors (Table 1).
The predictive losses on 3p and 13q involved the same correlating genes as the corresponding recurrent ones, whereas PCP4, RIPK4, and PDXK were correlating genes within the predictive 21q region (Table 2). To depict the correlating genes that most probably were involved in development of chemoradioresistance, we required that the gene was significantly associated with clinical outcome both at the gene dosage and expression level. Moreover, a clear difference in the survival curves should also be seen in an independent cohort of 41 patients when based on the Illumina gene expression data. The criteria were fulfilled for four genes; RYBP and GBE1 on 3p and MED4 and FAM48A on 13q, which were termed predictive genes (Figure 4). Two more genes, GTF2F2 and RNASEH2B on 13q, were correlated to outcome based on the cDNA data, but were not considered further since the tendency based on the Illumina data was weak (p>0.15). The relationship to outcome was not strong enough for PCP4, RIPK4, and PDXK on 21q to be included among the predictive genes either.
Biological processes associated with the recurrent and predictive gene dosage alterations were found by comparing the GO categories of the affected genes with those of all genes in the data set [15]. One or more biological processes were annotated to 155 of the correlating and predictive genes and to 5824 of all genes. The categories apoptosis, carbohydrate metabolism, translation, and RNA-protein complex biogenesis and assembly were significantly overrepresented among the correlating genes within the recurrent gains, whereas macromolecule localization, generation of precursor metabolites and energy, transcription from RNA polymerase II promoter, and establishment or maintenance of chromatin architecture were overrepresented among those within the recurrent and predictive losses (Table 4). Fifty-six genes were included in the significant categories and were candidate drivers of the biological processes. In addition, we included the predictive gene FAM48A, which was not associated to any process in the GO database, as a potential driver of chemoradioresistance together with RYBP and MED4 (transcription) and GBE1 (generation of precursor metabolites and energy).
We generated a map to visualize the connections between genetic events, affected genes, and biological processes (Figure 5). The processes carbohydrate metabolism and generation of precursor metabolites and energy were combined in metabolism, translation and RNA-protein complex biogenesis and assembly were combined in translation, and transcription from RNA polymerase II promoter was combined with establishment or maintenance of chromatin architecture in transcription. The combined categories were closely related, justifying this strategy. All but six of the recurrent alterations were associated with a process and represented in the map. The predictive 3p and 13q losses were merged with the corresponding recurrent losses, since the regions overlapped, and linked to metabolism (GBE1) and transcription (RYBP, MED4) in addition to chemoradioresistance. The predictive 21q loss was not connected to any known gene, but associated with chemoradioresistance. The map revealed features that seemed to be characteristic of recurrent and predictive alterations in cervical cancer. First, many of the genetic events were associated with clusters of genes in the same biological process. For example, gain on 3q affected three genes in apoptosis and three in translation, gain on 5p was linked to tree apoptosis genes, and loss on 6q was associated with four genes in transcription. Second, several events, like gain on 3q, 19q, 20q and loss on 2q, 6, and 11q, were connected to more than one biological process, either through the regulation of several genes or because some genes had multiple functions.
This work presents the first coupling of gene dosage and expression profiles in a large sample set of cervical cancers. We based our study on absolute gene dosages, which are more sensitive than the commonly used aCGH ratios in detecting gains and losses and enable comparisons across tumors with differences in ploidy and normal cell content [14]. This strategy and the large number of patients ensured reliable identification of recurrent gene dosage alterations, events associated with clinical outcome, and their intratumor heterogeneity. Further analysis based on GO categories provided an objective way of organizing the numerous correlating genes into biological meaningful information. We demonstrate a large potential of the integrative approach by the discovery and functional assessment of candidate driver genes that represent novel biomarkers of the disease. In particular, novel loci associated with clinical outcome were identified, providing the first evidence that gene dosage can be responsible for developing chemoradioresistance in cervical cancers.
The recurrent gene dosage alterations were consistent with earlier reports on advanced stage cervical cancer based on conventional CGH [8],[9],[17]. However, a more precise definition of the altered regions was achieved here due to the improved resolution of the array technique. The high frequency of the alterations suggests that they play a causative role in carcinogenesis. Hence, many of the alterations are common also in other squamous cell carcinomas, like head and neck cancers [18],[19]. Moreover, the recurrent loss on 3p and 13q overlapped with the losses associated with poor clinical outcome, strengthening the hypothesis of a central role in tumor evolution. Less frequent alterations can, however, also be crucial for tumor evolution, as was demonstrated by the recurrent gain on 11q22 in 14 patients and predictive loss on 21q in 23 patients.
The low intratumor heterogeneity of the recurrent and predictive gene dosage alterations indicated that they had occurred prior to many of the other alterations. The result was consistent with our previous cervical cancer study based on conventional CGH [9], showing a homogeneous intratumor distribution of the frequent gains on 3q, 5p, and 20q and losses on 3p and 11q14-qter. Moreover, regions overlapping with the 1p, 1q, 3q, 8q, 9q, and 20q recurrent gains and 2q, 3p, 4p, 11q, and 17p losses have been found to be altered in precancerous cervical intraepithelial lesions [17], [20]–[23], suggesting that the events had occurred at an early stage. It is therefore likely that the alterations identified here, and the consequently control of biological processes and development of chemoradioresistance, emerge early during carcinogenesis. It should be noted that a low heterogeneity was seen for some of the less common alterations as well, implying that they had occurred early. The affected genes in these regions may also be crucial for tumor evolution, however, other mechanisms than gene dosage alterations, such as epigenetic events or mutations, probably play the major role in their regulation. Moreover, some of the highly heterogeneous alterations may be important for disease progression a later stage, being a result of the continuing tumor evolution towards increased aggressiveness.
The gene dosage alterations were associated with specific biological processes that are closely related to known cancer hallmarks [3]–[5], indicating that the genes involved are drivers of carcinogenesis. Hence, gain of the genes in apoptosis, including the anti-apoptosis genes BIRC2, BIRC3, and ATF5, can help carcinoma cells to evade apoptosis [3]. Aberrations of the genes in metabolism, like gain of ARNT and IDH3G in carbohydrate metabolism, and loss of COX7C and ATP5J in oxidative phosphorylation, can be part of a metabolic reprogramming towards increased glycolysis and decreased mitochondrial function to meet the high energy demand linked to tumor growth [4]. In particular, gain of ARNT may increase hypoxia and hypoglycemia tolerance by signaling through the HIF1A pathway [24]. Loss of the genes in molecular localization, including HRB and TSG101, can lead to abnormal protein internalization and recycling and thereby abrogated degradation of proteins like growth factor receptors [25],[26]. Finally, aberrations of the genes in translation and transcription, such as gain of the translation initiation factors EIF4A2, EIF4G1, EIF2S2, and EIF2S3 and loss of the transcriptional repressors HDAC2 and HDAC4, can be a way to control the formation and activity of essential proteins. The EIF-proteins are central in adaptation to hypoxia and can stimulate MYC translation and thereby oncogenic processes like cell proliferation [27],[28]. Improper function of HDAC2 and HDAC4 may also increase proliferation [29]. Many of the genes, including BIRC2, BIRC3, ATF5, NUP62, FASTKD3, IDH3G, and POFUTI, have been found to be regulated by gains or losses in previous cervical cancer studies [30]–[33]. Our findings link each gene to one or more specific biological processes, and thereby indicate the functional meaning of the genetic events in carcinogenesis.
Loss and down regulation of GBE1 and RYBP on 3p and MED4 and FAM48A on 13q were associated with poor clinical outcome, suggesting that the genes are drivers of chemoradioresistance. The mechanisms underlying these findings and possible associations to known aggressive phenotypes like hypoxia and rapid proliferation [34]–[36] are not clear, but a tumor suppressor function of the genes has been indicated. GBE1, which plays a role in carbohydrate metabolism, has been found to be down regulated in ovarian cancers [37]. Loss of the transcriptional repressor RYBP may impair death receptor-mediated apoptosis [38],[39], and the encoded protein has been shown to be down regulated in many tumor types, including cervical cancer [40]. Loss of the transcriptional activators MED4 may impair transcription of genes with anti-cancer effect, like the vitamin D receptor [41],[42]. The function of FAM48A is less clear, but some studies indicate that loss of this gene can promote aggressiveness. Hence, FAM48A is required for activation of the MAPK p38 pathway [43], which represses cell proliferation [44]. We found no candidate driver gene of chemoradioresistance within the predictive loss on 21q. Only a few tumor suppressor genes have been identified in this region. One candidate is the transcriptional regulator PRDM15, which was not included in our cDNA data set [45]. Our data showed, however, no correlation between PRDM15 expression, assessed with the Illumina method in 52 patients, and gene dosage (data not shown), suggesting that the gene is not regulated by genetic loss. Further investigation with denser microarrays or possibly microRNA screening would be needed to find the drivers in this region.
The connection between genetic events, genes, and biological processes may provide insight into more general aspects of cervical carcinogenesis. Several genes were often associated with a single genetic event, supporting the hypothesis that there can be multiple drivers of an event that coordinately promote tumor evolution [11]. In cases of genes in the same biological process, like the anti-apoptosis genes BIRC2 and BIRC3 on 11q22, a broad and therefore efficient control of the process may be obtained. Hence, BIRC2 and BIRC3 may play complementary roles in apoptosis evasion, since upregulation of BIRC3, but probably not BIRC2, may impair hypoxia induced apoptosis [46],[47]. In cases of genes in different biological processes, such as metabolism (NDUFS1), macromolecule transport (HRB), and transcription (SMARCAL1, HDAC4) on 2q, the collective control of these processes through a single event is likely to give a growth advantage that is selected for in carcinogenesis. One or more genes in all biological processes were affected in most tumors due to the high frequency of the recurrent gene dosage alterations. All processes were therefore probably important, and the control of them through gains and losses seems to be a common feature of the disease.
The candidate driver genes represent novel biomarkers that may be utilized in the handling of cervical cancers. Diagnostic assessment of the biomarkers may help to understand the evolutionary status and therefore the biology of the cancer in individual patients. In particular, the predictive biomarkers may be used in addition to tumor size for classification of patients into risk groups in a personalized treatment regime. The biomarkers also open for the possibility to specifically repress biological processes in carcinogenesis by molecular targeting, and thereby interfere with tumor evolution. The use of drugs to inhibit translation by interaction with EIF-proteins has shown promising results [48] and been suggested as a tool to target tumor hypoxia [49]. The approach may be applied at all stages of the disease, since the genetic events probably emerge early. Moreover, improved outcome after chemoradiotherapy might be achieved by targeting the predictive biomarkers. Hence, viral-mediated delivery of RYBP has been shown to induce apoptosis in a number of cancer cell lines [38], and could be a useful strategy for the patients with loss of this gene.
A cohort of 102 patients was included for basic analyses to identify gene dosage alterations with aCGH (97 patients), affected transcripts with cDNA microarrays (100 patients), and to confirm the affected transcripts with the Illumina method (52 patients) (Table 1). An independent cohort of 41 patients was used to validate relationships between gene expression and outcome with the Illumina method (Table 1). All patients received external irradiation and brachytherapy combined with adjuvant cisplatin and were followed up as described previously [50]. Eighteen patients received extended radiation field due to enlarged common iliac and para-aortal lymph nodes. Progression free survival, defined as the time between diagnosis and the first event of locoregional and/or distant relapse, was used as end point. Six patients died of causes not related to cancer and were therefore censored. Tumor samples were collected at the time of diagnosis. One – four biopsies, approximately 5×5×5 mm in size, were taken at different locations of the tumor, immediately snap-frozen in liquid nitrogen and stored at −80°C until used for analyses. The study was approved by the regional committee of medical research ethics in southern Norway, and written informed-consent was achieved from all patients.
The aCGH experiments and generation of absolute gene dosage profiles have been described previously for all 97 patients (ArrayExpress accession no. E-TABM-398) [14]. The array slides were produced at the Microarray Facility at the Norwegian Radium Hospital and contained 4549 unique genomic BAC and PAC clones that covered the whole genome with a resolution of approximately 1 Mb. Genomic DNA was isolated from the biopsies, labeled, and co-hybridized with normal female DNA to the array slides. DNA from different biopsies of the same tumor was pooled. The biopsies of all except two patients had more than 50% tumor cells in HE stained sections from the middle part of the sample. Median tumor cell fraction was 70% (range 30–90%). After array scanning, image analysis, spot filtering, and ratio normalization, the GLAD algorithm was applied for ratio smoothing and breakpoint detection [51].
The cDNA microarray experiments have been presented previously for 48 of the 100 patients [50]. The array slides were produced at the Microarray Facility at the Norwegian Radium Hospital and contained more than 12000 unique cDNA clones, including most known oncogenes and tumor suppressor genes. Total RNA was isolated from the biopsies, labeled, and co-hybridized with reference RNA (Universal Human Reference RNA, Stratagene, La Jolla, CA) to the array slides. RNA from different biopsies of the same tumor was pooled. Only biopsies with more than 50% tumor cells in HE stained sections were utilized. Median tumor cell fraction was 70% (range 50–90%). All hybridizations were performed twice in a dye-swap design (ArrayExpress accession no. E-TABM-817). After array scanning, image analysis, spot filtering, and ratio normalization, the average expression ratios were calculated from the two data sets and used in the further analyses. The gene expressions were mapped to the gene dosages based on the exact chromosomal position of the cDNA and genomic clones, as derived from Ensembl (http://www.ensembl.org/Homo_sapiens/searchview).
Results based on cDNA data were validated with Illumina gene expression beadarrays in 52 of the patients subjected to aCGH and in the independent cohort of 41 patients. HumanWG-6 v3 beadchips (Illumina Inc., San Diego, CA) with 48000 transcripts were used. RNA was isolated from the biopsies as described above and amplified using the Illumina TotalPrep RNA amplification kit (Ambion Inc., Austin, TX) with 500 ng of total RNA as input material. cRNA was synthesized overnight (14 hr), labelled, and hybridized to the chips at 58°C overnight, according to the standard protocol. The hybridized chip was stained with streptavidin-Cy3 (AmershamTM, PA43001, Buckinghampshire, UK) and scanned with an Illumina beadarray reader. The scanned images were imported into BeadStudio 3.1.3.0 (Illumina Inc.) for extraction, quality control, and quintile normalization. The annotation file HumanWG-6_V3_0_R0_11282955_A was used.
The recurrent gene dosage alterations were identified based on a score that was calculated for each genomic clone by multiplying the average gene dosage amplitude with its frequency [16]. Gains and losses were considered in two separate procedures. Semi-discrete data were used, for which amplitudes lower than 1.1 were set to 1 when searching for gains and amplitudes higher than 0.9 were set to 1 when searching for losses. The score significance was assessed by comparison to similar scores obtained after data permutation [16], adjusting the p-value by a multiple testing procedure to control the false discovery rate (FDR) [52]. Recurrent alterations with an FDR q-value <5% were reported.
Gene dosage alterations associated with clinical outcome were identified with the LASSO method in the Cox proportional hazards model [53], as implemented in [54]. The LASSO is a method for variable selection and shrinkage in regression models when the number of covariates is larger than the number of individuals. By performing shrinkage in addition to selection, the LASSO is more stable than stepwise procedures where variables are either retained or discarded from the model sequentially, one at a time. In groups of highly correlated variables the LASSO tends to select only one variable in the group [55], and reported therefore one representative of each DNA region that jointly explained the outcome. Each region was thereafter found by selecting neighbouring genomic clones with strong correlation (r>0.9) to the one reported. Survival curves were generated by Kaplan-Meier analysis and compared by using log-rank test.
Spearman's rank correlation analysis with an FDR q-value <5% was used to search for significant correlations between gene dosage and expression. The analysis was based on semi-discrete data, retrieved as described above. To identify biological processes that were overrepresented among the correlating genes, the GO categories of the genes were compared with those of all genes on the array by using the master-target procedure with the Fisher's exact test in the eGOn software [15]. The GO categories were found in eGOn from public databases, based on the gene reporter EntrezGeneID.
|
10.1371/journal.pgen.1007299 | Cell polarity protein Spa2 coordinates Chs2 incorporation at the division site in budding yeast | Deposition of additional plasma membrane and cargoes during cytokinesis in eukaryotic cells must be coordinated with actomyosin ring contraction, plasma membrane ingression and extracellular matrix remodelling. The process by which the secretory pathway promotes specific incorporation of key factors into the cytokinetic machinery is poorly understood. Here, we show that cell polarity protein Spa2 interacts with actomyosin ring components during cytokinesis. Spa2 directly binds to cytokinetic factors Cyk3 and Hof1. The lethal effects of deleting the SPA2 gene in the absence of either Cyk3 or Hof1 can be suppressed by expression of the hypermorphic allele of the essential chitin synthase II (Chs2), a transmembrane protein transported on secretory vesicles that makes the primary septum during cytokinesis. Spa2 also interacts directly with the chitin synthase Chs2. Interestingly, artificial incorporation of Chs2 into the cytokinetic machinery allows the localisation of Spa2 at the site of division. In addition, increased Spa2 protein levels promote Chs2 incorporation at the site of division and primary septum formation. Our data indicate that Spa2 is recruited to the cleavage site to co-operate with the secretory vesicle system and particular actomyosin ring components to promote the incorporation of Chs2 into the so-called ‘ingression progression complexes’ during cytokinesis in budding yeast.
| Eukaryotic cells require division of all their cellular components before finally giving rise to two independent cells in a process named cytokinesis. Eukaryotic cells must build a physical barrier of plasma membrane, which is coupled with the remodelling of the extracellular matrix between the dividing cells. Plasma membrane is transported as part of vesicles by the secretory pathway. In addition, those vesicles carry key cargoes that need to be incorporated into the molecular machinery that promotes cytokinesis, the so-called actomyosin ring. We have previously described that a few components of the actomyosin ring form ‘ingression progression complexes’ or IPCs that coordinate late steps during cytokinesis such as the contraction of the actomyosin ring, ingression of the plasma membrane and extracellular matrix remodelling. Our findings provide new insights into the molecular mechanism by which cells coordinate the secretory pathway, essential for cytokinesis, with the other late cytokinetic steps mentioned above. The cell polarity protein Spa2 plays a key role in that coordination. Spa2 promotes the incorporation into IPCs of the glycosyltransferase Chs2, which is transported on vesicles and essential for extracellular matrix remodelling during cytokinesis in budding yeast. Taking into account that Spa2 protein contains conserved domains present in higher eukaryotes, our work contributes to an understanding of how these domains might play a key role in modulating membrane trafficking and the targeting of specific cargoes to their destination in higher eukaryotes.
| Before the end of mitosis, eukaryotic cells need to redirect the secretory machinery towards the site of division to ensure cells deposit additional plasma membrane between the two daughter cells. In addition, secretory vesicles transport key factors to enable cells to perform cytokinesis successfully [1–3]. Although the molecular mechanism is not understood, insertion of membrane and cargoes needs to be highly coordinated with the assembly and contraction of the actomyosin ring, ingression of the plasma membrane and the extracellular matrix remodelling [4–6].
In a process conserved from yeast to mammals, secretory vesicles are transported along actin cables by the type V myosin Myo2. Subsequently, the exocyst complex tethers secretory vesicles to sites of active exocytosis and membrane expansion. The exocyst complex was first identified in the budding yeast Saccharomyces cerevisiae and consists of eight subunits: Sec3, Sec5, Sec6, Sec8, Sec10, Sec15, Exo70 and Exo84. Two of them, Sec3 and Exo70, are located to where the secretory vesicle will be targeted and directly bind to PI(4,5)P2, which is situated at the inner leaflet of the plasma membrane. The remaining exocyst components are associated with secretory vesicles [7–9]. It has recently been reported that the exocyst has an open-hand conformation, which explains how the complex tethers secretory vesicles to put them in contact with the plasma membrane [10]. Finally, the association between vesicle and plasma membrane proteins of the SNARE complex (soluble N-ethylmaleimide-sensitive factor attachment protein receptor) leads to the fusion of secretory vesicles with the plasma membrane [7–9, 11, 12]
In S. cerevisiae, apart from the plasma membrane incorporation during cytokinesis, another essential role for the growth machinery is to deliver key factors such as the chitin synthase Chs2, which lays down a special extracellular matrix layer, the primary septum, between mother and daughter cells. The type V myosin and the exocyst are required for Chs2 localisation to the site of division [13], suggesting a mechanism by which Chs2 is targeted to the cleavage site. However, there must also be a capture mechanism to ensure that Chs2-containing vesicles are incorporated into the cytokinetic machinery while cells are assembling the actomyosin ring prior to the contraction.
To identify novel factors that control the activity of Chs2 at the division site during cytokinesis in budding yeast, we undertook a systematic analysis of regulators of the chitin synthase Chs2 in budding yeast [14]. We previously reported the first part of this work, in which we identified that components of the actomyosin ring, including Chs2, form the so-called ‘ingression progression complexes’ or IPCs. We proposed that the function of these complexes is to coordinate contraction of the actomyosin ring, plasma membrane ingression and primary septum deposition [14]. Here, we describe the role during cytokinesis of the cell polarity protein Spa2, whose molecular details were previously unknown. We show that Spa2 binds to actomyosin ring components during cytokinesis in budding yeast, by interaction between conserved domains within Spa2 and the IPC components Cyk3 and Hof1. Localisation of Spa2 at the cleavage site requires the presence of both IPC components and the growth machinery. We found that Spa2 co-operates with the secretory vesicle system and specific IPC components to promote the incorporation of Chs2 into the cytokinetic machinery.
To understand how cells control the activity of the chitin synthase Chs2 at the division site during cytokinesis, we previously isolated proteins that were able to interact at the same time with Chs2 and one of its regulators, the protein Inn1 [14]. Using mass spectrometry, we identified a specific set of proteins that interact with Inn1-Chs2 complexes at the cleavage site during cell division. Initially we focused on the known core components of the budding yeast actomyosin ring, which we named ‘ingression progression complexes’ (IPCs). The IPCs contain, together with Chs2, the type II myosin Myo1, the IQGAP protein Iqg1, the F-BAR protein Hof1 and the cytokinesis regulators Inn1 and Cyk3 [14].
Targeting secretory vesicles to the cleavage site is essential for cytokinesis and occurs from yeast to animal cells. Cells need to incorporate new plasma membrane in order to expand the cell surface and create a physical barrier between mother and daughter cells [1, 4]. In addition, secretory vesicles carry essential cargoes such as the protein Chs2 in budding yeast, a transmembrane protein that is transported to the site of division at the end of the cell cycle [4, 13]. To understand how secretory vesicles carrying Chs2 are incorporated at the site of division and allow Chs2 to be part of the IPCs, we focused our attention on the list of specific proteins that simultaneously interact with Chs2 and Inn1. We identified two proteins by mass spectrometry that could potentially help us to understand the process, since they were previously known to be involved in polarised growth and vesicle transport. The more abundant of the two factors was Spa2 (Fig 1A (i)), which had been suggested to have a role during cytokinesis, although the molecular details were largely unknown. spa2 mutants show a mild defect in cell separation, and genetic interactions have been described between SPA2 and other genes involved in cytokinesis, including septin CDC10 and IPC components MYO1, CYK3 and HOF1 [15–21]. In addition, we identified myosin type V Myo2 (Fig 1A (i)), which is involved in vesicle transport and delivers essential cargoes such as the chitin synthase Chs2 to the site of division [13]. Together with Spa2, cell polarity proteins Bud6, Pea2 and Bni1 have been described to form the so-called ‘polarisome‘, which plays a role in cell growth [22, 23]. However, we were unable to identify any of the other polarisome components in our mass spectrometry analysis. To confirm that Spa2 and Myo2 were able to interact with the IPC component Inn1, control cells and cells expressing Inn1 fused to TAP were grown at 24°C in the presence of glucose, and cells were synchronised in G1 phase with mating pheromone. Subsequently, we released cells from G1 block and monitored the progression through the cell cycle (Fig 1A (ii)). We pulled down the fusion protein Inn1-TAP from cells released from G1 block for 90 minutes to enrich for cells undergoing cytokinesis and showed that Inn1 co-purified with Spa2 and Myo2 (Fig 1A (iii)). Although the role of Myo2 as a motor protein has been described, the function of Spa2 during cytokinesis was not understood. This prompted us to investigate the Spa2 protein further to understand how secretory vesicles are incorporated into the cleavage site and how the chitin synthase Chs2 is integrated into the IPCs.
To confirm that Spa2 could immunoprecipitate components of the IPCs, we grew TAP-SPA2 and control cells as described above for Fig 1A. We immunoprecipitated TAP-Spa2 and found that Spa2 interacted with all IPC components (Fig 1B (i) and (ii)). To determine precisely when during the cell cycle Spa2 interacts with IPCs, the protein TAP-Spa2 was pulled down from extracts of cells that had been arrested in G1 phase, cells that were going synchronously through S phase, or cells undergoing cytokinesis (Fig 1C (i)). We found that Spa2 only interacted with IPC components at the end of the cell cycle, which suggests a role for Spa2 during cytokinesis (Fig 1C (ii) and (iii)).
To understand the function of Spa2 at the site of division, we first constructed a strain that expressed Spa2-GFP and in which the type II myosin, Myo1, was fused to the red fluorescent protein tandem tomato, Myo1‐Tomato. These cells were released from G1 arrest at 24°C and time-lapse video microscopy was then used to examine when exactly Spa2 localises at the site of division. We found that Spa2 is recruited to the site of division a few minutes before the actomyosin ring starts to contract (Fig 1D), which would suggest precisely the time when CDK-associated kinase activity is inactivated and exactly when Inn1 and Chs2 localise at the division site [24–29]. IPC components appear at the site of division as medial rings and contracted dots, which shows them to be part of the contracting ring. However, Spa2 co-localised with Myo1 at an early stage, but it seems that Spa2 did not share the same localisation with Myo1 later, as it did not appear as a contracted dot (Fig 1D). Spa2 interacts with septins [20], which act as a barrier to compartmentalize proteins around the cleavage site [30]. Therefore, the septin ring might play a role to keep Spa2 at the site of division during contraction. Our data indicate that Spa2 may share a role with IPC components before actomyosin ring contraction starts.
To understand the role of Spa2 during cytokinesis, we first determined which of the IPC components were able to interact with Spa2. Although the biological significance was not found, genome-wide screens and genetic analysis showed genetic evidence that SPA2 could share a role with CYK3 and HOF1 [19, 31, 32]. To verify that there is a synthetic lethality between SPA2 and CYK3 genes, the meiotic progeny of spa2Δ cyk3Δ diploid cells were analysed by tetrad analysis. We confirmed that deletion of the SPA2 gene in cells lacking the CYK3 gene led to cell death (S1A (i) Fig).
To explore whether Spa2 and Cyk3 can interact physically, two different approaches were taken. First, we used the yeast two-hybrid assay and subsequently studied whether both proteins were able to interact directly in an extract of E. coli cells. Spa2 protein contains a predicted coil-coiled region and 25 time 9-amino-acid repeats. In addition, Spa2 contains five so-called Spa2 Homology Domains (SHD-I to V), which are conserved domains with the budding yeast Sph1 protein (Fig 2A (i)) [33]. Interestingly, it is the SHD-I that seems to be conserved in higher eukaryotes [33, 34]. It has been reported that SHD-I, SHD-II and SHD-V are the relevant domains for the described dynamics of Spa2 [21, 33, 35, 36]. On the other hand, Cyk3 protein comprises two domains: an N-terminal SH3 domain and a transglutaminase-like domain located in the middle of the protein (Fig 2A (i)). Using a yeast two-hybrid assay, we determined that full-length Cyk3 was able to interact with a fragment of Spa2 that contains the SHD-I (Spa2-1-145) (Fig 2A (ii)). In addition, we found that protein fragments containing the SH3 domain (Cyk3-1-74) and the rest of the protein including the transglutaminase-like domain (Cyk3-68-885), bind to the same fragment of Spa2 (Spa2-1-145) (Fig 2A (ii)). We determined that the Cyk3 transglutaminase-like domain (Cyk3-475-764) and the C-terminal end of Cyk3 (Cyk3-765-885) could interact with Spa2 (Spa2-1-145) as well. Therefore, it seemed that Cyk3 contains multiple sites that bind to Spa2. We found that both Cyk3 domains (SH3 domain and the transglutaminase-like domain) share a function with Spa2, since inactivation of both SH3 (cyk3-SH3Δ) and the transglutaminase-like domain (cyk3-2A; [14]) induced cell death in spa2Δ cells (S1A (ii) and (iii) Fig). This confirms the functional importance of both the SH3 and the transglutaminase-like domains of Cyk3 for Spa2 function.
Next, we generated an E. coli strain that expressed Strep-tagged SH3 domain of Cyk3 (Strep-tag-Cyk3-SH3) and, in parallel, another strain that produced a truncated version of Spa2 fused to 6His (6His-Spa2-1-552; note that His-tags throughout this work were only used for protein detection, not for purification purposes). We then mixed the cultures and generated a single cell extract containing the SH3 domain of Cyk3, Spa2-1-552 and all native E. coli proteins. We purified Strep-tag-Cyk3-SH3 from the cell extracts and found that the N-terminal half of Spa2 containing the SHD-I domain co-purified with SH3 domain of Cyk3 (Fig 2B). In addition, using the same experimental design, we expressed a fragment of Cyk3 that contained the transglutaminase-like domain fused to 6His tag (6His-Cyk3-475-885) and, in parallel, a strain that produced the strep-tagged N-terminal truncated version of Spa2 as above (Strep-tag-Spa2-1-552). We purified Strep-tag-Spa2-1-552 from the cell extracts and found that the fragment of Cyk3 containing the transglutaminase-like domain co-purified with Spa2 (Fig 2C). Our data indicate that Cyk3 can bind directly to the N-terminus of Spa2.
We confirmed the synthetic lethality between SPA2 and HOF1 by tetrad analysis of the meiotic progeny of spa2Δ hof1Δ diploid cells (S1B (i) Fig) [19]. As mentioned above in relation to Cyk3, this genetic evidence suggests that Spa2 might share a function with Hof1. In fact, it had been previously shown that Spa2 interacts with Hof1 using fluorescent reporters, although the role of such association was unknown [37]. Hof1 contains an N-terminal F-BAR domain followed by an unstructured region and a C-terminal SH3 domain (Fig 3A). This same structure is observed in other Hof1 orthologues including Cdc15 in Schizosaccharomyces pombe, which is involved in actomyosin ring assembly and membrane dynamics [38, 39]. We performed a yeast two-hybrid assay with three different fragments of Hof1 and the first 145 amino acids of Spa2 containing its SHD-I domain (Fig 3A). We found that Spa2 SHD-I interacts with the SH3 domain of Hof1 and with its F-BAR domain, which has been recently crystallised to show that the F-BAR domain of Hof1 is formed of an elongated crescent-shaped dimer [40]. We narrowed down the area within the F-BAR domain that binds to Spa2-1-145. Amino acids 200 to 272 of the F-BAR domain, which corresponds to the convex side of the F-BAR dimer [40], are sufficient to enable such an interaction (Fig 3A). To determine whether Spa2 and Hof1 bind each other directly, we checked, as illustrated in Fig 2B, whether these factors were able to interact in E. coli extracts expressing the indicated fragments of Spa2 and Hof1. We found that the SH3 and F-BAR domains of Hof1 both bind directly to the N-terminal half of Spa2 containing the SHD-I domain (Fig 3B and 3C).
To detect amino acids that may be relevant in Spa2 interactions, Psi-BLAST searches and secondary structure analysis were performed to show that the amino terminal part of Spa2, which contains the SHD-I, is conserved in other fungal orthologues of Spa2. In addition, it is precisely the SHD-I domain that is conserved in higher eukaryotes [33, 34] (Fig 3D). We identified a stretch of amino acids that is well conserved between fungal and higher eukaryotic orthologues of Spa2 comprising positively charged amino acids, which may be important for Spa2 to bind to its partners. To examine the role of SHD-I, four consecutive conserved basic amino acids were mutated into alanines. A fragment of 145 amino acids containing the wild-type Spa2 SHD-I and the mutated Spa2 SHD-I (Spa2-1-145-GAGA) were used to perform a yeast two-hybrid assay (Fig 3D). We found that the change in these amino acids blocked interactions with SH3 domains of Cyk3 and Hof1, and with the fragment of Cyk3 containing the transglutaminase-like domain of Cyk3 (Cyk3-475-764) (Fig 3D). In contrast, Spa2-1-145-GAGA was able to interact with the C-terminal end of Cyk3 (Cyk3-765-885) and a minimal protein fragment of the F-BAR domain of Hof1 (Hof1-200-272) that we found to interact with Spa2 (Fig 3D).
Taken together, these data indicate that Cyk3 and Hof1 interact directly with Spa2. The SH3 domains of Cyk3 and Hof1, and the transglutaminase-like domain of Cyk3 play a key role with Spa2 SHD-I domain. Positively charged residues within Spa2 are essential for those interactions. In addition, Spa2 SHD-I is able to interact with the F-BAR domain of Hof1 and the C-terminal fragment of Cyk3, but the stretch of positively charged residues that lies within SHD-I seems to be irrelevant for these interactions. Additionally, we found that simultaneous deletions of the SH3 and F-BAR domains of Hof1 induced cell death in spa2Δ cells (S1C Fig (iii)), unlike what occurs when single deletions of the SH3 or F-BAR domains of Hof1 are combined with the lack of Spa2 (S1C Fig (i) and (ii)). This confirms the functional importance of both Hof1 domains for Spa2 function (S1C Fig).
Spa2 interacts with components of the IPCs during cytokinesis (Fig 1). To determine whether the localisation of Spa2 is dependent on IPC components, yeast strains were generated in which the protein Iqg1 or Myo1 was fused to the degron cassette to permit depletion of Iqg1 or Myo1 protein levels [41, 42]. Initially, we grew asynchronous cultures of iqg1-td SPA2-GFP and control cells at 24°C before synchronising cells in G1 phase with mating pheromone (Fig 4A; S2 Fig, ‘-td‘ denotes temperature sensitive degron). After the induction of Ubr1 E3 ligase and a shift to 37°C to rapidly deplete Iqg1-td protein, cells were released from G1 block at 37°C [42]. We observed that both mutant and control cells progressed to anaphase in a similar manner. Unlike control cells, iqg1-td SPA2-GFP accumulated as binucleate cells and contained 4C DNA content as shown by flow cytometry (Fig 4A; S2 Fig), which reflects a failure of cell division. We found that rapid inactivation of Iqg1 prevents localisation of Spa2 at the site of division (Fig 4A (ii)). Similarly, we showed that Myo1 protein is required for Spa2 localisation (S3A Fig). To investigate whether Inn1 promotes Spa2 localisation at the cleavage site, we used the mutant inn1-td in which we were able to deplete Inn1 protein levels. We grew SPA2-GFP inn1-td and control cells in an identical fashion to that described for Fig 4A and found that Spa2 localisation was independent of Inn1 (S3B Fig).
We determined that Spa2 binds directly to the IPC components Hof1 and Cyk3. Consequently, to explore whether localisation of Spa2 at the cleavage site depends on the interaction with Hof1 and/or Cyk3, we used strains in which we fused the temperature-sensitive degron cassette to the N-terminus of Hof1 or Cyk3 proteins [43]. Cultures of hof1-td SPA2-GFP or cyk3-td SPA2-GFP strains were grown and synchronised as described for the experiment in Fig 4A. Cells were released after inactivation of Hof1-td or Cyk3-td. Next, we observed that the localisation of Spa2 is indeed defective in hof1-td cells, in which binucleate cells accumulated (Fig 4B). In addition, as localisation of Spa2 was only partially altered in cyk3-td cells (Fig 4B), we confirmed that Cyk3 protein had been depleted under the restrictive conditions (S4A Fig).
To understand what other factors promote Spa2 localisation at the site of division, TAP-SPA2 hof1-td and TAP-SPA2 control cells were grown as described for the experiment depicted in Fig 4B. After collecting cells going through cytokinesis synchronously in the presence or absence of Hof1, we made cell extracts and immunoprecipitated TAP-Spa2 on IgG-beads to find that in control cells Spa2 co-purified with components of the IPCs, as described above (Fig 1), and the type V myosin Myo2. In Hof1-depleted cells, Spa2 preserved the interaction only with Myo2 and Chs2, however, Spa2 was unable to interact with other IPC components (Fig 4C), which suggests that Spa2 is still able to interact with factors involved in secretory vesicle transport in the absence of any interaction with the IPCs at the site of division. It also suggests that Cyk3 and Inn1 interact with Spa2 through Hof1. To determine whether the Spa2 localisation defect in Hof1-depleted cells reflects a failure in secretory vesicle docking at the site of division or whether it is specific to certain secretory vesicles, we monitored the localisation of the exocyst component Sec8. The exocyst complex drives the delivery of secretory vesicles to the sites of growth during the cell cycle. Before actomyosin ring contraction starts, cells redirect the exocyst to the cleavage site [9]. Using iqg1-td cells, we confirmed that Sec8 localisation at the site of division is entirely dependent on the presence of an actomyosin ring (S5A Fig). Subsequently, we carried out similar experiments, inactivating Hof1 instead of Iqg1 (S5B Fig). We found that the exocyst component Sec8 has similar dynamics and localises equally at the site of division in the presence or absence of Hof1, which implies that Hof1 has no role in the docking of the exocyst at the site of division, although it is involved in Spa2 localisation.
Given that motor type V myosin, Myo2, can be pulled down by Spa2 in the absence of Hof1, we analysed the contribution of the secretory vesicle system to the localisation of Spa2 at the site of division. We fused the temperature-sensitive degron cassette to the N-terminus of Myo2 to permit conditional inactivation of the protein. Control and myo2-td SPA2-GFP cells were synchronised in G1 phase and then released into S phase in the presence of 0.2 M hydroxyurea, which inhibits ribonucleotide reductase and prevents chromosome replication. Under these conditions, polarised growth continues and buds are able to grow, allowing us to determine the role of Myo2 in the localisation of Spa2 at the site of division during cytokinesis, independently of Myo2 function in bud growth. Once cells had fully grown their buds, Myo2 was depleted by shifting cells to grow at the restrictive temperature of 37°C and expressing the E3 ubiquitin ligase Ubr1. In the absence of functional Myo2, we confirmed that the chitin synthase Chs2 was unable to be targeted to the site of division, as it was previously found using standard temperature sensitive mutants [13] (S6 Fig). We also determined that Spa2 localisation during cell division is only partially defective in the absence of Myo2-driven transport (Fig 4D (i) and (ii)). Spa2 localisation at the tip of new buds in myo2-td cells was totally diminished (Fig 4D (iii)), which confirmed that Myo2 was fully inactivated under these experimental conditions. We also confirmed that Myo2 protein levels were depleted soon after cells were transferred to the restrictive conditions (S4B Fig). Taken together, these findings indicate that Spa2 localisation at the site of division partially depends on Hof1 and the secretory vesicle transport, since inactivation of only one of them alone is not enough to block Spa2 localisation at the site of division.
To determine whether Hof1 and Myo2 share a role in the localisation of Spa2, we grew control and hof1-td myo2-td cells under the same conditions as described for Fig 4C and found the localisation of Spa2 to be completely dependent on both Hof1 and Myo2 proteins (Fig 5A; S7 Fig). There appear to be two populations of Spa2, one that binds to the IPCs via Hof1, and another that interacts with the secretory vesicle transport system. We speculated that these two populations might associate through Spa2 itself. To investigate whether Spa2 might interact with itself, a diploid yeast strain was generated in which one SPA2 gene expressed TAP-Spa2, while the other expressed Spa2-5FLAG. These diploid cells, together with diploid cells that lacked the expression of TAP-SPA2 as control, were grown asynchronously. Yeast protein extracts were made and TAP-Spa2 was subsequently pulled down on IgG beads. We showed that TAP-Spa2 interacted with Spa2-5FLAG (Fig 5B), suggesting that Spa2 might be able to form dimers. Using the yeast two-hybrid assay, we found that the amino-terminus of Spa2 containing the SHD-I domain was able to interact with a fragment of Spa2 comprising the SHD-II, which would leave open such a possibility (Fig 5C (i)). Furthermore, we found that both fragments were unable to interact if the SHD-I domain was mutated to eliminate positively charged amino acids (Fig 5C (ii)). Taken together, these experiments indicate that Hof1 and Myo2 share a role in the localisation of Spa2 and that Spa2 proteins might form dimers.
To determine whether Spa2 domains control localisation of Spa2 at the site of division in a different manner, the chromosomal SPA2 locus was modified so that cells expressed either Spa2-1-552-GFP (containing the SHD-I and SHD-II domains) or Spa2-553-1466-GFP (comprising multiple 9-aminoacid repeats and the SHD-V domain) under the control of SPA2 promoter. We grew SPA2-GFP cells in parallel with cells expressing either the N-terminal (1-552-GFP) or C-terminal fragment of Spa2 (553-1466-GFP). Cells were arrested in G1 phase and subsequently released synchronously to monitor the localisation of each construct at the site of division (Fig 5D; S8 Fig). We found that the localisation of the N-terminal half of Spa2 (1-552-GFP) was clearly defective (Fig 5D (ii); S8A Fig). Expression of Spa2 truncations was confirmed using immunoblotting analysis (S8B Fig). This suggests that the C-terminus of Spa2 is required for Spa2 to localise at the site of division. To investigate whether Spa2-553-1466 can interact with Hof1 in the same manner as shown above for Spa2-1-552 (Fig 2), we used the yeast two-hybrid analysis to show that, indeed, the fragment of Spa2 containing amino acids 553–1466 interacts with Hof1 (Fig 5E). We found that the SHD-V domain is sufficient to bind to Hof1 (Fig 5E). Taken together, these results indicate that Spa2 domains are able to interact with the IPC component Hof1. The C-terminal of Spa2 protein seems to be more relevant in the recruitment of Spa2 to the site of division.
The protein Spa2 is essential in the absence of either Cyk3 or Hof1 (S1 Fig) [19, 31, 32]. We proposed that understanding why Spa2 becomes essential in spa2Δ cyk3Δ cells and spa2Δ hof1Δ cells could reveal the molecular details of Spa2 role during cytokinesis. We and others have reported that Cyk3 and Hof1 proteins regulate the function of the chitin synthase Chs2, which lays down the primary septum between mother and daughter cells during cytokinesis [14, 28, 44–46]. To determine whether Spa2, together with Hof1 and/or Cyk3, has a role related to Chs2 function, we performed genetic analyses in which we tried to rescue defects associated with the double mutants spa2Δ cyk3Δ and spa2Δ hof1Δ. We constructed a diploid strain lacking one copy of CYK3, one copy of SPA2 and harbouring a hypermorphic allele of CHS2, which has enhanced chitin synthase activity [44]. The meiotic progeny was then analysed by tetrad analysis (Fig 6A). We found that hypermorphic Chs2 (CHS2-V377I) suppressed the growth defect at 24°C caused by the lack of the Cyk3 and Spa2 proteins (Fig 6A). Following the same strategy we also showed that CHS2-V377I rescues defects associated with spa2Δ hof1Δ cells at 24°C (Fig 6B).
Therefore, we found that growth defects associated with the lack of Cyk3 and Spa2, or Hof1 and Spa2 could be rescued by the hypermorphic allele of CHS2 (Fig 6C (i)). To complete the genetic analysis of CYK3, HOF1 and SPA2, we showed that growth defects associated with double degron hof1-td cyk3-td strains under restrictive conditions were rescued by the hypermorphic allele CHS2-V377I (Fig 6C (ii)). It seems that HOF1, CYK3 and SPA2 form a network of factors whose defects associated with the lack of function in pairs are always rescued by CHS2-V377I (Fig 6C (i)). Finally, we determined that SPA2 becomes essential for CHS2-V377I to rescue the absence of Hof1 and Cyk3, since spa2Δ hof1-td cyk3-td CHS2-V377I cells were unable to grow (Fig 6C (ii) and (iii), compare strains (2) and (5)). Taken together, these genetic analyses indicate that Spa2 plays a role related to the chitin synthase Chs2 during cytokinesis.
To look in greater detail at the Spa2-Chs2 functional relationship, we examined whether Spa2 physically interacts with Chs2. First we used the yeast two-hybrid assay to show that a fragment of Chs2 comprising its catalytic domain (Chs2-215-629) interacts with Spa2 truncations containing the SHD-I, SHD-II or SHD-V domains (Fig 6D). Subsequently, we used E. coli cells to express 6His-Spa2-1-552, which contains two of the domains that we found to interact with Chs2 in the yeast two-hybrid assay (SHD-I, SHD-II), in parallel with another strain that expressed Strep-tag-Chs2-215-629 that comprises its catalytic domain. We found that both proteins interacted directly (Fig 6E), which supports the hypothesis that Spa2 function during cytokinesis is related to chitin synthase Chs2.
Spa2 localises at the site of division a few minutes before the actomyosin ring contraction starts (Fig 1D). In addition, we found that inactivation of the IPC components Hof1 or Cyk3 alters the localisation of Spa2 and that the defect is clearly more severe when Hof1 is not present (Fig 4B). Furthermore, our genetic analysis showed that growth defects associated with the lack of Hof1/Spa2, Cyk3/Spa2 and Hof1/Cyk3 are rescued by the expression of a hypermorphic allele of CHS2. Finally, we found that Spa2 directly binds to Hof1, Cyk3 and, independently of Hof1, to Chs2 when it is being transported via motor protein Myo2. Therefore, we hypothesised that Hof1 and Cyk3 contribute to the docking of Chs2-containing vesicles at the site of division, for which Spa2 could play a role.
To test this hypothesis we first determined whether both Hof1 and Cyk3 shared a coordinated role in the localisation of Spa2 at the cleavage site. Cultures of SPA2-GFP and hof1-td cyk3-td SPA2-GFP cells were grown at 24°C and cells were synchronised in G1 phase with mating pheromone, before rapidly inactivating Hof1 and Cyk3 at 37°C. Upon release from G1 arrest at 37°C, hof1-td cyk3-td cells completed mitosis but were unable to divide in the same way as control cells (Fig 7A (i)). Localisation of Spa2 at the site of division was not observed in the absence of Hof1 and Cyk3 (Fig 7A (ii)), despite the presence of Spa2 protein in mutant cells (S9A Fig).
Since we have previously described that Chs2 localisation is defective in Hof1-depleted cells [14], and Hof1, Cyk3 and Spa2 seem functionally connected to Chs2, we aimed to confirm whether the localisation of Chs2 at the site of division was co-ordinately dependent on both Hof1 and Cyk3. We grew hof1-td cyk3-td CHS2-GFP and control cells as described above for Fig 7A. Chs2 protein levels were confirmed in control and mutant cells (S9B Fig). Interestingly, we found that localisation of Chs2 was completely defective if Hof1 and Cyk3 had been previously depleted (Fig 7B and S9C Fig). This was unexpected since the hypermorphic allele CHS2-V377I rescues hof1-td cyk3-td cells (Fig 6C (ii)).
On the other hand, we have previously shown that Chs2-V377I is constitutively active in vitro [44]. We anticipated that Chs2-V377I should localise at the site of division in order to rescue any defects associated with hof1-td cyk3-td cells. We investigated this by growing hof1-td cyk3-td CHS2-V377I-GFP and hof1-td cyk3-td CHS2-GFP cells as described for Fig 7B. Indeed, Chs2-V377I partially recovered Chs2 localisation at the site of division (Fig 7C (ii) and (iii)), as was reflected by the partial rescue of the cell division defect associated with hof1-td cyk3-td cells (Fig 7C (i)). Therefore, we concluded that structural changes promoted by the change in amino acid V377 were enough to drive Chs2 localisation during cytokinesis.
We have previously described that the Inn1 protein contributes to the localisation of Chs2 at the actomyosin ring [14] and that the hypermorphic allele CHS2-V377I rescues the defect in Inn1-depleted cells [44] (Fig 8A, compare (1) and (2)). To determine whether Inn1 is a key protein in the rescue of hof1-td cyk3-td cells by CHS2-V377I, we generated a strain that contained Hof1, Cyk3 and Inn1 fused to the temperature-sensitive degron cassette in order to deplete levels of all three proteins. We performed a growth assay and found that CHS2-V377I was no longer able to rescue hof1-td cyk3-td cells in the absence of Inn1 (Fig 8A, compare (4) and (6)). Our data would suggest that the three key factors that enable Chs2-V377I to dock at the actomyosin ring are Hof1, Cyk3 and Inn1.
Therefore, Inn1 might explain why CHS2-V377I localises at the site of division and rescues hof1-td cyk3-td cells. Inn1 directly binds to Chs2 [14]. We hypothesised that the association between Chs2-V377I and Inn1 might be stronger than between the wild-type Chs2 and Inn1, which would explain why CHS2-V377I rescues the defect in these cells. We generated E. coli strains that produced Strep-tag-Chs2-215-629, Strep-tag-Chs2-V377I-215-629 and 6His-tagged-Inn1. We then mixed the cultures in pairs (Strep-tag-Chs2-215-629/6His-tagged-Inn1, Strep-tag-Chs2-V377I-215-629/6His-tagged-Inn1 and empty vector control/6His-tagged-Inn1) and generated a single cell extract containing Chs2/Inn1, Chs2-V377I/Inn1 or no tagged protein/Inn1, together with all native bacterial proteins (Fig 8B). Next we purified the same amount of Strep-tag-Chs2 or Strep-tag-Chs2-V377I from the cell extracts, and subsequently determined that the amount of Inn1 protein purified with Strep-tag-Chs2-V377I was almost three times that of the Inn1 protein isolated from the cell extract with Strep-tag-Chs2 (Fig 8B). These findings suggest that Chs2-V377I interacts more strongly with Inn1, which must be enough to be incorporated in the IPCs in the absence of Hof1 and Cyk3.
The proposed model for Spa2 localisation at the cleavage site suggests dual mechanisms: Spa2 interacts with IPC components and with members of the secretory vesicle transport. To test whether Spa2 localisation at the site of division is recovered in cells in which Hof1 and Cyk3 proteins had been depleted, we investigated whether artificial recruitment of Chs2 (via Chs2-V377I) to the actomyosin ring was sufficient to induce Spa2 localisation. We grew SPA2-GFP CHS2 hof1-td cyk3-td cells and SPA2-GFP CHS2-V377I hof1-td cyk3-td cells as described above for Fig 7C. We observed that Spa2 protein can partially localise at the site of division when cells are expressing the hypermorphic allele of CHS2 in the absence of Hof1 and Cyk3 (Fig 8C). Our findings would suggest that part of the Spa2 protein population might require the arrival of Chs2-containing vesicles at the site of division in order to finally localise before the actomyosin ring contraction starts and successful cytokinesis takes place.
To investigate whether overexpression of Spa2 might increase Chs2 incorporation at the site of division, strains overexpressing SPA2, together with control, were grown at 24°C and cells were synchronised in G1 phase. Cells were then released from G1 arrest into medium containing galactose to allow overexpression of Spa2. Subsequently, samples were used to examine the incorporation of Chs2-GFP at the division site by fluorescence microscopy (Fig 9A). Cells overexpressing Spa2 increased the number of Chs2-GFP rings (Fig 9A (ii)), mainly due to the increase of rings with a fainter fluorescent signal associated with Chs2-GFP (Fig 9A (iii)). Normally, Chs2-GFP signal at the site of division initiates as a faint ring that turns into a ring with a stronger fluorescent signal before actomyosin ring contraction starts. As cell cycle progression seemed to be similar in control and GAL-SPA2 cells (Fig 9A (i)), we hypothesised that those faint rings might indicate that there is slightly more Chs2 protein at the site of division in cells overexpressing Spa2, and Chs2-GFP signal passed the threshold of detection under fluorescent light. To confirm our hypothesis, time-lapse video microscopy was used to examine Chs2-GFP localisation (Fig 9B). Cells were grown in the same way as described above for Fig 9A. 75 minutes after the release from G1 block, cells were shifted to Synthetic Complete (SC) medium and subsequently placed in a time-lapse slide to examine the localisation of Chs2 at 24°C every 2 minutes (see Methods). Both control and GAL-SPA2 cells were treated in an identical fashion, since the cultures were mixed before the cells were transferred to the time-lapse slide (the control cells expressed Spc42-eQFP and could therefore be distinguished from GAL-SPA2 cells). Twenty movies each were examined for control and GAL-SPA2 CHS2-GFP cells. We were able to detect Chs2-GFP signal 2 minutes earlier in 16 out of 20 cells that overexpressed Spa2 (Fig 9B). The kinetics of actomyosin ring contraction were similar in control and cells overproducing Spa2. The average period of contraction was similar in both types of cells (a mean value of 6 min in control compared with 6.15 min in the GAL-SPA2 cells). This finding would suggest that Spa2 promotes Chs2 incorporation at the site of division.
Since chitin synthase Chs2 promotes primary septum formation during cytokinesis, we investigated whether increased Spa2 protein can also induce higher levels of primary septum. Chs2 chitin synthase activity assays require the use of chs3Δ cells, as most of the chitin content is synthesised by chitin synthase Chs3 in budding yeast cells [47]. Cells were grown at 24°C and were synchronised in the G1 phase of the cell cycle. Then, we released cells from G1 block into medium containing calcofluor to stain primary septa and galactose to allow overexpression of Spa2. Progression through cytokinesis was similar in control and GAL-SPA2 cells (Fig 9C (i)). To observe calcofluor-stained chitin in cells completing mitosis, cells were collected 135 minutes after release from G1 block when the percentage of cells containing primary septa peaks [14]. We showed that the relative signal intensity of primary septa was more than twice as strong in cells overexpressing Spa2 (Fig 9C (ii) (iii)), which support that Spa2 induces Chs2 incorporation. Taken together, these findings suggest that Spa2 has a direct role in recruiting the chitin synthase Chs2 to the site of division in budding yeast.
Our data highlight a key role for the cell polarity protein Spa2 during cytokinesis in budding yeast, and provide the first evidence of how specific factors contained within secretory vesicles, such as the protein Chs2, are incorporated into the cytokinetic machinery. Spa2 has previously been reported to form the so-called polarisome complex, which includes Pea2, Bud6, and the formin Bni1 [22, 23]. The polarisome functions in actin cytoskeletal organisation during polarised cell growth, which is important for numerous cellular functions including differentiation, proliferation, and morphogenesis [22, 23]. However, we found none of the other polarisome components in our purified material associated with Chs2-Inn1, suggesting an independent function during cytokinesis for Spa2.
Spa2 directly interacts with IPC components during cytokinesis (Fig 10). Spa2 localisation requires the presence of an actomyosin ring and a functional secretory pathway contributes to Spa2 targeting at the division site. As has recently been described for the orthologue of Spa2 in Candida albicans [48], it seems very likely that kinase activity associated with mitotic CDK/Cyclin blocks Spa2 translocation to the cleavage site before chromosome segregation is resolved in S. cerevisiae, in a similar manner to that described for the assembly of the exocyst [49].
In Xenopus cells, disruption of the actomyosin ring blocks cleavage furrow formation, however the addition of membrane proceeds. This suggests that the actomyosin ring is important for restricting new membrane incorporation at the site of division [50–52]. Therefore, there must be a capture mechanism that allows the incorporation of transmembrane proteins transported on secretory vesicles into the actomyosin ring (Fig 10). We showed that the Spa2 protein might play such a role since it directly interacts with IPC components, Hof1 and Cyk3, and the chitin synthase Chs2. Spa2 could function as a bridge to allow specific Chs2 incorporation and membrane fusion of Chs2-containing vesicles at the site of division. It is of particular note that we found that increased levels of Spa2 promote Chs2 incorporation and primary septum formation. We have also previously shown that Inn1 regulates function and localisation of Chs2 [14, 44] and found that Inn1 plays a fundamental role in the incorporation of Chs2 at the cleavage site in the absence of Hof1 and Cyk3 (Fig 10). In addition, the C2 domain of Inn1 might also influence the fusion of Chs2-containing vesicles since the C2 domain of the human protein synaptotagmin is thought to contribute to the fusion of target membranes with synaptic vesicles [53].
The role of Spa2 protein during cytokinesis may well be conserved since S. pombe Spa2 was found to interact with Cdc15 [54], a Hof1 orthologue in fission yeast. Super-resolution microscopy and FRET techniques have recently revealed the nanoscale spatial organisation of fission yeast actomyosin ring components relative to the plasma membrane [55]. Spa2 is present in the same layer as proteins like Cyk3 and Fic1, the orthologue of budding yeast Inn1 [55]. Consistent with our results, Gould and colleagues were unable to find any association between other polarisome components and the cytokinetic machinery [54]. They found that Spa2 pulled down the Chs2 counterpart in fission yeast, the glycosyltransferase Bgs1 ((1,3)beta-D-glucan synthase catalytic subunit), which is required for primary septum formation [54, 56], although no molecular significance was noted. The Spa2-homology domain (SHD) is present in the mammalian GIT protein family, which is involved in cytoskeletal dynamics and membrane trafficking [34]. Thus, it seems that the role of the Spa2-homology domains in coordinating the assembly of larger complexes may be fundamental to modulating membrane trafficking and the targeting of specific cargoes to their intracellular destination.
The budding yeast S. cerevisiae strains used in this study were all based on W303 and are listed in S1 Table. Cells were grown in rich medium containing 1% yeast extract, 2% peptone and supplemented with 2% glucose (YPD), or 2% raffinose (YPRaff), or 2% galactose (YPGal). For all synchronisation experiments, asynchronous cultures of cells were grown overnight. The following morning cells were counted and diluted to a concentration of 4 x 106 cells per ml before allowing them to grow to a density of 7 x 106 cells per ml. To achieve synchrony of the yeast cultures we followed the previously described protocol [57]. To arrest cells in the G1 phase of the cell cycle, the mating pheromone α-factor (Pepceuticals Ltd) was added to a final concentration of 7.5 μg per ml. After 2 hours, additional 2.5 μg per ml aliquots of α-factor were added every 20 minutes and cells were checked using phase contrast microscopy until at least 90% of cells were unbudded. To release cells synchronously from G1 arrest, cells were pelleted, washed twice and released into fresh medium. Cells performing cytokinesis synchronously were collected 90 minutes after the release from alpha factor arrest when cells were grown in YPD medium at 24°C or 105 minutes when cells were grown in YPRaf at 24°C. On the other hand, cells were collected 75 minutes after the release from alpha factor arrest when cells were grown in YPGal medium at 37°C. Temperature and carbon source determine progression through the cell cycle. We arrested cells in the G2/M phase of the cell cycle by adding nocodazole to the medium at a final concentration of 5 μg per ml.
To stain primary septa of living cells, calcofluor was added 30 minutes after release from G1 block to a final concentration of 0.05 mg per ml and culture was incubated further for at least 60 minutes [58].
For time-lapse video microscopy in Fig 1D, cells were grown in YPD at 24°C, arrested in the G1 phase of the cell cycle. Cells were then released into YPD for 30 minutes before switching to Synthetic Complete (SC) medium to perform the time-lapse video microscopy. For time-lapse video microscopy depicted in Fig 8B, cells were grown in YPRaff at 24°C and arrested in G1 phase. Cells were then released into YPGal medium containing mating pheromone for 30 minutes before releasing cells from G1 block in YPGal medium for 75 minutes. Afterwards, cells were switched to SC medium in order to perform the time-lapse video microscopy.
In all experiments with temperature-sensitive degron strains (td), 0.1mM CuSO4 was included in the growth medium of exponential cultures before changing the carbon source to galactose to induce degradation [42]. To degrade proteins fused to the degron cassette, cells were transferred to YPGal medium at 24°C for 35 minutes to induce expression of GAL-UBR1, and then transferred to 37°C for 1 hour before release from the arrest [42]. For time-lapse video microscopy in S9C Fig, cells were grown in YPRaff at 24°C, arrested in G1 phase and degron protein depletion was achieved as described above. Subsequently, cells were released from G1 block into YPGal medium at 37°C for 30 minutes before switching to SC medium and placing them on the time-lapse slide to perform the time-lapse video microscopy. Cell cultures were maintained to monitor progression through the cell cycle for both strains under the same circumstances as time-lapse microscopy.
For experiments with myo2-td allele, cells were initially arrested in G1 phase using the mating pheromone α-factor. Subsequently, cells were released and blocked in early S phase of the cell cycle using hydroxyurea (Molekula Limited), which was added to a final concentration of 0.2 M. Cells were checked using phase contrast microscopy until at least 90% of cells contained a fully grown bud. At this point, Myo2-td inactivation was induced for 50 minutes, after which cells were released from hydroxyurea arrest.
Tenfold serial dilutions of fresh colonies of yeast cells were made and spots of cells containing between 50,000 and 50 cells were plated on the appropriate media. Plates were incubated for 2–3 days at the indicated temperature before the scan.
Two-hybrid analysis was performed using the vectors pGADT7 and pGBKT7 (Clontech). Cells were grown at 30°C on SC medium lacking leucine and tryptophan (non-selective) or lacking leucine, tryptophan and histidine (selective). They were scanned after 3–4 days of growth, as indicated.
The plasmids to express recombinant proteins in E. coli used in this study were based on the ‘pET’ series (Novagen) and are listed in S2 Table. Recombinant proteins were expressed in Rosetta cell line at 37°C for 2 hours after induction with 1mM IPTG. Subsequently, pairs of cultures with induced proteins of choice were mixed so that each cell extract would contain two recombinant proteins. As control, a culture with an empty vector was mixed with the corresponding cultures expressing recombinant proteins. In all cases, after mixing, cell pellets were frozen at -20°C. To study the interaction between proteins, the strep-tagged fusions were isolated from a cell extract on 1 ml of Strep-Tactin Superflow resin (2-1206-025, IBA GmbH) before eluting with 2.5mM d-Desthiobiotin (D1411 Sigma). We detected the indicated proteins by immunoblotting with the previously described anti-StrepMAB Classic (2-1507-001, IBA GmbH) and Penta-His (34660, QIAGEN) antibodies.
To monitor the association of proteins in yeast cell extracts, we used 1000 ml samples (1010 cells). Frozen cell pellets were ground in the presence of liquid nitrogen, using a SPEX SamplePrep LLC 6850 freezer/mill as described previously [59]. We isolated tagged proteins by immunoprecipitating with magnetic Dynabeads M-270 Epoxy (Invitrogen) coupled at 4°C to rabbit anti-sheep IgGs (Sigma S-1265). We detected the indicated proteins by immunoblotting with the previously described polyclonal antibodies to Inn1, Chs2, Cyk3 and Hof1 [14], or by using polyclonal, anti-Spa2-yC-16 (Santa Cruz sc-15578), anti-FLAG antibody (Sigma F-7425), M2 anti-FLAG monoclonal antibody (Sigma F3165), peroxidase-antiperoxidase (PAP) (Sigma P1291), monoclonal 9E10 (anti-MYC) or 12CA5 (anti-HA).
For mass spectrometry analysis of protein content, the digested peptides were analysed by nano LC/MS/MS with an ‘Orbitrap Velos’ (ThermoFisher). Data were processed as described previously (MS Bioworks) [14, 60, 61]. The total identification list was filtered at 1% FDR.
We prepared samples to measure the DNA content or to determine the proportion of binucleate cells by fixing cells with 70% ethanol and staining with propidium iodide as described previously [57, 62]. Flow cytometry was performed with a Becton Dickinson FACSCanto II. For binucleate cell analysis, samples were then processed and images acquired with an upright fluorescence microscope (Axio Imager M1; Carl Zeiss, Inc.) using a 63x 0.95NA objective, an HRm camera, a Rhodamine specific filter set (em:546/12, exc: 608/65) and Axiovision software. We examined at least 100 cells at each time-point.
Pictures of colonies on agar were taken after 24 hours (YPD medium) or 30 hours (YPGal medium) with a Nikon CoolPix 995 camera attached to a Nikon Eclipse E400 microscope. To observe GFP-tagged proteins cells were fixed with 8% formaldehyde for 10 minutes and subsequently washed twice with ice-cold PBS. Phase contrast and fluorescence microscopy images of cells grown in liquid culture were performed with a Nikon A1R Microscope and an Orca R2 camera (Hamamatsu) with objective lens Plan Apo TIRF 100x oil DIC 1.49NA, and LightLine single-band filter set FITC Semrock. The illumination source was a Nikon Intensilight C-HGFIE (ultrahigh presure 130W mercury lamp). We used NIS elements software. We analysed eleven z-sections with a spacing of 0.375 μm to facilitate the examination of whole cells for all experiments. Exposure time, sensor gain and digital adjustments were the same for control and experimental samples. We examined 100 cells for each time-point. Each experiment was carried out at least three times.
For time-lapse video microscopy cells were grown in an IBIDI cells in focus 15 μ -slide (8 well glass bottom; 80827) after their release from G1 arrest. The base of each well is formed of a glass coverslip that we coated with a 5 mg per ml solution of the lectin Concanavalin A (Sigma L7647), and then washed with water and dried for 30 minutes. A suspension of cells was then placed on the glass coverslip and incubated for 5 min in order to allow cells to attach. The coverslip was then washed with pre-warmed SC medium. Finally 300 μl pre-warmed SC medium was added. Time-lapse video microscopy illustrated in Fig 1D was performed using the DeltaVision system with an Olympus IX-71 microscope and a CoolSNAP HQ2 Monochrome camera. A Plapon 60X0 1.42 NA objective lens was used. The 300W xenon system with liquid light guide was used for illumination. Images were captured with Softworx Resolve 3D acquisition software. We analysed 8 z-sections with a spacing of 0.4 μm. For time-lapse video microscopy shown in Fig 9B and S9C Fig, 9 z-sections with a spacing of 0.4 μm were acquired with a widefield epifluorescence microscope (Nikon Eclipse Ti2) equipped with an APO TIRF x100/1.49 objective and an sCMOS camera (Hamamatsu Orca-Flash4.0). Brightfield and fluorescence images were sequentially acquired every 2 minutes for 1 hour at 24ºC or 37ºC. Focus drift was avoided with the help of a hardware-based focusing system (Nikon´s Perfect Focus System).
The microscopy data were deconvolved using Huygens (SVI) according to the “Quick Maximum Likelihood Estimation” method and a measured point spread function. The deconvolved data set was viewed with “ImageJ” software (National Institute of Health, USA) [63]
|
10.1371/journal.pgen.1006275 | Reverse Chemical Genetics: Comprehensive Fitness Profiling Reveals the Spectrum of Drug Target Interactions | The emergence and prevalence of drug resistance demands streamlined strategies to identify drug resistant variants in a fast, systematic and cost-effective way. Methods commonly used to understand and predict drug resistance rely on limited clinical studies from patients who are refractory to drugs or on laborious evolution experiments with poor coverage of the gene variants. Here, we report an integrative functional variomics methodology combining deep sequencing and a Bayesian statistical model to provide a comprehensive list of drug resistance alleles from complex variant populations. Dihydrofolate reductase, the target of methotrexate chemotherapy drug, was used as a model to identify functional mutant alleles correlated with methotrexate resistance. This systematic approach identified previously reported resistance mutations, as well as novel point mutations that were validated in vivo. Use of this systematic strategy as a routine diagnostics tool widens the scope of successful drug research and development.
| One of the most profound outcomes of fast, reliable genome sequencing is the ability to tailor drug therapy to an individual’s genotype. This ‘personalized’ or ‘precision medicine’ is the realization of a decades-long effort to maximize drug effect and limit unwanted side effects. An undesirable consequence of such targeted therapies, however, is the emergence of drug resistance. This outcome is the result of an evolutionary process where mutations in the drug target render the drug perturbation allow such mutant cells to proliferate. Because of the unbiased, and stochastic nature of the emergence of drug resistance, it is impossible to predict. We developed a test where hundreds of thousands of mutant cells are exposed to a drug simultaneously and those cells that modulate resistance survive. This method is innovative because it partners a high-throughput experimental protocol with a tailored statistical model to identify all mutations that modulate resistance. Finally, we used synthetic biology to re-create these mutations and demonstrate that they were, in fact, bona fide drug-resistant variants. These mutations were further extended and confirmed to also be resistant in the human orthologue. This combined biological-computational approach allows one to identify drug’s degree of resistance to both guide treatments and future drug discovery.
| Drug resistance is a worldwide health concern that affects all drug classes, including anti-infectives and anti-cancer agents [1–3]. Recent reports illustrate that first-line antibiotic treatment failure rates have increased 12% from 1991–2012 [4]. Cancer drug resistance has increased, in part due to the use of highly specific targeted therapeutics [1,5]. While attempts to combine drugs into “smart cocktails” hold some promise to reduce emergence of resistance, in the majority of cases drug resistance is inevitable. Therefore, it is important to understand the causative mechanisms of resistance to improve the use and targeting of therapeutics.
Current strategies for understanding the mechanisms of resistance include: i) observational trials [1,6], ii) in situ mutagenesis [7–10] and iii) computational approaches [7,11–13]. However, each of these methods suffer from limitations with respect to throughput, resolution and accuracy. Hence, a rapid, systematic and cost-effective strategy to identify gene variants that modulate drug resistance over time is required to improve our understanding of resistance mechanisms.
Here, we present such a streamlined method to identify the emergence and persistence of modulators of drug resistance. Our integrative approach combines a strategic parallel competitive in vivo resistance assay with a Bayesian statistical model [14,15] that is both systematic and quantitative. We applied this assay to the anti-cancer drug methotrexate (MTX) in its well-characterized target, dihydrofolate reductase. Our pipeline takes advantage of the S. cerevisiae variomics collection, which contains libraries of 2 x 105 random plasmid-borne point mutation alleles for every yeast gene [16]. These alleles are packaged within haploid-convertible heterozygous diploid yeast gene knockouts which can be grown competitively and quantified with massively parallel sequencing.
Yeast dihydrofolate reductase (DFR1) is a validated functional orthologue of human dihydrofolate reductase (hDHFR), which is commonly used to study the MTX mechanism of action and enzymology [15,17,18]. In previous work, yeast has been employed to study genome-wide gene-drug interactions [19–23], and is a well-established model for anticancer drug research [6,17,24,25]. Methotrexate acts as an antimetabolite that targets the enzyme dihydrofolate reductase, which functions to maintain folate homeostasis in nucleus and mitochondria by reducing dihydrofolate into tetrahydrofolate as a key element of thymidylate and protein synthesis [15]. Due to the high degree of conservation between yeast and human cellular pathways, the results obtained for the yeast dihydrofolate reductase can provide insights into how tumors acquire drug resistance, which is a major barrier to effective cancer treatment [26–28] and point mutations in DHFR active site have been shown to affect MTX binding affinity altering in turn MTX efficacy [8–10,29–37]. Thus, systematically surveying the causative DFR1 point mutations that correlate with poor MTX response and understanding how resistant dfr1 alleles interact with MTX will help develop MTX analogues with a potentially lower likelihood of resistance.
We first describe our novel integrative experimental and statistical analysis method. We then apply this method to the identification of variants that modulate resistance to methotrexate in its target, dihydrofolate reductase. We next present validation studies using reconstituted individual mutations grown in isolation. Finally, we use a DFR1 protein model to provide structure/function relationship analysis of the validated mutations.
The functional variomics technology was adapted in our study by using the original dfr1 variomics library, which contains 2 x 105 point mutations in DFR1 [16]. To recover as many distinct dfr1 MTX resistant-alleles as possible, we exploited the variomics tool by screening the diploid and haploid dfr1 pools using an improved screening assay (Fig 1 and Methods). Specifically, we wanted to test if the resulting alleles differed depending on if the wild-type DFR1 allele was present, as is the case for the DFR1/dfr1Δ heterozygote strain, or absent as in the haploid dfr1Δ strain. For haploids, the dfr1 allele must maintain viability and provide drug resistance whereas in the diploid case, the wild type allele can in principle allow separation-of-function alleles (i.e. resistance without viability) to be recovered.
We tuned the parameters of the drug resistance assay to maximize for the enrichment of dfr1 alleles in parallel competitive conditions in an attempt to mimic the environment in which heterogeneous tumors are exposed to cytostatic drugs [38,39] (Fig 1). The dfr1 diploid library was first grown without drug selection to generate a dfr1 pool with ~50-fold coverage per variant for each of the 2 x 105 independent variants (see Methods for details). The pool was then induced to sporulate to generate a haploid dfr1 pool of 2.2 x 104 viable dfr1 alleles which were then challenged with drug in liquid media. To minimize the loss of rare dfr1 alleles, drug exposure was limited to a 6-day treatment of the diploid and haploid pools in liquid media at a MEC100 dose of MTX (Fig 1 and S1 Fig). Treated samples were collected every 2 days (equivalent to 8 generations of growth) and the remaining dfr1 pools were further propagated in fresh media with MTX (S2 Fig). MTX-treated pools were harvested at each time point and plasmid-borne dfr1 alleles were PCR amplified and sequenced at a median coverage of 10K (Fig 1 and S1 Table; Methods).
The sequencing data was collected in separate runs for the diploid and haploid experiments and each processed independently (see Methods for details). To call variants and estimate their associated allele frequencies in the mixed dfr1 pools, we used our previously published rare variant detection statistical model (RVD2) [14]. We estimated the parameters of the model for each time point and for the wild-type control using the default Gibbs sampling and Markov chain monte carlo parameters (4000 Gibbs samples, 10 Metropolis-Hastings samples per gibbs sample, 20% warm-up, thinning rate of 2). Finally, we called variants using the somatic test function in RVD2. This test identifies variants where the difference in the non-reference read rate is between 0.1% and 100% between a designated case and control sample (95% posterior confidence). This test also filters for variant loci that have non-uniform, non-reference read counts to eliminate false-positive calls due to generally elevated sequencing error rates.
We used RVD2 to compare the non-reference read rate at the starting time point “T0” to that of all later time points (T1, T2, and T3). We denote the model’s estimate of the true non-reference read rate at each locus the Variant Allele Frequency (VAF) at that locus. This analysis identified 66 variant positions in the DFR1 locus in the diploid pool and 49 variant positions in the haploid pool (Fig 2; S2 and S3 Tables; and S3 Fig). Among the 35 (53%) coding mutations in the diploid pool, 28 were missense mutations. Exactly 11 of these 28 mutations (39%) correspond to highly conserved residues (Fig 2; S2 Table; and S3 and S4 Figs). We noted that missense mutations that affect M35V and M35T residues, which were previously shown to affect MTX binding affinity and/or MTX resistance, were recovered in our screen [14] (S2 Table). In the haploid pool in contrast, only 8 out of 17 coding mutations (47%) were found to be missense mutations, 3 of which correspond to residues that are conserved in hDHFR (Fig 2; S3 Table; S3 and S4 Figs).
We estimated the diversity in the diploid and haploid pools at each time point by comparing the number and frequency of the variants under selection to the number and frequency of variants in the background strain, which carries only the wild-type allele on the parental plasmid. This procedure accounts for changes in the number and relative abundance of variant alleles, and sequencing variation (see Methods for details). Because the wild-type allele-bearing strain was only sequenced with one replicate, the sensitivity of RVD is low and few variants were called. We record a—when fewer than two variants were called at a time point and no diversity score can be computed. For the haploid strain, the diversity scores for T0 - T3 were 1.37, -, -, - respectively. For the diploid strain, the diversity scores for T0 –T3 were -, 9.20, 7.13, 4.66 respectively. The diversity of the haploid strain is lower than the diversity of the diploid strain at any time point, as would be expected due to the viability requirement imposed on any haploid allele. The diploid diversity score decreases monotonically from T1 to T3. These results align with our expectation that diversity is higher for the diploid pool than the haploid pool and that diversity decreases with time under drug selection.
The nucleotide coding sequence positions called in the diploid and haploid backgrounds do not overlap, except for one (627T>C) (S2 and S3 Tables). We reasoned this is likely due to the random genetic drifts introduced by the sporulation and haploid conversion events (see Methods for details). Furthermore, two positions in the coding sequence identified in the diploid strain showed two minor allele changes over the MTX timecourse e.g. 187T>G (at T1 time point) and 187T>A (at T2 time point), which result in non-synonymous mutations S63A and S63T, respectively (S3 Fig and S2 Table). The emergence of all of these nucleotide changes at given time points suggests that these silent and non-synonymous mutations can have marginal effects in modulating MTX resistance in the diploid genetic background.
We compared the spectrum of variants in the haploid pool to those in the diploid pool prior to selection (T0), because dfr1 mutants that survive conversion to the haploid state must be viable. We found some DFR1 positions had high variant allele frequencies (VAFs) in the diploid and haploid pools (Fig 2). The pre-existing dfr1 haploid allele frequency did not predict the emergence of MTX-resistant variants in later time points (Fig 2; S3 Fig; and S3 Table). In contrast, in the diploid state, 5 out of 6 positions with high initial VAFs (over 10%) increased in abundance upon MTX exposure (Fig 2; S3 Fig; and S2 Table). These observations are consistent with a model in which pre-existing mutations required for viability are no more likely to confer drug resistance, while resistant alleles can be found as pre-existing in the diploid state. Our strategy of surveying dfr1 alleles in both diploid and haploid backgrounds allowed us to distinguish dfr1 mutations with dominant resistance phenotypes regardless of whether or not such pre-existing mutants with competitive fitness advantages were already present.
To validate the alleles identified by our high-throughput parallel assay as individual variants, all mutant alleles were reconstructed de novo and assayed for MTX resistance in individual growth assays (Methods). We selected all of the coding (non-synonymous and silent) mutations that increased in VAF at the earliest time point (S2 and S3 Tables) and integrated a full length synthetic gene into the chromosomal DFR1 locus of the isogenic DFR1/dfr1Δ or dfr1Δ strains, such that these alleles were under the control of the endogenous promoter (Methods). All of the DFR1 mutations in the haploid background were viable, however 6 out of 10 dfr1 mutants exhibited a slow growth phenotype (Fig 3A and S5 Table). Also, 9 out of 10 dfr1 alleles (not Q16Q) were reproducibly resistant to a sublethal dose of MTX (Fig 3A; Methods). The majority of the alleles (7 out of 10) are recessive based on our observation that they were no longer MTX resistant when a wild-type DFR1 copy was expressed (Fig 3A; Methods). In the diploid background, 10 out of 27 DFR1/dfr1 mutants that express non-synonymous mutations were confirmed to exhibit strong resistance to MTX, with at least 80% retained growth relative to each corresponding DMSO-treated strain (Fig 3B). Of these 10 DFR1/dfr1 mutants, 2 exhibit competitive fitness advantages while 3 represent hypomorphic alleles, (with predicted DFR1 catalytic defects) given their inability to survive under obligate respiratory conditions [40] (Fig 3B; S4 Table and Methods). These results, combined with the experimental differences between initial screen vs. validation experiments (e.g. re-screened individually vs. growth in competitive mixed culture) (Methods) suggest that some of the DFR1 mutations with marginal MTX resistance, including I55M, F68L, T156A, F178F, E187K and N209H might manifest resistance only when expressed in combination.
Ten of the validated MTX-resistant dfr1 alleles cluster in the functional binding pocket for folate, the substrate of Dfr1p, or its NADPH cofactor (Fig 4). Specifically, mutations in L27, M35, F38 and T141 correspond to residues that directly interact with MTX (or folate) in hDHFR (Fig 4B and 4C). We hypothesize that these mutations likely reduce MTX affinity to render the drug ineffective. Similarly, mutations found in V127 and T156 correspond to residues situated in the NADPH binding cleft (Fig 4C). Previous work has shown that specific C. galbrata DHFR inhibitors act by displacing the NADPH cofactor [41,42], suggesting that a similar mechanism could be at work for the V127 and T156 mutations identified in our screen. Other non-synonymous mutations identified in our screen, including F38Y, M35T and M35V, have also been previously reported to lead to MTX resistance [30,32,33,35]. Further, we identified a W29R mutant, a residue known to be essential for enzyme function [43]. Specifically, the side chain of W25 (W29 in yeast) forms hydrophobic aromatic stacking interactions with both MTX (PDB ID 1U72)[10] and folate (Fig 5A) (PDB ID 1DHF)[44]. In the latter case, the protein-ligand interaction is further stabilized by a hydrogen bond formed between the Nɛ of W25 and the hydroxyl group of the folate pteridine ring. We also noted that a Trp residue is conserved at this position, with the notable exception of the recently-identified hDHFR-like 1 (hDHFRL1) protein [45], which has an Arg residue at this position (S3 and S4 Figs). This is the same substitution that we obtained in our screen (W29R) (Fig 2 and S2 Table). An arginine at this position is unable to form the key hydrophobic interactions with MTX [46] and therefore we hypothesized that hDHFRL1 may be resistant to MTX. To test this and to explore the possibility that resistance arises from destabilizing hydrophobic contacts with antifolate [46] (Fig 5A), we investigated the growth fitness of a wild-type human hDHFRL1 construct bearing an arginine at position 25 in a DFR1/dfr1Δ heterozygote strain (Methods). This change did indeed render the cells resistant to MTX (Fig 5B and S6 Fig). To extend this observation, we constructed a hDHFRL1 construct containing the putative loss-of-resistance allele, R25W, in a DFR1/dfr1Δ heterozygote strain (Fig 5B and S6 Fig). Although this mutant DFR1/hDHFRL1 (R25W) has a comparable growth rate to its wild-type counterpart DFR1/hDHFRL1, MTX resistance was abolished by introducing the R25W allele. Conversely, we tested the growth fitness of a hDHFR construct bearing the same mutation W25R (equivalent to W29R in yeast) in a DFR1/dfr1Δ heterozygote strain (Methods). This human W25R variant was reproducibly MTX resistant (Fig 5B). Of note, the W29R mutant in a yeast construct did not yield MTX resistance in our validation assay (Fig 3B), likely due to the weak ability of W29R variant to persist and modulate MTX resistance in the pool over the timecourse (Fig 2 and S3 Fig). Future work will address the mechanistic differences between the various MTX-resistant dfr1 alleles and their implications for folate metabolism.
Here we report a combined experimental and statistical approach capable of rapidly achieving high coverage of a targeted region to reliably identify bona fide drug resistant variants of dihydrofolate reductase. As a unique result of this strategy, we achieved increased throughput, resolution and sensitivity critical to detect dfr1 point mutation alleles that emerge or persist upon exposure to lethal MTX conditions over time. Although we cannot demonstrate we have exhausted all possible resistance conferring dfr1 mutations in the original dfr1 variomics library, our innovative approach proves to advance our understanding of the molecular basis of MTX resistance with the identification of a significant fraction of the resistance dfr1 variant space previously unknown in S. cerevisiae. By performing parallel competitive screening on diploid and haploid dfr1 libraries, we also uncovered pre-existing dfr1 hypomorphic alleles in the diploid state, which are as likely to modulate MTX resistance as haploid dfr1 mutations with dominant phenotypes. These observations further validate the relevance of interrogating gene variants in the diploid background given that in a human therapeutic target, mutations in only one allele suffice to provide drug resistance.
Our variant calling algorithm, RVD2, is specifically designed to call rare variants in pooled sequencing data. It does so by leveraging replicates of a sample to estimate a baseline error rate at each locus. Then the test sample error rates are compared, in a hypothesis testing framework, to the locus-specific baseline error rate. In previous work, we compared the performance of RVD2 to state-of-the-art variant calling methods using in vitro mixtures of synthesized DNA fragments at defined fractions. These experiments showed that RVD2 has higher sensitivity and specificity for detecting rare variants and equivalent sensitivity and specificity for higher frequency (more than 10%) variants.
We also identified the emergence of rare variant alleles, at starting frequencies lower than 1% that are capable of conferring resistance to MTX over time. As variant diversity is lost upon lethal MTX selective pressure, many of these rare alleles do not persist in the pooled conditions suggesting that the presence of epistatic mutations that can affect the evolution trajectory of adaptive MTX resistant alleles [47].
In addition, several novel functional MTX-resistant dfr1 alleles that disrupt the conserved active-site residues were identified in vivo, providing additional genetic insights into the determinants of MTX resistance. Importantly, we show that the yeast-based assay used here is capable of interrogating functional homologs such as the human enzyme. Out of 42 Dfr1 residue changes identified in our screen, 14 (33%) conserved residues are known to display key interactions with antifolate compounds and/or NADPH cofactor, which can affect the potency and selectivity of antifolates [29,33,35,42,48,49]. The remaining 28 (67%) residues identified are novel sites capable of modulating MTX resistance. The catalytic function of such DHFR mutations remains to be explored.
With current ever-improving gene synthesis approaches, determining the consequences of non-coding SNPs will become tenable as will assessing the concomitant effects of causative dfr1 mutations when expressed in combination. As sequencing technology becomes cheaper and more practical, this platform should in principle be extensible to uncovering linked mutations in small drug targets like DHFR that can confer specific resistance in yeast to address the growing problem of resistance to otherwise effective compounds and FDA drugs.
The homozygous diploid reference strain BY4743 MATa (his3Δ1 leu2Δ0 LYS2 met15Δ0 ura3Δ0)/MATalpha (his3Δ1 leu2Δ0 lys2Δ0 MET15 ura3Δ0) was used to determine the MEC100 (minimum effective concentration to cause 100% cell killing) of methotrexate (MTX) to use for resistance screening and validation growth assays. For MTX resistance screens, a dfr1 variomics library was used [50]. MTX was purchased from Sigma (M9929) and single-use MTX aliquots were prepared by dissolving MTX in DMSO solvent to 100 mM and stored at -80°C. To counterselect cells carrying the plasmid-borne dfr1 mutations, 5-Fluoroorotic acid (5’-FOA; Sigma F6625) was added to the media at a final concentration of 1 g/L. To select DFR1/dfr1 cells with the correct dfr1 integration event, growth sensitivity to G418 sulphate (Geneticin) was verified by adding G418 (Fisher 142480) to a final concentration of 400 mg/L to YPD agar plates.
For assessing growth fitness in the presence of MTX, yeast strains and pools were cultured to mid-log phase (OD600 ~0.5) in synthetic complete (SC) liquid media before adjusting the cultures to an initial OD600 of 0.0625. Cells were then transferred to a 96-well microtiter plate containing liquid SC media with either MTX or DMSO solvent (2% v/v) as control. To determine the MEC100 dose of MTX, a range of doses (0.025, 0.05, 0.1, 0.3 and 2 mM) were tested against the wild-type BY4743 strain. Cell growth upon MTX treatment relative to DMSO solvent was assayed in three biological replicates using a spectrophotometer Tecan shaker-reader that measured OD595 values over 24 hours at 30°C. Cell growth was inhibited at ~100% at 2 mM, which was the determined lethal dose for the MTX resistance assay (S1 Fig). To confirm MTX resistance, the reconstructed yeast strains were cultured in rich YPD media at a sublethal dose (1 mM) of MTX. Cell growth upon MTX treatment relative to DMSO solvent was assayed in three biological replicates using a Tecan shaker-reader over 24 hours at 30°C. For each mutant, the percent of growth rate in MTX relative to DMSO was calculated and the average and standard error of three biological replicates reported in S4 and S5 Tables. Mutant strains that showed a reproducible growth in the presence of the drug were confirmed to be true MTX-resistant strains (Fig 3). Cell growth in obligate respiratory media was used as a proxy to assess mitochondrial folate metabolism. Dfr1 mutant cells with non-synonymous mutations were cultured in obligate respiratory growth media using YPG media prepared with a non-fermentable carbon source (glycerol at 3% v/v). Growth or lack of growth in YPG media was assayed in three independent assays. Cultures that displayed two doublings or fewer after 24 hours in YPG were scored as respiratory-defective. The respiratory-proficient strains DFR1/dfr1Δ and wild-type BY4742 were included, in parallel, as controls.
The starting dfr1 variomics library consists of at least 2 x 105 independent dfr1 variant alleles, with single and multiple point mutations per allele. The variomics library is cloned into a CEN-based plasmid under control of native upstream and downstream regulatory regions, and transformed into a DFR1/dfr1Δ heterozygote convertible diploid strain [50]. The diploid variomics pool is cultured in synthetic dropout medium lacking uracil (SD-URA) at 30°C to generate a working stock of OD600 1, equivalent to an average of 50-fold coverage (independent cells) for each of the 2 x 105 independent variants. The latter calculation assumes that each DFR1/dfr1Δ cell harbors one variant dfr1 allele and that all doubling times are similar. To generate a dfr1 pool in haploid MATa cells lacking the chromosomal wild-type DFR1 gene, the dfr1 diploid pool was sporulated and subsequently haploid converted using the previously described optimized procedures [16], with the following modifications. Diploid cells were cultured in 200 ml sporulation medium at room temperature with vigorous shaking (200 rpm) for 5 days in the dark to increase sporulation. ~ 2 x 105 sporulated cells, with an average 10-fold variant coverage were cultured in 400 ml haploid selection medium to enrich for MATa dfr1 G418R URA+ haploid cells at 30°C for 2 days [51]. Sporulation and haploid conversion recovered a dfr1 haploid pool with ~11% of the initial dfr1 alleles, which represents a total of 2.2 x 104 viable dfr1 alleles. The selection for haploid cells and genetic “bottle necks” introduced in the methodology (Fig 1 and S2 Fig) are likely to exert additional selective pressure on the haploid pool. Hence, we predict a smaller proportion of MTX-resistant variants that sustain cell viability is present in the haploid pool.
The DFR1/dfr1Δ diploid and dfr1Δ haploid pools were cultured to an initial OD600 of 0.01, with an average 10-fold variant coverage observed in the dfr1 variomics library. Each pool was screened in triplicate wells (n1, n2, n3 technical replicates) in 6-well microtiter plates over a 6-day time course in 10 ml of SD-URA media supplemented with MTX at 2 mM (MEC100) (Fig 1 and S2 Fig). Media supplemented with DMSO (2% v/v) was prepared in parallel as a negative control. The screening assay consists of 3 time points, where the MTX and DMSO treated pools were propagated at 30°C with vigorous shaking (200 rpm). Sampling was done every 2 days, which typically resulted in 8-generation propagation for the DMSO-treated pools (S2 Fig). The first replicate (n1) set of cultures for MTX and DMSO was used for propagating the subsequent time point: at each time point, MTX-treated cells from the n1 replicate well were diluted to OD600 0.01 with fresh SD-URA medium supplemented with either MTX or DMSO, and transferred to the equivalent 3 replicate wells in a new microtiter plate (S2 Fig). At each time point, MTX-treated cultures from all replicate wells were harvested for DFR1-targeted sequencing and analysis (see below). The initial dfr1 variomics libraries of both diploid and haploid pools (time point 0) were also split in three technical replicates to assess sample-to-sample variation. The sample size per condition (n = 3) and expected sequencing median read depth (20,000x) was selected based our published power analysis from a previous version of our statistical model [52] and experience with this version of the model [14] to detect a minimum variant allele frequency of 0.1%.
Plasmids were extracted from harvested MTX-treated cell pools at each time point using the DNA extraction protocol described previously [50]. PCR reactions were performed using Phusion High Fidelity polymerase, according to the manufacturer’s instructions (Thermo Fisher Scientific) with the following modifications. To amplify the dfr1 amplicons from the plasmid-containing extracts, PCR reactions were performed in 50 μl containing 100 ng of plasmid extract and universal plasmid-specific oligonucleotides at 0.5 μM [16]. The cycling protocol was as follows: 1× (98°C for 30 sec), 30× (98°C for 10 sec, 52°C for 30 sec, 72°C for 45 sec),1× (72°C for 5 min). For colony PCRs, a fraction of each yeast colony was picked using a plastic micropipet tip and placed at the bottom of the reaction tube containing 10 μl of 20 mM NaOH. Samples were boiled for 5 min and 1 μl of each sample was used for the PCR reactions in a total of 25 μl containing oligonucleotides (2.5 μM). For a complete list of oligonucleotides used, see S1 Table. The cycling protocol for colony PCR amplification was as follows: 1× (98°C for 30 sec), 30× (98°C for 10 sec, 48°C for 30 sec, 72°C for 10 sec),1× (72°C for 5 min). All reaction products were analyzed on a 1% (w/v) agarose gel.
Dfr1 amplicons prepared by PCR were first purified using the Thermo Fisher Scientific PCR purification kit, according to the manufacturer’s instructions, quantified using Qubit fluorometry (Life Technologies) and diluted for sequencing library preparation. Libraries were constructed using plexWell library kit technology (seqWell, Beverly MA). In this approach, each 1+ kb pool of diverse amplicons is tagged with a pool-specific barcode via a transposase-mediated adapter addition at random locations. After this tagging, the pools of amplicons are then pooled into a single meta-pool, and subjected to a second transposase-mediated adapter addition. Fragments of this pool containing sequence from each of the two iterative adapter additions are then amplified to yield a final sequence library representing identifiable fragments from each original amplicon pool. Sequencing data is available upon request. We have deposited the raw fastq files at the NCBI SRA under the accession number SRP072709.
DFR1-targeted sequencing data was collected and processed separately for the diploid and haploid experiments. To align the raw (fastq) sequencing data to the S. cerevisiae genome, we first trimmed the Illumina adaptor sequences using cutadapt (v 1.7.1) (--anywhere AGATCGGAAGAGC). Then the paired-end reads were aligned to the April 2011 UCSC S. cerevisae reference genome (sacCer3) using bwa (v 0.7.12) mem with the -M flag set. Finally, the resulting bam files were indexed and sorted for subsequent processing and visualization.
First, we used samtools (v 1.2) mpileup to generate pileup files for the DFR1 gene region chrXV:780367–782084. We also set the -A and -BQ0 flag to get high quality read depth estimates without discarding anomalous reads, and we set the maximum depth to 10,000,000 to ensure no truncation of read depth occurs. We used a custom program, described previously [14], to provide the count of each base pair at each position in the region of interest. Then, we ran RVD2 gibbs on the wild-type, T0, T1, T2, and T3 data sets separately with the default warm-up, thinning and sample size parameters to estimate the model parameters and latent variables in the RVD2 statistical model. Finally, we called variants between all pairs of data sets using RVD2 somatic test with an interval of [0.001, 100] and a significance level (α) of 0.05. The somatic test calls a provisional variant at a position if the Bayesian posterior probability (estimated from a sample size of 1000 from the model posterior distribution) that the difference between the VAF in two data sets (e.g. T0 and T1) is in the interval is greater than 1-α (two-sided). The provisional variant is called a variant if the distribution over the non-reference bases is non-uniform by a chi-squared test with a significance level of 0.05. Calls based on these posterior credible intervals were not adjusted for multiple comparisons and we did not detect any gross deviations from the assumptions of the statistical model for this data. Further details on the estimation procedures and hypothesis test are provided in the RVD2 study [14].
Our statistical model and variant calling method as described previously [14] is publicly and freely available at https://bitbucket.org/flahertylab/rvd.
Given a set of called variants, V, we compute the diversity as follows. To compute the diversity, we first compute the KL divergence from μ^j to p where p = 1/|V| where μ^j is the estimated non-reference read rate at called position j. The KL divergence is zero if and only if μ^j is equal to p. In that case, each variant is equally represented in the pool and the entire pool was made up of variant clones. Otherwise, the KL divergence is greater than zero. We compute the diversity as
D=∑j∈V1+tanhDKL(p‖μ^j).
The tanh function is -1 when DKL(p‖μ^j)=∞ and 0 when DKL(p‖μ^j)=0. So, 1+tanhDKL(p‖μ^j) is 0 when DKL(p‖μ^j)=∞ and 1 when DKL(p‖μ^j)=0. Summing over all of the called variants means that the maximum value of the diversity grows with the number of variants. Therefore, this diversity measure captures both the uniformity of the distribution of the variants as well as the total number of variants.
Gene fragments containing coding sequence point mutations flanked by DFR1 specific homology sequences were synthesized by IDT (S6 Table). Gene fragments containing wild-type yeast DFR1 and human DHFR and DHFRL1 sequences were included as controls. Each gene fragment was resuspended in water (molecular grade, Thermo Fisher Scientific) to make a 10 ng/μl stock and stored at -20°C. Prior to yeast transformation, the gene fragments were PCR amplified using Phusion High Fidelity, according to the manufacturer’s instructions (Thermo Fisher Scientific). For each PCR reaction, 10 ng of the IDT gene fragment was used as template in a 50 μl reaction and amplified with the oligonucleotides listed in S1 Table.
To confirm MTX resistance, each dfr1 point mutant fragment was integrated into the dfr1:: kanMX locus of the haploid and diploid progenitor strains, which harbours the dfr1 variomics library. The diploid DFR1/dfr1Δ progenitor strain was first outgrown in YPD containing 5’-FOA for 2 days at 30°C to counterselect for the Ura+ dfr1 plasmids prior to the transformation. The haploid dfr1 strain was propagated in SD-URA medium to maintain the plasmid-borne dfr1 pool in order to maintain its viability. A high efficiency transformation protocol was used to create the mutants by mitotic recombination [53]. The progenitor strains were first cultured to mid-log phase in liquid SD-URA media and subsequently transformed with the dfr1 fragments according to the standard heat shock protocol [53]. Human DHFR and DHFRL1 point mutations were also integrated into the dfr1::kanMX locus to generate dfr1 yeast hybrid strains. To confirm that the yeast transformants have the correct integration, both diploid and haploid clones were 1) confirmed for the appearance of PCR products of the expected size using oligonucleotides that span the upstream and downstream junctions of the dfr1:: kanMX locus (S1 Table); 2) confirmed for loss of G418 resistance; and 3) confirmed for MTX resistance using a sublethal dose of MTX in liquid growth assays (Fig 3). Additionally, the haploid clones were counter-selected in 5’-FOA containing YPD agar plates to kill any cells carrying the plasmid-borne dfr1 mutations and confirmed for the absence of plasmid-borne dfr1 PCR products by colony PCR using plasmid specific oligonucleotides (S1 Table). To make the yeast/human mutant constructs in the heterozygous diploid DFR1/dfr1 background, the haploid dfr1 mutants and wild-type BY4741 control were mated with the wild-type haploid BY4742 (MATalpha his3Δ1 leu2Δ0 lys2Δ0 MET15 ura3Δ0) using standardized yeast manipulation procedures [54]. The diploid constructs were confirmed by selectively growing in agar plates containing synthetic dropout medium that lacks lysine and methionine amino acids (SD-LYS-MET) for 2 days at 30°C.
The multiple sequence alignment for the dihydrofolate reductase protein was obtained with ClustalW [55] and the S4 Fig generated with ESPript [56]. The consensus sequence for all DHFR homologues (S5 Fig) was built using WebLogo [57]. The Saccharomyces cerevisiae DFR1 structural model was generated with Modeller [58] using the closest homologue of known structure (Candida glabrata DHFR, 54% identity, PDB ID: 3CSE) [41] as a template. The coordinates of NADPH were obtained by superimposing the structure of the C. glabrata DHFR structure in complex with NADPH (PDB ID: 3CSE), and the coordinates of methotrexate were obtained by superimposing the structure of the E. coli DHFR in complex with methotrexate (PDB ID: 4P66) [59]. The colour gradient for the sequence conservation was generated with ConSurf [60], using the aforementioned multiple sequence alignment. All structure figures were obtained with PyMol (Schrodinger, LLC).
|
10.1371/journal.ppat.1007836 | A broadly neutralizing germline-like human monoclonal antibody against dengue virus envelope domain III | Dengue is the most widespread vector-borne viral disease caused by dengue virus (DENV) for which there are no safe, effective drugs approved for clinical use. Here, by using sequential antigen panning of a yeast antibody library derived from healthy donors against the DENV envelop protein domain III (DIII) combined with depletion by an entry defective DIII mutant, we identified a cross-reactive human monoclonal antibody (mAb), m366.6, which bound with high affinity to DENV DIII from all four DENV serotypes. Immunogenetic analysis indicated that m366.6 is a germline-like mAb with very few somatic mutations from the closest VH and Vλ germline genes. Importantly, we demonstrated that it potently neutralized DENV both in vitro and in the mouse models of DENV infection without detectable antibody-dependent enhancement (ADE) effect. The epitope of m366.6 was mapped to the highly conserved regions on DIII, which may guide the design of effective dengue vaccine immunogens. Furthermore, as the first germline-like mAb derived from a naïve antibody library that could neutralize all four DENV serotypes, the m366.6 can be a tool for exploring mechanisms of DENV infection, and is a promising therapeutic candidate.
| Dengue virus infects 50–100 million people each year. Infection is initiated by entry of the virus into cells mediated by the viral envelope glycoproteins. There are four closely related DENV serotypes, but they all are antigenically distinct, with each comprising several genotypes that exhibit differences in their infection characteristics in both the mosquito vector and in the human host. One of the confounding problems that has faced vaccine and biological drugs development for decades is the inability of antibodies to one serotype to protect against infection by another one. Instead, the induced humoral immune response to one dengue virus infection can enhance the infection and disease processes brought by a subsequent infection with another dengue serotype. In this study, by using a competitive sorting strategy to interrogate a human naïve antibody library, we identified a cross-reactive mAb, designated as m366.6, against the four DENV serotypes. The mAb m366.6 possesses only few somatic mutations from the closest VH and Vλ germline genes and high affinity to DIII. Most importantly, the germline-like m366.6 demonstrated a broad spectrum of neutralization against the four DENV serotypes. Thus, m366.6 is a promising candidate therapeutics and its epitope may imply on the design of effective vaccine immunogens to elicit m366.6-like antibodies in vivo.
| Dengue virus (DENV) causes the most prevalent mosquito-borne viral disease. Over 2.5 billion people are at risk for infection in over 100 countries, 50–100 million are infected with symptoms, and up to 50,000 die from dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS) each year [1,2]. No specific antiviral drug has been available against DENV infection; the only approved vaccine, Dengvaxia, has caused considerable controversy regarding its safety and potential benefits [3–6]. For decades, anti-DENV vaccine and biological drugs development has been hampered by the high sequence divergence (25–40%) among the four DENV serotypes [7,8]. Such divergence leads to the fact that one antibody may not be sufficient to neutralize all DENV infection. Instead, the induced humoral immune response to one DENV infection can enhance the infection and disease processes brought by a subsequent infection with another DENV serotype [2–4]. These findings suggest that the development of new and broadly neutralization antibodies against all the serotypes of DENV could be promising candidate anti-DENV agents, and may also guide the design of effective and safe vaccine immunogens.
The DENV envelope glycoprotein (E protein), which mediates virus entry into cells, is the major neutralizing target of antibodies [9–13]. E protein is a type II fusion protein and consists of three domains: DI, DII, and DIII of which DIII has been proposed to contain a receptor binding domain [14–17]. Recent studies revealed that cross-reactive conserved epitopes exist on DII as well as DIII of the DENV E protein [14,16–18]. During the naturally-occurring primary DENV infection, a large fraction of the antibody repertoire consists of DII-specific antibodies which are, unfortunately, typically poor in neutralization and may increase the likelihood of severe disease upon subsequent infection through a mechanism known as antibody-dependent enhancement (ADE) [18–20]. In contrast, antibodies targeting DIII have proven to be the most potent neutralizing antibodies, but very few could be elicited in naturally infected individuals [18,19,21–35]. Despite this, previous studies indicated that anti-DENV DIII serotype-specific and cross-reactive antibodies could be elicited using DENV DIII as vaccine immunogen [36–43] and in infected humans [44–47]. It has also been demonstrated that the lysine at position 310 on DIII is the critical residue in the cross-reactive epitope [24]. Therefore, the conserved epitope on DIII represents an attractive target for the development of broadly neutralizing DENV antibodies.
Here, we report the isolation of a potent DENV DIII-specific human monoclonal antibody (mAb), designated as m366.6, from a large naïve antibody library constructed by the blood of healthy adult donors. A competitive sorting strategy using a DIII mutant as competitor was applied to identify antibodies precisely targeting the conserved neutralizing epitope. To our knowledge, m366.6 is the first human mAb isolated from a naïve antibody library which could neutralize all the four serotypes DENV viruses. Importantly, both heavy and light chain genes of m366.6 are very close to their putative germline predecessors. Its fully human origin, the germline-like nature, combined with high-affinity and broad neutralizing activity toward all DENV serotypes, suggest that m366.6 is a promising candidate antiviral agent and may also provide a unique template for designing effective dengue vaccine immunogens.
We previously prepared some large naïve antibody libraries using peripheral blood B lymphocytes of non-immunized healthy donors and used them for panning/screening against viral and cancer targets [48–55]. In this study, we used a competitive library sorting strategy to isolate broadly neutralizing antibodies against DENV1-4 (Fig 1A and 1B). The yeast-displayed naïve single chain antibody fragment (scFv) library was used to screen against the biotinylated DENV DIII, and, importantly, ten times concentration unbiotinylated DIII K310E mutant was used as the competitor. The yeast cells were selected to present the antibody-expressing cells that could bind well to the wild-type DIII instead of the DIII mutant, resulting in the isolation of antibodies that can target the cross-reactive neutralizing epitopes covering the residue Lysine310 [55]. Potent enrichment was achieved after four rounds of sorting, and a panel of antibodies were identified (Fig 1B). Two antibodies, designated as m360 and m366, bound potently to DENV DIIIs. Their scFv gene were fused with human IgG1 Fc for protein expression, and surface plasmon resonance (SPR) experiments were used to evaluate the antigens binding. The equilibrium dissociation constant (KD) of m360 for the DENV1-4 DIIIs were 5.8 nM, 5.1 nM, 0.1 nM and 8.3 nM, respectively. The mAb m366 displayed a broader binding profile compared with that of m360, with the KD of 3.3 nM, 1.2 nM, 1.1 nM and 12 nM to DENV1-4, respectively (Table 1, S1 and S2 Figs).
To further improve the affinity of m360 and m366 with the four DENV serotypes, we constructed a mutant library using the error-prone PCR technologies. Following three cycles of mutagenesis and selection, two clones were identified from the enriched pool of yeast sorting, designated as m360.6 and m366.6. Biacore analysis showed that the cross-reactive binding activities of m360.6 and m366.6 to all 4 DIIIs were preserved after the affinity maturation process. The KD of m360.6 for the DENV1-4 DIIIs were 0.3 nM, 42 pM, 2.3 pM and 33 nM, respectively (S3 Fig). Although the binding to DENV1-3 DIIIs was improved, the m360.6 had only slightly increased binding affinity to DENV4 DIII compared to its parental mAb m360. Notably, the m366.6 exhibited high affinity to all the DENV DIIIs. The KD of m366.6 for the DENV1-4 were 0.8 nM, 0.3 nM, 0.3 nM, and 1.9 nM respectively, which demonstrated that m366.6 could bind to all the four serotype DENV viruses with high avidity (Table 1, Fig 2). We also assessed the binding specificity of m366.6 by ELISA, and the results showed that m366.6 had weak cross-reactivity with Zika virus (ZIKV) DIII and no binding with other irrelevant antigens (S4 Fig).
Next, we assessed the neutralization capacity of m360.6 and m366.6 against the four DENV serotypes using a DENV luciferase reporter viral particle (RVP) neutralization assay. We used DENV RVPs against the four dengue serotypes that are common strains in DENV research: DENV-1 (WestPac 74), DENV-2 (S16803), DENV-3 (CH53489), and DENV-4 (TVP360). The luminescent reporter expression was proportional to the number of RVPs added to BHK DC-SIGN cells, confirming the linear correlation between the extent of RVP infection and reporter gene expression. In consistent with the Biacore binding results, both m360.6 and m366.6 could neutralize all the four serotype DENV, and m366.6 displayed better neutralization than m360.6, with the 50% neutralization titers (IC50) of 22, 2.4, 0.85, and 0.36 μg/ml against DENV1-4 respectively (S5 Fig).
To further evaluate the neutralization breadth of m366.6 IgG against the four DENV serotypes, a standard plaque reduction neutralization assay (PRNT) on BHK-21 cells was performed using DENV1-4 live viruses, including DENV-1 128 (GenBank FJ176780), DENV-1 GZ01/2017 (S6 Fig, isolated from a DENV-1 infected patient in Guangzhou, China), DENV-2 43 (GenBank AF204178), DENV-3 80–2 (GenBank AF317645), and DENV-4 B5 (GenBank AF289029). An irrelevant human mAb G12 was used as the negative control [56], and 2A10G6, a broadly neutralizing mAb against all the four DENV serotypes, was used as the positive control [57,58]. As shown in Fig 3, m366.6 IgG could neutralize all the four DENV serotypes. The 50% neutralization titers (IC50) of m366.6 against DENV1-4 was 12.7, 4.57, 5.23, and 23.31 μg/ml respectively (Table 2).
We next used a well-established ADE assay to detect the in vitro ADE effect of m366.6 IgG. A mutated form of m366.6 IgG was also generated containing the leucine to alanine substitutions at positions 234 and 235 (m366.6 IgG-LALA), which lacked binding to Fcγ receptors. The ADE effects of DENV-1 or DENV-2 by m366.6 IgG, m366.6 IgG-LALA, as well as 2A10G6 were measured. Interestingly, neither m366.6 IgG nor m366.6 IgG-LALA presented any ADE effect against different serotypes of DENV (Fig 3F, S7 Fig). In contrast, potent ADE effects were observed for the DII-specific mAb 2A10G6. These results showed that m366.6 IgG is a DENV DIII-specific mAb without detectable ADE effect.
We further analyzed the sequences of mAbs using the IMGT tool to identify their closest VH and Vλ germline genes. The results indicated that m360.6 and m366.6 originated from different B-cell lineages (Table 3). The m360.6 VH gene was derived from the IGHV2-70 and the Vλ gene was from IGLV1-51. In contrast, the m366.6 VH gene was derived from the IGHV3-21 and the Vλ gene was from IGLV3-21. Interestingly, we found that the encoding genes of both m360.6 and m366.6 closely resembled their corresponding germline gene segments. Notably, m366.6 VH and Vλ gene shared 95.8% and 95.2% sequence identities with the IGHV3-21*01 and IGLV3-21*01 germlines respectively (Fig 4A and 4B). These results indicated that the mAb m366.6 is a germline-like antibody, which, in general, could show better drug properties and lower immunogenicity compared to somatically hypermutated antibodies [59].
To further investigate the immunogenetic characteristics of m366.6-like antibodies, we analyzed in detail the IGHV3-21 recombination frequencies with specific IGHD and IGHJ genes families from naïve immunoglobulin M (IgM) repertoires of 33 health adult donors and neonatal IgM repertoires of 10 newborn babies, using next-generation sequencing data previously generated from our antibodyome studies [48]. By querying the m366.6 sequence from the IgM repertoires, 39 sequences were found to display m366.6-like V(D)J recombination from the genes IGHV-3-21, IGHD1, and IGHJ3 out of a total of 10,498,301 sequences from healthy adult IgM repertoires. In IgM repertoires of newborn babies, a similar recombination frequency was also observed, in which 111 sequences with m366.6-like V(D)J recombination were found from 5,617,227 sequences. Our analysis showed that IGHV3-21 is one of the most frequently used IGHV genes, and identified that many of those sequences sharing a significant degree of resemblance to m366.6 (Fig 4C). In brief, analysis of these data showed the potential of eliciting robust immune responses with the m366.6-like germline antibodies by vaccination.
To determine whether m366.6 can protect DENV infections in vivo, we firstly used a lethal DENV1-4 infection suckling mouse model. The mice were challenged with DENV1-4 at 200 PFU/mouse via intracranial injection. Four hours later, the mice were treated intracranial with a single dose (100 μg) of m366.6 IgG, m366.6 IgG-LALA mutant and G12 (unrelated antibody control). These animals were monitored for morbidity and mortality daily. As shown in Fig 5, all the mice in control groups died from DENV infection, and most of them died within the first two weeks of viral challenge. Interestingly, there was no significant difference in therapeutic efficacy against DENV1-4 infection between m366.6 and the LALA-mutated m366.6. The m366.6 IgG protected 100% DENV-1, DENV-3, DENV-4 and 83% DENV-2 infection whereas LALA-mutated m366.6 protected 83% DENV-1, DENV-4 and 67% DENV-2, DENV-4 respectively. Therefore, m366.6 has no detectable ADE as confirmed in both in vitro and in vivo experiments. We also used the AG129 (types-I and -II IFN receptor deficient) mice to test the therapeutic effect of m366.6 against DENV-2 (S8 Fig). The results showed that all the mice in the control antibody treatment group died while the survival rate of mice can reach 67% in m366.6 treatment group, indicating that the antibody can also protect the lethal infection of DENV-2 in AG129 mice. Taken together, these results indicated that m366.6 can protect DENV1-4 infections in vivo.
To map the epitope of the germline-like mAb m366.6 and identify in greater detail the structural basis of DENV neutralization, we employed multiple approaches (Fig 6).
Sequence alignment of different DENV genotypes and mapping of the conserved amino acid residues of DENV DIII showed that four serotypes DENV DIIIs amino acid residues were different from one another between amino acids 300–393 (Fig 6A). Subsequently, serotype 2 derived DIII consensus gene was randomly mutated to construct a yeast-displayed mutant library. Two rounds of sorting of those yeast cells showing expression on surface but lacking the binding to m366.6 was performed. A total of 35 binding escape mutants were aligned with the serotype 2 consensus protein sequence. Mutation frequency at each position was plotted against the residue position number. Similarly, 193 unique DIII sequences derived from naturally isolated serotype 2 dengue viruses from GeneBank were also aligned with the consensus sequence (Fig 6B and 6C). The superimposed profiles of the two set of sequences showed that many of the escaped mutations located in the well-conserved area, indicating the broad cross-reactivity of m366.6 to naturally isolated dengue viruses. Besides, the epitope mapping shows that the m366.6 epitope is at close to or partially overlaps the dimerization interface between domains II and III. These results may explain why m366.6 is a potent cross-reactivity antibody to all the four DENV serotypes.
Furthermore, computational docking of DENV DIII-m366.6 antibody complex was performed using ZDOCK method. We selected the three top scored docked complexes that contained the key residues identified from an experimental epitope mapping approach. One of the top scored docked models exhibited minimum clashes with appropriate protein interface parameters and was used to demonstrate the lactation the potential epitopes and their interactions with m366.6 antibody, which might shed light on the molecular mechanisms of broadly cross-reactive neutralization. Fig 6D showed the docking model of the DIII-m366.6 antibody complex in which these epitopes are highlighted in green surfaces. The docking model revealed a different orientation of antibody binding as compared to the DIII complex structure with Fab 1A1D-2 that was previously determined [35]. The epitope comprised of three structurally proximal regions, residues 305–311 in green, 325, 327 and 361 in dark green, and 383–385 at the C-terminal in lime. One of the key residues, K310, contacts the CDR-L1 of m366.6 which has a germline mutation. In Env-DIII-Fab-1A1D-2 complex structure, the residue K310 contacts the CDR-H1. The hydrophobic residues, Ile and Trp, of CDR-H3 contact the center part of the epitope, and other loops H1, H2, L2 and L3 also involve in the binding. The surface area of the interface between DIII and m366.6 antibody in the model complex is 716 Å2, a typical of antibody-antigen interactions. There are six hydrogen bonds likely to form and no salt bridges at the interface. In brief, the binding regions of the m366.6 may be close to or partially overlaps the dimerization interface between domains II and III, which might indicate the broad cross-reactivity of m366.6 to the four serotypes of DENV.
Dengue is a disease with a complex immune response orchestrated by host cells partially due to the presence of four serotypes of DENV. Importantly, after a primary DENV infection, one can be protected against or aggravate of a secondary infection with a different serotype, which bring many difficulties to develop an effective vaccine. Thus, it is very urgent to develop an effective and cross-reactive antiviral therapy against DENV infection.
Monoclonal antibodies (mAbs) are of growing importance for protective and pathogenic immune responses to viruses. At present, there are many therapeutic antibodies to treat viral infections under development, such as antibodies for HIV-1, SARS-CoV, MERS-CoV, Nipah and Hendra viruses, and H7N9 influenza virus [48,50,60–66]. Fortunately, screening antibodies from the large naïve libraries has enabled the rapid development of high-affinity human mAbs, especially for the rapid response to the outbreak of emerging viruses and diseases. We recently successfully identified two human germline-like mAbs against MERS-CoV and H7N9 influenza virus from the naïve library, named m336 and m826, respectively [48,50]. They both can naturally exist with very low level of somatic hypermutation in the naïve library with which they have potent binding activity against the envelop proteins of MERS-CoV and H7N9 influenza virus. Most importantly, m336 and m826 all showed highly therapeutic effective in the animal models. Therefore, the naïve library screening can be quickly used to isolate germline-like antibodies that effectively bind to complex protein targets like those in DENV viruses.
How to increase the neutralization breadth is a key issue in developing anti-DENV antibodies. Previous studies revealed two classes of broadly neutralizing antibodies to flaviviruses, including antibodies targeting the conserved epitopes in DII or DIII [16–18,20]. While the conserved fusion loop epitope (FLE) in DII is the immunodominant epitope in E protein, unfortunately, this epitope frequently induced poorly neutralizing and strongly infection-enhancing antibodies via ADE [18–20]. Therefore, DIII represents the ideal target for neutralizing antibodies. In this study, we applied a highly efficient yeast-display-based sorting strategy by using the highly diverse DENV DIIIs as antigen and the competitive sorting technique. By applying this method, we quickly and efficiently identified two human germline-like broad-spectrum anti-DENV mAbs (m360 and m366) from the naïve scFv yeast library using the DIII antigen that make them as promising candidate therapeutics as well as the template for vaccine development. Another class of highly efficient broadly neutralizing antibodies that target the envelope dimer epitopes (EDE) from the secondary acute DENV infection plasmablasts has been identified by Dejnirattisai et al. [67]. These antibodies may especially get through with high somatic mutations from the secondary virus infection. Compared with the highly somatically mutated antibodies, germline-like antibodies typically have better safety and drug-related property [59]. Importantly, the Hendra and Nipah antibody m102.4 is a near germline antibody and exhibited a very good drugability, which was from a similar library that was also used to isolate our m366 and m366-like antibodies. m102.4 was successful as a candidate therapeutic mAb in animal models and was also completed the phase I clinical trial (ACTRN12615000395538) without side effects [64]. To further improve the affinity of m366 with the four serotypes DENV DIIIs, we subjected m366 to affinity maturation process, and named it as m366.6. Subsequently, we analyzed m366.6 sequence using the IMGT tool to identify its closest VH and Vλ germline genes. Interestingly, we found that m366.6 is still a germline-like antibody although it went through the mutation process in vitro, with over 95% identities of its VH and Vλ genes to the IGHV3-21*01 and IGLV3-21*01 germline respectively.
In order to evaluate the neutralization effect of the m366.6 IgG, we used a standard plaque reduction neutralization with BHK21 cells to measure DENV infection and neutralization. The m366.6 IgG showed broadly neutralization towards the four serotypes DENV as well as a recent DENV isolate from clinical samples. More importantly, m366.6 did not present any ADE effects in different serotypes of DENV. The in vivo study results demonstrated the therapeutic potential of m366.6 against severe DENV1-4 infections. In brief, the m366.6 could neutralize the four serotypes DENV in vitro and protect the DENV infection mouse model in vivo without detectable ADE effects. We therefore expect that m366.6 has a likeness drugability of m102.4 and could be developed as a candidate therapeutic in the future.
We have also localized the m366.6 epitope by using a combination of computational structural modeling, display-based antigen mutagenesis, and sequence-based analysis of mutants. The epitope appears to overlap with the epitope previously explored as targets for cross-reactive murine mAbs and close to or partially overlaps the dimerization interface between domains II and III. This further indicates that this epitope could be an important component of vaccine immunogens intended to elicit cross-reactive neutralizing antibodies. In progress are our experiments to crystallize the complex of m366.6 with DENV DIII that would allow precise determination of the m366.6 epitope.
The major result of this study is the identification of a germline-like human mAb, m366.6, from a naïve yeast antibody library which binds with high (picomolar) affinity to DIIIs from all serotypes and neutralizes the four DENV serotypes. There are two major implications from this finding: 1) m366.6 is a potential candidate therapeutic which could be further developed in preclinical and clinical settings. 2) the epitope of the germline-like mAb m366.6 could guide the design of effective candidate vaccine immunogens capable of eliciting m366.6 and/or m366.6-like antibodies.
BHK21 cells were cultivated in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% fetal bovine serum (FBS) (Biowest). Mosquito cells C6/36 were cultured in RPMI-1640 medium supplemented with 10% FBS. All cells were maintained in a humidified atmosphere of 5% CO2 at 37°C incubator, except for C6/36 cells, which were cultivated at 28°C. DENV-1 128 (GenBank FJ176780), DENV-1 GZ01/2017 (isolated from DENV-1 infected patient in Guangzhou), DENV-2 43 (GenBank AF204178), DENV-3 80–2 (GenBank AF317645), and DENV-4 B5 (GenBank AF289029) were propagated in C6/36 cells by using RPMI 1640 medium and the titers were measured by standard plaque forming assay in BHK21 cells.
DENV DIII genes from all 4 serotypes were synthesized by Genescript, Inc (Nanjing, China), fused with IgG1 Fc and a C-terminal Avi-tag, and cloned into pSecTag expression vector. The DIII.3 (serotype 3) K310E mutant was generated through overlapping PCR. For the conversion of IgG1 from scFv, the heavy and light chains of scFv were amplified and recloned into the PTT-IgG1 vector. The plasmids were transfected into Expi293 cells (Thermo Fisher) for transient expression, and purified using protein G column (GE Healthcare, Piscataway, NJ) according to the manufacturer’s instructions. The purified protein was biotinylated by mixing with biotinylation reagents in PBS for 30 min on ice, according to the manufacturer’s instructions (Pierce).
A large yeast-displayed scFv library was used for antibody screening, and the screening protocols were essentially carried out as described previously [55]. Briefly, 10 μg of binotinylated DIII.3-Fc and 1010 cells of the initial naïve library were mixed and washed by PBSA, and incubated with 100 μl streptavidin conjugated microbeads (Miltenyi Biotec, Auburn, CA) before loading onto the autoMACS system (Miltenyi Biotec) for sorting. After three rounds of sorting, the downsized library was further sorted against binotinylated DIII.3-Fc (1 μg/ml) but also using unbiotinylated K310E mutant (1 μg/ml) as the competitor. The cells were stained by the addition of mouse anti-c-Myc antibody (Roche), Alexa-488 conjugated goat-anti-mouse antibody (Invitrogen), and PE-conjugated streptavidin (Invitrogen) for sorting on a FACSAria II cell sorter (BD Biocsiences, San Jose, CA) to isolate the positive binders. The plasmids of the positive clones were prepared by using Zyppy Plasmid Miniprep Kit according to the manufacturer’s instructions (Zymo Research).
To generate the m360 scFv and m366 scFv mutant libraries, random mutagenesis of the scFv genes were performed through error-prone PCR by using a GeneMorph II kit (Stratagene) following the manufacturer’s instructions with minor modifications. To further diversify the mutation profile, 3 uM of each of the two nucleotide analogues (8-oxo-deoxyguanosine triphosphate and 2'-deoxy-p-nucleoside-5'-triphosphate) was mixed in the PCR reaction mixture. For the second and third cycle library constructions, an extra step of DNA shuffling PCR was inserted into the regular PCR cycles to combine the beneficial mutations obtained from previous maturation process. DNA shuffling PCR step was performed as following: 20 cycles of denature at 94 °C for 15 seconds followed by annealing/extension at 68°C for 1 second on the Bio-Rad MyCycler.
Binding affinities of m360 scFv, m366 scFv, m360.6 scFv, and m366.6 scFv to the 4 DENV DIIIs were analyzed by surface plasmon resonance technology using a Biacore X100 instrument (GE healthcare). The antibodies were covalently immobilized onto a sensor chip (CM5) using carbodiimide coupling chemistry. A control reference surface was prepared for nonspecific binding and refractive index changes. For analysis of the kinetics of interactions, varying concentrations of antigens were injected at flow rate of 30 μl/min using running buffer containing 10mM HEPES, 150 mM NaCl, 3 mM EDTA, and 0.05% Surfactant P-20 (pH 7.4). The association and dissociation phase data were fitted simultaneously to a 1:1 Langumir global model by using the nonlinear data analysis program BIAevaluation 3.2. All the experiments were done at 25°C.
Neutralizing activity of mAbs was measured using a standard plaque reduction neutralization with BHK21 cells as previously described [57]. Briefly, 5-fold serial dilutions of mAbs were added to approximately 200 PFU of a variety of dengue virus strains and incubated for 1 h at 37°C. Then, the mixture was added to BHK21 cell monolayers in a 12-well plate in duplicate and incubated for 1 h at 37 °C. The mixture was removed, and 1 ml of 1.0% (w/v) LMP agarose (Promega) in DMEM plus 4% (v/v) FBS was layered onto the infected cells. After further incubation at 37 °C for 4 days, the wells were stained with 1% (w/v) crystal violet dissolved in 4% (v/v) formaldehyde to visualize the plaques. PRNT50 values were determined using non-linear regression analysis. PRNT50 data were calculated by doing a non-linear regression analysis using Sigmaplot (Version 9.01, Systat Software, Inc., CA) as previously described [57].
DENV RVPs from all four serotypes were pre-incubated with an equal volume of serially diluted antibodies (25 μg/ml to 0.0012 μg/ml pre-dilution or 12.5 μg/ml to 0.0006 μg/ml pre-dilution, as measured based on the dilution of antibody prior to combining with RVPs) in DMEM infection media for 1 h at room temperature and transferred to wells of a 96-well plate. An equal volume of DENV RVPs were added to each well followed by slow agitation for 1 h at room temperature. BHK DC-SIGN cells were added to each well at a density of 30,000 cells per well followed by incubation at 37°C in 5% CO2 for 48 h. Cells were subsequently fixed in lysed and analyzed for luminescent reporter expression using the Wallac Victor. The percent infection for each concentration of mAb or serum was calculated, and the raw data was expressed as percent infection versus log10 of the mAb concentration or the reciprocal serum dilution. The data were fit to a sigmoidal dose-response curve using Prism (GraphPad Software, La Jolla, CA) to determine the titer of antibody that achieved a 50% reduction in infection. Maximum infection was determined in the absence of antibodies.
The in vitro ADE assay was performed using K562 cells [57]. Briefly, serial 10-fold dilutions of antibodies under concentrations ranging from 100 to 0.01 μg/ml were mixed with DENV-1 or DENV-2, and incubated for 1 h at 37 °C. Mixtures were then added to 2×105 K562 cells at multiplicity of infection of 0.1~0.25 for 2 h in 24-well plates. The cells were subsequently washed 3 times with serum free RPMI-1640 medium. After collecting cells by centrifugation, the cell pellets were re-suspended with RPMI-1640 medium containing 2% FBS and added to 24-well plates, then incubated for 4 days at 37 °C with 5% CO2. The titer of viruses in the supernatant was then measured using a plaque assay. The ADE effect was calculated as different viral yields in the supernatant after infection in the presence of the added antibodies.
The epitope mapping of m366.6 was performed using previously described protocols [55]. Briefly, random mutagenesis of the DENV DIII.2 gene was performed using a GeneMorph II kit (Stratagene). As described above, the yeast-displayed mutant library was mixed with biotinylated m366.6 scFv-Fc, washed, and stained by mouse anti-c-Myc antibody (Roche), Alexa-488 conjugated goat-anti-mouse antibody (Invitrogen), and PE-conjugated streptavidin (Invitrogen). After two rounds of sorting on a FACSAria II cell sorter (BD Biocsiences, San Jose, CA), the sorted cells were amplified and their plasmids were prepared and sequenced.
Homology modeling of variable regions of heavy (VH) and light (VL) chains for m366.6 scFv antibody was carried out using the SWISS-MODEL workspace [68] by selecting the closest template structures (PDB codes: 3QOS for heavy chain and 2DD8 for light chain), whose sequence similarities were 92% and 87% respectively. The VH-VL orientation of m366.6 scFv structure was assigned similar with one of the templates (PDB code: 2DD8) that showed minimum steric clash for creating the final m336.6 scFv model. The crystal structure of DENV DIII serotype 2 (PDB code: 2R29) was used for docking with the modeled scFv antibody m366.6. Docking of scFv m336.6 to the dengue Env-III was performed by ZDOCK server (http://zdock.bu.edu) that uses a fast Fourier transform (FFT)-based rigid-body protein docking algorithm with scoring functions combining pairwise shape complementarity, desolvation and electrostatic energies. Based on the escape mutants that led to the loss of epitopes and available crystal structure of DENV DIII, we selected a list of residues as biological constrains, 307, 309, 310, 311, 327, 361 and 383, on the surface of Env-DIII as potential contacting residues for docking constraints. Similarly, one or two residues from each of CDR-H1, H3 and L3 loops were chosen at the docking interface. CDR-H1 and H3 loops had dominant hydrophobic residues whereas CDR-L1 had a germline mutation, and they all had high antigen-contacting propensities [69]. Results from the top 2000 ZDOCK predictions were filtered using the user-defined residues and a 6 angstrom distance cutoff. Three predicted complexes were only kept as all residues selected come together at the interface and were further examined by PDBePISA (Protein Interfaces, Surfaces and Assemblies). PyMOL was used for the analysis of docked model and graphical illustration [70].
The suckling mice were purchased from B&K Universal Group Limited (Shanghai, China) and housed under specific pathogen-free conditions at the animal facilities of the Shanghai Public Health Clinical Center, Fudan University (Shanghai, China). Before infection, the mice were transferred to the Animal Biosafety Level 2 (BSL-2) Laboratory (Shanghai, China). One day mice were used for all experiments. All mice were intracerebrally injected with 200 PFU of DENV1-4. At 4 h post infection, mice were passively transferred a single dose of 100 μg antibody m366.6 IgG, m366.6 IgG LALA mutant or G12 IgG as the negative control via intracerebrally injection. Survival rates and disease sings were monitored daily. The AG129 mice (type I and type II interferon receptor-deficient) were purchased from B&K Universal Group Limited (Shanghai, China) and housed under specific pathogen-free conditions at the animal facilities of the Shanghai Public Health Clinical Center, Fudan University (Shanghai, China). Before infection, the mice were transferred to the BSL-2 Laboratory (Shanghai, China). Groups of mixed-sex 4- to 6-week-old mice were used for all experiments. All mice were intraperitoneally injected with 2x106 PFU of DENV-2 in a volume of 200 μL. At 16 h post infection, mice were passively transferred a single dose of 500 μg antibody m366.6 IgG-LALA, or G12 antibody as the control via i.p. injection. Survival rates, weight loss, and disease sings were monitored daily.
Specific-pathogen-free AG129 mice (4–6 weeks old) and suckling mice were used for all experiments. All experimental protocols were reviewed and approved by the institutional committee of Fudan University (Permit Number: 2018-A056-02) in accordance with the Guideline for Ethical Review of Animal Welfare (GB/T 35892–2018) of the Chinese National Health and Medical Research Council (NHMRC).
|
10.1371/journal.pgen.1004432 | Genetic Background Drives Transcriptional Variation in Human Induced Pluripotent Stem Cells | Human iPS cells have been generated using a diverse range of tissues from a variety of donors using different reprogramming vectors. However, these cell lines are heterogeneous, which presents a limitation for their use in disease modeling and personalized medicine. To explore the basis of this heterogeneity we generated 25 iPS cell lines under normalised conditions from the same set of somatic tissues across a number of donors. RNA-seq data sets from each cell line were compared to identify the majority contributors to transcriptional heterogeneity. We found that genetic differences between individual donors were the major cause of transcriptional variation between lines. In contrast, residual signatures from the somatic cell of origin, so called epigenetic memory, contributed relatively little to transcriptional variation. Thus, underlying genetic background variation is responsible for most heterogeneity between human iPS cell lines. We conclude that epigenetic effects in hIPSCs are minimal, and that hIPSCs are a stable, robust and powerful platform for large-scale studies of the function of genetic differences between individuals. Our data also suggest that future studies using hIPSCs as a model system should focus most effort on collection of large numbers of donors, rather than generating large numbers of lines from the same donor.
| Human induced pluripotent stem (hiPS) cells are a potentially powerful model system for studying human disease and development, and a resource for personalized medicine. However, it has been reported that hiPS cells exhibit substantial heterogeneity which could limit their use as model systems. Clearly, knowledge of the source of heterogeneity is key for deeper understanding of the use of human iPS cells for basic and therapeutic applications. One source of this heterogeneity has been presumed to be “memory” of the adult somatic cell from which the hIPS cells were derived, but the evidence to support this view is scant. We have generated a set of human iPS cells from a set of somatic cell types from different donors. Our study shows that cell lines from different somatic sources but from the same donor (i.e. with the same genome) are more similar than cell lines isolated from the same tissue type but from different donors. Once genetic changes are accounted for, all aspects of gene expression, including mRNA levels, splicing and imprinting are highly similar between iPS cells derived from different human tissues. Thus, most of the previously described transcriptional variation between cell lines is likely to be genetic in origin.
| Induced pluripotent stem cells (iPSCs) are the subject of tremendous interest as model systems for studying human disease and development [1], [2]. However, cellular reprogramming to iPSCs is an inefficient process in which stochastic events during clonal selection may fix a variety of alternative epigenetic and transcriptional states[3]. Some reports have described significant variation between iPS cells and ES cells, while others have suggested that iPS cells retain a memory of the somatic tissue from which they were derived that may negatively affect their differentiation efficiency into certain cell lineages [4]–[10]. However, comparisons between human iPS cells and ES cells are confounded with differences in genetic background because the lines are derived from different donors. Likewise, because collection of multiple primary tissues from the same individual is frequently impractical, studies of cellular memory in hiPS cells have often confounded iPS source tissue type and donor genetic background. This is important because many cellular phenotypes, including transcription and methylation, are substantially impacted by genetic differences between individuals [11]–[13].
In this study we set out to understand the basis of this variation by establishing a set of iPS cells from a panel of tissues isolated in parallel from several different donors. RNA-seq data sets from these lines, the corresponding adult somatic cells and human ES cells have been systematically compared. This has enabled us to investigate patterns of expression, splicing and imprinting between these iPS cells, their adult cell progenitors and compare these with hES cells. Using a statistical model we estimated the relative contributions of genetic background and tissue of origin to transcriptional variability between human iPS cell lines.
We established primary fibroblast, keratinocyte and endothelial progenitor cell (EPC) somatic cell lines from three healthy male organ donors, labeled S2, S5 and S7, and one healthy female donor (S4). From each primary adult tissue cell line, we derived at least three independent iPS cell lines for each donor. For the adult cell cultures we extracted RNA following each of three passages to give a total of 18 RNA samples from adult donor cells (6 fibroblast, 3 keratinocyte and 9 EPCs). We also extracted RNA from the iPS cell lines derived from each of these tissues to give a total of 9 RNA samples from fibroblast-derived iPSCs (F-iPSCs), 6 from keratinocyte-derived iPSCs (K-iPSCs) and 10 EPC-derived iPSCs (E-iPSCs). Finally, we also extracted 4 RNA samples from two ES cell lines, H9 and Val9. Throughout our study highly standardized conditions were used. This included the isolation, derivation and culture of primary cell cells as well as use of the same batches of reprogramming viruses, serum and media. Each of the primary cell lines were reprogrammed using the four Yamanaka factors to generate a series of 25 iPS cell lines derived from each of the three adult tissues (hereafter F-, K- and E-iPS cells). The differentiation properties of our iPS cell lines were checked by differentiating a subset to endoderm, mesoderm and neuroectoderm derived cell types (Figure S1). Polyadenylated mRNA was prepared from each iPS cell line after 10 passages, the somatic cell lines at passage 3 and from two human ES cell lines which was subjected to high-depth RNA-seq (Fig. 1a). We checked a subset of lines for Sendai virus persistence using RT-PCR with virus-specific primers, but found no evidence for viral presence following passage (Figure S2). For each sample, we performed 75 bp paired end sequencing on the Illumina HiSeq2000 platform. In total we generated 7.3 billion reads, with between 85.3 and 229.8 million reads sequenced in each sample (Figure S3). We mapped reads to assembly h37 of the human genome using Bowtie2 [14] and constructed spliced alignments using Tophat2 [15]. Following read alignment and QC filtering, between 49% and 89% of reads mapped uniquely to the human genome (Figure S4).
Initially we examined the global pattern of transcription in the different cell lines. Previous work has suggested that embryonic stem cells are more transcriptionally active than differentiated cells [16]. Within protein coding regions there was a clear bimodal distribution of gene expression levels in all samples reflecting abundantly expressed and transcriptionally repressed genes (Figure S5). Adult cell lines exhibited more completely repressed and very highly expressed genes compared with hiPS cells and hES. Relative to differentiated lineages approximately twice as many genes could be classified as coming from the repressed mode of the distribution in the adult cell lines compared with hiPS cells and ES cells (Figure S 5, inset). We also found that more transcription appeared to be originate from repetitive elements in pluripotent stem cells (Figure S6), although it is unclear whether this arises from a slightly more relaxed chromatin structure in stem cells, or is an artifact resulting from the slight excess of very highly expressed genes in somatic cells relative to pluripotent stem cells.
Next, we sought to quantify the contribution of multiple biological and experimental factors to transcriptional variation in hiPSCs. Hierarchical clustering clearly placed adult somatic cells and pluripotent cells in two distinct clades (Fig. 1b). However other sources of variation, such as tissue of origin, were more difficult to discern. To quantify the relative importance of different sources of global transcriptional variation more precisely we employed a variance component analysis. Here, transcriptional variation was decomposed into five separate components. These components comprised: (1) a random intercept term (2) a component to capture variation in transcription between the three adult somatic tissues, hESCs and hiPSCs (3) a component modeling differences between F-, K- and E-iPSCs (4) a component capturing transcriptional variation between different donors or genetic backgrounds and (5) a component captures differences between the two sequencing batches in our data set (see Text S1). We quantified the contribution of each of the five components using intraclass correlation, which measures the proportion of total transcriptional variance explained (VE) by different experimental groups holding other model factors constant. As such, the estimated VEs for each component are not constrained to sum to 100%. Throughout, we modeled the effect of sequencing batch to disentangle its potential influence from the other variance components in the model.
We found that inter-individual transcriptional variation in hiPS cells (VE ∼38%) is considerably larger than that between somatic tissue of origin (VE∼4%) with an even smaller fraction of transcriptional variation (<1%) explained by differences between iPSCs and ESCs (Fig. 1c). Strikingly, when we didn't correct for variation between individuals, transcriptional variation between iPSCs and ESCs and between different iPSC tissues of origin appeared to be much larger (<1% vs 12.7%, iPSCs vs ESCs, 4% vs 13.5%, iPSC tissue of origin, with versus without individual included in the model, respectively) (Fig1c). This suggests that some previous observations of cellular memory and transcriptional differences between iPSCs and ESCS may in fact arise from changing genetic backgrounds rather than experimental effects. Confounding with donor genetic background seems particularly plausible for comparisons between iPSCs and ESCs where controlling for genetic differences is often impossible [4], [17], or in cases where iPSC tissue of origin is confounded with donor genetic background [9], [18]. We found that fibroblast- and keratinocyte-derived iPS cells (F- and K-iPSCs) were highly similar at the transcriptional level with tissue of origin explaining <1% of the transcriptional variation in the between them. In fact, the majority of the tissue of origin effect we observe arises from differences between F/K-iPS cells and EPC-derived iPS cells (E-iPS cells). Some of this signal could reflect transcriptional memory. However it is possible that the reprogramming method could account for this difference because the EPCs were resistant to Sendai virus infection and were therefore derived using the retroviral method. Individual transcriptional variation was slightly greater in adult somatic cells (VE∼42%) than in stem cells. We speculate that this could potentially be explained by non-genetic differences between individuals, such as varying methylation status, which are present in somatic cells but erased during cellular reprogramming.
We also examined the amount of residual transcriptional variation that remained unexplained by any of the known factors using a heteroscedastic model (Text S1). This extension of the model allowed us to capture differences in residual variance between different subsets of our data set. In our experiment, the residual variance captures transcriptional variation between different cell lines from the same donor, either as growths of different IPS cell lines derived from the same donor, from different growths of either the ESCs or as different passages of adult cells from the same donor. We compared two heteroscedastic models with a homoscedastic model, which contained a single error term for the entire data set (Figure 1d, “All”). Heteroscedastic model 1 (“model 1”) contained three terms representing variation between biological replicates of adult cells, IPS cells and ESCs. The results of this analysis illustrate that the residual variation between replicates of IPS cells is lower than that in adult cells and ESCs (Figure 1d: Heteroscedastic Model 1 “F/K/E-iPSCs” versus “ESC”). Heteroscedastic model 2 further divided error into components for each of the three different adult cell types, for each of the three IPS cell types and for ESCs. This analysis demonstrated that variation between biological replicates of IPS lines derived from different tissues was relatively consistent (Figure 1d: Heteroscedastic Model 2). Overall, our data suggest that transcriptional variation between biological replicates of iPS cells was not substantially, and may be somewhat lower, different from that between passages of an established hES cell line or of adult primary cells.
Given that our variance component analysis was based on FPKM values, we also reanalyzed data excluding highly expressed genes, as this may impact our results. We found that our component estimates and correlation heatmap were qualitatively very similar when the top 1 and 5% of genes were removed from the data set (Figure S7, S8). We also attempted a similar analysis of mitochondrial gene expression. We found similar proportions of reads coming from mitochondrial genes in adult, IPS and ES cells (Figure S9). However, the very low number of expressed genes (13) resulted in extremely noisy estimates of the correlation between sample gene expression profiles (Figure S10) and prohibited variance components analysis, due to failure of the model to converge. Our differential expression analysis classified all mitochondrial genes as “invariant expression” (data not shown). Our experiment did not include multiple replicates from the same cell line, and so we were unable to formally address the issue of variation between lines from the same donor. Visual inspection of our read coverage plots did suggest that some cell lines might be more variable than others. Heatmaps of the differentially expressed genes illustrate that while most cell lines were quite consistent, one cell line derived from keratinocytes formed an outgroup (K-iPSC-S2-1) with other keratinocyte cell lines (Figure S11).
Although our variance components analysis suggested relatively small global effects of tissue of origin, this could mask effects at individual genes. We next sought to identify those genes whose expression in hiPS cells more closely resembled their somatic progenitors or were improperly silenced or activated, relative to ES cells (Fig. 2a). For each of the three somatic tissues in turn, we performed a three-way comparison of expression levels in the somatic tissue, in the hiPS cells derived from that tissue, and in the hES cells. At each gene we tested for departures from a null hypothesis of equal expression levels in the somatic cell, the iPS cells and the ES cells using a negative binomial generalized linear model similar to that outlined in [19] (Text S1). When the null hypothesis (“invariant expression”) was rejected we further classified genes into one of four possible categories, “correctly reprogrammed”, “transcriptional memory”, “aberrantly reprogrammed” and “complex” by selecting the alternative with the highest log likelihood ratio.
We used a hierarchical model (Text S1), similar to that in [20], to estimate the true proportion of transcribed genes in each category without setting a threshold on statistical significance or effect size. Our results suggested that transcriptional memory is very uncommon, occurring at 0.06, 0.06 and 0.20% of all expressed genes in F-, K- and E-iPS cells (Fig 2b). Aberrant or complex expression patterns were also infrequent, although aberrant expression appeared slightly more often in E-iPSCs (0.15% of genes) than in the other two tissues. The fractions falling into the complex category were similarly low (0.10, 0.02 and 0.74% of genes, respectively). The remaining genes either showed invariant expression or were classified as correctly reprogrammed (99.83, 99.92 and 98.92%). Core pluripotency markers, including Sox2, Nanog and Oct4, were all classified as correctly reprogrammed (Figure S12). Our analysis suggested that the fraction of correctly reprogrammed genes is approximately 30-60% of expressed genes. Although comparison between different studies and methodologies difficult, this is not dissimilar to fraction of genes that are differentially expressed in reprogramming found by other studies [21]).
Next we identified individual loci that were differentially expressed at a genome-wide false discovery rate (FDR) of 5% (estimated from permuted data), and that exhibited a 1.5-fold or greater change in expression level between hiPS cells and hES cells. Using these criteria we identified a total of 61, 5 and 103 transcribed regions that showed some form of differential expression in F-, K-, or E-iPS cells respectively (Table 1), the majority of which exhibited either complete or partial transcriptional memory (“TM” or “PTM”). Most genes we identified were very weakly expressed in IPS cells, with mean FPKMs 74–79% lower than the average (Figure S13). The most enriched gene ontology (GO) terms in the TM and PTM gene sets were for mesodermal migration in F-iPS cells (enrichment q-value <0.002;), for hemidesmosome development in K-iPS cells (q<0.03) and for inflammatory and defense response in E-iPS cells (q<3×10−5) (Figure S14). Aberrantly expressed genes were less frequent than genes exhibiting transcriptional memory. One interesting exception to this was the long noncoding RNA, H19, which was highly expressed in the majority of hiPS cells in our data set relative to the hES cells (Fig. 2d). Our differential expression analysis also suggested that aberrant activation occurs more often than aberrant silencing (21 versus 6 genes at FDR 5%), and that transcriptional memory more frequently involves memory of an active rather than a silenced gene (101 versus 19 genes at FDR 5%) (Table 1). However, this may simply reflect better power to detect differential expression when a gene is highly expressed rather than silenced. Removal of highly expressed genes had almost no impact on our differential expression analysis (Figures S6, S7). Finally, differential expression analysis of mitochondrial genes classified all 13 genes we found be expressed in any of our samples as “invariant expression”. Overall, our results suggest that, although transcriptional memory and aberrant reprogramming do occur occasionally, relatively few genes are involved, those that are affected are weakly expressed.
Although whole gene expression levels appeared to be relatively stable across different tissues of origin, cellular memory could also manifest at the level of RNA splicing. Using the statistical framework we developed for gene expression levels, we next tested whether isoform abundance ratios showed evidence of memory of their cell type of origin (Text S1). In this analysis, we tested the null hypothesis that the ratio of the top two most abundant isoforms was equal in adult tissues, iPS cells and ES cells. We computed abundances using two popular approaches, Cufflinks2 and MISO [22], [23], and analysed the subset of genes where the ranking of isoform abundances agreed between the two methods (Figure S15). We found no significant memory of adult cell splice patterns or aberrant alternative splicing in IPS at an FDR of 5% estimated from permuted data (Figure S16). This suggests that transcriptional memory of cellular splice patterns or aberrant splicing induced during reprogramming is a relatively weak effect, smaller than that observed at the level of whole gene expression, and potentially masked by larger technical and genetic effects.
Previous studies have suggested that imprinted gene expression patterns may be unstable in hES cells and hIPS cells [24]. We genotyped the four individuals from whom our hIPS cell lines were derived using an Illumina Omni2.5 genotyping chip. Genotypes were phased, and SNP genotypes were imputed using Beagle [25]. Using the phased haplotypes, we computed estimates of allele-specific expression for the paternal and maternal chromosomes of each individual. We began by investigating whether patterns of imprinting observed in the adult somatic tissue remained conserved in the iPS cells that were derived from them in a set of 210 putatively imprinted genes obtained from the http://www.geneimprint.com/database. We found that many genes were expressed at a low level or lacked sufficient coverage of heterozygous SNPs, which were subsequently excluded. The remaining genes (26, 23 and 23 in donor S2, S4 and S7, respectively) exhibited allelic expression patterns in adult cells that were conserved in their derived iPS cell lines (Pearson r2 = 0.46; p<6.3×10−10), Fig 3. Many genes did not demonstrate the characteristic mono-allelic expression of imprinted genes in either adult or pluripotent cells (Fig 3). In a small number of cases, such as the paternally expressed zinc finger gene, ZDBF2, we observed a loss of imprinting and reversion to bi-allelic expression in many IPS lines. We note that, in the cases where we observe loss or alteration of imprinting, the genes involved are relatively weakly expressed in either the adult or the IPS cell, making ascertainment of imprinting status more difficult.
Finally we returned to the effects of the genetic background in our hiPS cells. Extensive maps of genetic variants whose genotype correlates with gene expression (expression quantitative trait loci, eQTLs) have been generated in adult human tissues and cell lines [26]. We investigated whether similar effects could be observed in our iPS cells. Since the number of individuals in our data set was small, a standard eQTL mapping experiment was not possible. Instead, we tested whether genetic associations ascertained in lymphoblastoid cell lines (LCLs), a model system for eQTL detection in humans, were also detectable in iPS cells.
We reanalyzed an existing LCL RNA-seq dataset derived from the same source population (162 GBR+CEU individuals) as our hiPS cells (Figure S17) and identified 4,350 eQTLs at an FDR of 5% [27]. Variance component analysis revealed a greater amount of variation between individuals in genes that were ascertained to have an eQTL in LCLs, than in genes where the null of no eQTL could not be rejected (Fig 4a). A substantial fraction (17%) of this variation could be explained by the lead eQTL SNPs (eSNPs). We tested whether the direction of the genetic effect at the LCL eQTL genes replicated in iPS cells by grouping the four individuals in our data set according to their eSNP genotype. We found that the expression level of genes with an eQTL ascertained in LCLs follows the expected direction in hiPS cells (Fig. 4b). Likewise, for individuals in our dataset that are heterozygous at the eSNP we see a corresponding, highly significant allelic imbalance also in the expected direction (Fig. 4c). The correlation between genotype and expression level at ascertained eQTLs in hiPS cells was highly significant (Pearson r = 0.44, p<9.8×10−68), as was the allelic imbalance at heterozygous ascertained eSNPs (Student t, p<1.3×10−8). Although our data set is small, we do find convincing examples where a correlation between genotype and gene expression replicated a known eQTL identified in LCLs such as the exonic eSNP, rs1059307, located within the noncoding RNA gene SNHG5. At this gene, we also observe clear allelic imbalance in iPS cells derived from S5 and S7 individuals, who are heterozygous for this eSNP (Fig. 4e), which is in the same direction as the eQTL effect in LCLs (Fig. 4f). These genetic effects on gene expression in iPS cells were detectable across multiple independent iPS cells from the same genetic background, despite the variety of different tissues sources and reprogramming methods.
We have shown that epigenetic memory of the adult progenitor cell is a rare phenomenon in hIPS cells, and that cellular heterogeneity between different hIPS lines is more likely to be driven by changing genetic background. Our study has important implications for future attempts to use iPSCs as cellular model systems for drug discovery and other applications. Encouragingly, our results suggest that genetic effects are readily detected in hIPSCs and that cell phenotypes are highly reproducible within individuals. Equally important, however, is the fact that the noise introduced by genetic background could potentially obscure small genetic signals of interest in small samples. A clear implication of this result is that, in iPSC-based studies of genetic disease, most effort should be expended on collection of samples from different donors rather than generation of large numbers of lines from the same individual. Collection of multiple individuals, perhaps with a shared genotype at a single locus of interest, will allow the effects of genetic background to be averaged over and separated from that of the putatively causal locus.
Our study also highlights how the effects of genetic background cannot be ignored when considering cellular variability between pluripotent stem cell lines. Previous studies have attributed cellular variability in IPS to a range of sources, including epigenetic memory [5]–[10], [17], inherent differences between IPSCs and ESCs [4], [28], artifacts of reprogramming [7] or lab environment [29]. Perhaps surprisingly, the effects of genetic background have been less well appreciated, although more recent work has highlighted its potential importance in differentiation [30]. It seems likely that at least some of variability previously reported to exist in IPS cells could in fact have arisen from genetic differences. This is particularly true of comparisons of IPS with ES that are typically derived from different individuals. It is also notable that studies including larger numbers of donors tend to find fewer transcriptional differences between IPS and ES [29], [31]–[33]. Studies that have not controlled for genetic background when investigating epigenetic memory, such as by confounding tissue of origin and donor, may also have mistakenly attributed genetically driven differences in transcription to epigenetic memory. Our study explicitly incorporates multiple tissues from the same donors, allowing us to correct for the effects of changing genetic background. This is likely to explain why we do not find extensive apparent epigenetic memory. We note, however, that other studies that have also explicitly controlled for genetic background still report some variation in transcription, methylation and differentiation efficiencies that appear to arise from cell type of origin effects [5], [8], [10]. A possible explanation for the discrepancy between the results of these studies and our own is that we have also taken our samples from cells at between 10 and 13 passages and epigenetic memory effects may be transient and disappear following multiple passages [8].
An important caveat for our study is that influential cellular differences may simply not manifest as transcriptional variation but reside at, for example, the epigenetic level as changes in methylation status or histone tail modifications. Such differences may harbor a “hidden” functional role that only becomes apparent upon differentiation into a specific cell lineage. Our study suggests that epigenetic differences are likely to be more plausible candidates as drivers of variation in IPS cell differentiation ability. However, our results also illustrate that current iPS cell technology is robust enough to enable detection of genetic effects on important cellular phenotypes such as mRNA levels. Although further technological hurdles remain, an exciting area for future work will be detection of regulatory variation that influences transcription during cell lineage specification and differentiation, employing iPS cells as a model system.
All primary tissue samples and blood for this project were obtained from adult cadaveric organ transplant donors referred to the Eastern Organ Donation Services Team (part of NHS Blood and Transplant). Ethics approval was obtained from the local Research Ethics Committee (REC No. 09/H306/73).
For each subject included in this study, a sample of skin was excised from the midline surgical incision. The skin was transported to the lab and washed in iodine and ethanol and was cut into approximately 1 mm3 pieces. These were dispersed evenly on a 90 mm plate and incubated with fibroblast media (Knockout DMEM and 10% FBS). Outgrowths of fibroblasts and keratinocytes from the skin explants were usually apparent at around 14 days. The cells were separately harvested using 5 min treatment with Versene (15040-066, Invitrogen), which detached the fibroblasts leaving the keratinocytes on the plate. The fibroblasts were cultured on non-coated plates using fibroblast media and keratinocytes were cultured on plates coated with matrix (R011K, Invitrogen) and using EpiLife media plus Defined Growth Supplement (M-EPI-500-CA and S-012-5, Invitrogen).
Endothelial Progenitor Cells were derived from 100 mL of peripheral blood as previously described [11]. Briefly, the mononuclear cells of the blood sample were separated using Ficoll. The cells were cultured on collagen-coated plates using EPC media (EGM-2MV supplemented with growth factors plus 20% Hyclone serum; CC-3202, Lonza and HYC-001-331G; Thermo Scientific Hyclone respectively). Colonies of EPCs appeared at around 10 days.
Four pseudotyped Moloney murine leukemia retroviruses containing the coding sequences each of human OCT-4, SOX-2, KLF-4 and C-MYC were obtained from Vectalys (Toulouse, France). A multiplicity of infection of 10 was used in all retroviral reprogramming experiments. For each hiPS cell derivation, 1×105 EPCs were transfected with the 4 viruses in the presence of 10 ug/mL of polybrene (TR-1003-G, Millipore). After 24 hrs the viruses were washed off with PBS and the cells were re-fed with EPC media that remained for the next 4 days. On day 5 after transduction, the cells were re-plated using trypsin onto a 10 cm dish of fresh MEF feeders. After 2 days the media was changed from primary cell-specific to hiPSC media (KSR + FGF-2). The media was changed every 2 days until colonies emerged after which the media was changed daily. Colonies were picked once they had reached sufficient size, typically from day 25 following transduction. The colony was split into quarters and the segments gently lifted off the plate and transferred to one well of a 12 well plate of fresh MEF feeders containing hiPS cell media (KSR + FGF2) supplemented with ROCK inhibitor (Y-27632, Sigma).
Four Sendai viruses containing the coding sequences of each of human OCT-4, SOX-2, KLF-4 and C-MYC were obtained from DNAVec (Ibaraki, Japan). The protocol for reprogramming was identical to that of retroviruses with the exceptions that 5×105 primary cells (fibroblasts or keratinocytes) were used at MOI 3 and polybrene was not used.
hiPS cells were grown on irradiated MEF feeders, using human embryonic stem cell media (termed KSR + FGF-2): Advanced DMEM (12634-010, Invitrogen) was supplemented a follows: 10% Knockout Serum Replacement (10828028, Invitrogen), 2 mM L-glutamine (25030024, Invitrogen, 0.1 mM β-mercaptoethanol (M6250, Sigma-Aldrich) and 4 ng/µL of recombinant human basic Fibroblast Growth Factor-2 (233-FB-025, R&D systems, Minneapolis, MO, USA). Media was changed daily and the cells were passaged every 5–10 days depending on the confluence of the plates. To passage hiPS cells, the plates were washed in PBS and colonies detached using collagenase and dispase (Collagenase IV 1 mg/mL, Invitrogen 17104-019; Dispase 1 mg/mL, Invitrogen 17105-041). The colonies were washed in media and mechanically broken up before being re-plated onto fresh MEF feeders.
Total RNA was extracted using the RNeasy Mini Kit protocol (Qiagen, Hilden, Germany). RNA-seq libraries were constructed according the manufacturers guidelines, with minor modifications, using the Illumina mRNA-seq and TruSeq mRNA sample preparation kits (Illumina, Inc., San Diego, CA). Briefly, mRNA was enriched from total RNA using oligo dT beads before fragmentation via zinc and heat hydrolysis. mRNA was subject to first and second strand cDNA synthesis before end repairing and A-tailing. Double-stranded cDNA was then adapter-ligated before size-selecting fragments with inserts ranging from 200–300 bp using a LabChipR XT (Perkin Elmer, Waltham, MA). Size-selected material was then PCR-amplified using KAPA HiFi polymerase (Kapa Biosystems, Boston, MA) before sequencing on an Illumina HiSeq2000 (Illumina, Inc., San Diego, CA).
We mapped reads to assembly h37 of the human genome using Bowtie2 [14] and constructed spliced alignments using Tophat2 [34] with default settings. We also used known gene annotation information given by Ensembl release 69 as a guide for the alignment. Following read mapping, we selected fragments (read-pairs) where at least one of mate-pairs had a quality score of >10, aligned with no gaps, with three base mismatches or less. Any read pairs with an insert size less than 150 bp or greater than 1 Mb, or on different chromosomes, were excluded from subsequent analyses. Computational analysis was carried out using a combination of existing packages, such as DESeq [19] and our own analysis tools. For the variance components analysis, transcription level at each gene j was modeled as a linear combination of five normally distributed random effects (b1–5) and a single error term:where is a vector of log normalized fragments per kilobase per million reads sequenced (FPKMs) for gene j in each of the 46 samples in our data set, b1 is an intercept term, b2 models variation in transcription between the three adult somatic tissues, hESCs and hiPSCs, b3 models differences between F-, K- and E-iPSCs, b4 captures transcriptional variation between different donors, b5 captures differences between the two sequencing batches in our data set, ε is the error term and Z1-Z5 are design matrices. For full details of computational and statistical analyses, see Text S1.
Our raw sequence data are available from the European Genotype Archive under study ID EGAS00001000367. A variety of processed data, including raw read counts, log2 FPKMS and the results of our differential expression analysis are available from our lab website (http://www.sanger.ac.uk/research/projects/genomicsofgeneregulation/) under the “Data” tab.
|
10.1371/journal.pntd.0002991 | Risk Factors and Spatial Distribution of Schistosoma mansoni Infection among Primary School Children in Mbita District, Western Kenya | An increasing risk of Schistosoma mansoni infection has been observed around Lake Victoria, western Kenya since the 1970s. Understanding local transmission dynamics of schistosomiasis is crucial in curtailing increased risk of infection.
We carried out a cross sectional study on a population of 310 children from eight primary schools. Overall, a total of 238 (76.8%) children were infected with S. mansoni, while seven (2.3%) had S. haematobium. The prevalence of hookworm, Trichuris trichiura and Ascaris lumbricoides were 6.1%, 5.2% and 2.3%, respectively. Plasmodium falciparum was the only malaria parasite detected (12.0%). High local population density within a 1 km radius around houses was identified as a major independent risk factor of S. mansoni infection. A spatial cluster of high infection risk was detected around the Mbita causeway following adjustment for population density and other potential risk factors.
Population density was shown to be a major factor fuelling schistosome infection while individual socio-economic factors appeared not to affect the infection risk. The high-risk cluster around the Mbita causeway may be explained by the construction of an artificial pathway that may cause increased numbers of S. mansoni host snails through obstruction of the waterway. This construction may have, therefore, a significant negative impact on the health of the local population, especially school-aged children who frequently come in contact with lake water.
| It is estimated that more than ten percent of the world's population is at risk of schistosome transmission, with over 90% of infections occurring in sub-Saharan Africa. In Kenya, schistosomiasis remains a major public health concern particularly around Lake Victoria. The objective of this study was to identify the risk factors associated with Schistosoma mansoni infection among schoolchildren on the shores and adjacent islands of Lake Victoria in Mbita district, western Kenya. High local population density was identified as an important risk factor for S. mansoni infection. Socio-economic factors were not found to be significantly associated with infection risk. Our study suggests that environmental changes related to causeway construction and the dense human population around Mbita town may result in favourable ecological conditions for S. mansoni transmission.
| Schistosomiasis is a parasitic disease affecting 249 million people worldwide. It is endemic in 78 countries with over 90% of cases occurring in sub-Saharan Africa [1]. About 779 million people, more than 10% of the world's population, were estimated to have been at risk of schistosome infection in mid-2003 [2]–[5]. In Africa, schistosomiasis is due predominantly to infection with Schistosoma mansoni, which causes intestinal schistosomiasis, and Schistosoma haematobium which causes urinary schistosomiasis [6].
Small scale spatial heterogeneity is a typical epidemiological feature of schistosomiasis [7], [8]. Such heterogeneity is closely associated with the distribution of the snail intermediate host, and with human contact with infective water [9], [10]. Past studies showed correlation between schistosome transmission and several epidemiological and socio-economic factors such as age [11], [12], sex, sources of drinking water, latrine availability, sanitation, hygiene [13]–[17], urbanization and population growth. [18]–[21]. Moreover, some works have merged both our understanding of demographic risk factors together with environmental transmission dynamics in order to create large scale (national, regional, continental) maps that are instrumental in designing control programmes for the disease [22]–[25], while small scale analysis is important in contributing to the local distinct need [26]–[28].
Schistosomiasis is increasingly a major health problem among communities around Lake Victoria. Geographical patterns of S. mansoni infection have been described in this area in relation to proximity to the lake [25], [29]–[31] and in comparison of islands versus mainland habitation, where risk is higher on the islands [32]. Identifying local risk factors of infection at multiple levels is crucial so as to understand how transmission varies within small spatial scales and how it changes over time. In addition, identifying risk factors may facilitate disease control by targeting high risk groups or by informing possible intervention strategies. The main objective of this study was to identify the risk factors associated with S. mansoni infection among schoolchildren in Mbita and the two adjacent islands (Rusinga and Ngodhe) of Lake Victoria, Kenya.
The study was reviewed and approved by the scientific steering committee and ethical review committee of the Kenya Medical Research Institute, Kenya (KEMRI, SSC No. 2084), and the ethical review board of Institute of Tropical Medicine, Nagasaki University, Japan (No. 10121666). Written informed consent was obtained from parents/guardians and schoolchildren prior to the study. Children infected with schistosomes were treated with 40 mg/kg praziquantel and those infected with soil transmitted helminths (STHs) were treated with 400 mg albendazole by a clinical officer in accordance with WHO guidelines [33]. All children positive for malaria were treated with artemether/lumefantrine (AL) according to national guidelines for uncomplicated malaria [34]. A study feedback meeting was held with parents or guardians of participants, as well as the head masters and health teachers of the schools.
This study was conducted on the shores and islands of Lake Victoria in Nyanza province, Mbita district, western Kenya (Figure 1) in an area covered by a health and demographic surveillance system (HDSS) [35], [36]. The Mbita HDSS includes Rusinga east and west on the island and Gembe east and west on the mainland. Ongoing HDSS data showed that the total population in Mbita was 55,929 during our survey conducted in 2011. Notably, population density on Rusinga Island was twice as high as in the Gembe region. In Mbita district, the waterway separating Rusinga Island from the mainland was filled in 1985 and a road to Rusinga Island was constructed to facilitate transportation of people, goods and services. Economic activities are high around and within a 5 km radius from the centre of Mbita causeway, referred to as an urban area, while rest of the study area was treated as rural. Mbita is dominated mainly by the fishing communities living in the immediate vicinity of the lake. The temperature in Mbita ranges from 15°C to 30°C. Rain seasonality is bimodal with a short rainy season starting from October to December, while a longer rainy season lasts from March to May. The average annual rainfall ranges between 800–1,200 mm in the western part of the study area in Rusinga Island while Gembe receives slightly higher rainfall of 800–1,900 mm. In our study, HDSS data was used for obtaining household locations and population density.
A cross-sectional study was conducted between September and November 2011. According to the education office in Mbita district, the primary school enrollment rate was 91.6%. The inclusion criteria of the schools were to be a full grade primary school and not to have received mass-chemotherapy for a year prior to the study. As most of the private primary schools received mass drug administration for STHs a year prior to the study, they were excluded. The schoolchildren in 4th grade were targeted in this study and the total number of them were 1,747 in 2011. Of these, 888 were females and 859 were males. Among the 64 public primary schools, 39 schools met inclusion criteria and 8 schools were randomly selected as clusters (Figure 1). Parent/guardian and teacher association meetings were held in all selected schools prior to the survey for communicating the study purpose and obtaining their consent with full understanding. Ninety-eight percent of parents/guardians consented and consequently, 310 of all 4th grade children were enrolled in the study.
All children were instructed to provide stool specimens in a labeled specimen cup on three consecutive days. The school health teacher or class teacher guided students on stool sample collection during container distribution, a day before the survey. A trained field worker visited the school during morning break time with a registration sheet to ensure all students provided samples. Those who did not provide samples were followed up to ensure each child provided maximum possible samples. The Kato-Katz fecal thick smear technique was used for the detection and the quantification of S. mansoni eggs and the presence of STHs. Intensity of infection was estimated as the number of eggs per gramme of feces (epg) [37], . Slides were prepared and examined by two independent readers within an hour for hookworm egg detection and within 24 hours for the rest of the parasites in focus. Parasite eggs were counted and the arithmetic mean of 3 slides per child was calculated to give the intensity of infection. The extent of S. mansoni infection was categorized as light (1–99 epg), moderate (100–399 epg) or heavy (≥400 epg) according to WHO guidelines [33]. The intensity of infection per school was calculated as the geometric mean of egg excretion among all children testing within the school. In addition, the presence of S. haematobium and Plasmodium spp. were examined to assess their association with S. mansoni infection. Midday urine was collected for the detection of S. haematobium eggs using direct microscopy examination since S. haematobium is known not to be endemic in the study area [6]. Venous blood was collected for the microscopic examination of Plasmodium spp. by thick and thin Giemsa stained smears. Additional haematological and serological examinations were also carried out for a separate study.
To identify risk factors for S. mansoni infections, trained interviewers administered a questionnaire to children during the parasitological survey at the school, while parents/guardians were interviewed in the household setting. Information about individual treatment history for schistosomiasis and the water contact behaviour of each child was collected in the school setting. For socio-economic factors, the household head or the most informed adult present during the household interview gave information on: ownership of land, household size, total number of rooms in house, mother's/female guardian's education level and the main source of drinking water. The age of each child was confirmed by cross checking with official birth certificates or church baptism cards during household visits. In addition, an observation checklist was used to collect information on house structure, latrine and electricity availability in each household. Houses were categorized into two groups: traditional houses with grass roofs and modern houses with iron sheet or cemented roofs. The number of persons per room was obtained as one of the indicators of socio-economic status by dividing the household size by total number of rooms in the house. Households with more than two persons per room were categorized as overcrowded, since the average number of persons per room was two in the study population.
Participants were defined as being positive for each helminthic infection if at least 1 egg was detected in their stool for S. mansoni, STHs or in urine for S. haematobium. For P. falciparum infection, both thick and thin blood smears were examined using a light microscope at ×100 with an oil immersion objective. Positive cases were defined as those with at least one malaria parasite detected in the microscopic field of 200 white blood cells for thick film or 2,000 red blood cells for thin film [39]. The intensity of helminth infection was expressed as the arithmetic mean of three slides per child, while the intensity of helminth infection per school was expressed as the geometric mean. Total egg counts of S. mansoni in fecal samples were analyzed in relation with potential risk factors by using both a generalized linear model (GLM) and a generalized linear mixed model (GLMM) with school as a random factor. Since fecal egg count is over-dispersed, a negative binomial generalized linear model (NB-GLM) and a negative binomial generalized linear mixed model (NB-GLMM) were used. The mixed model was employed to account for the potential lack of independence among samples that emerges from children attending the same school [40]. As children attending the same school were clearly clustered around the school, the school effects might be interpreted as the effects of the areas where children reside.
Local population density was obtained for each child using the HDSS population data. The number of people living within a radius of 1 km was counted for all participants' houses using Quantum GIS version 1.7.4. [41]. This scale was selected because population density showed the strongest association with S. mansoni infection when the radius was set at 1 km. The HDSS population data was incomplete for Ngodhe Island and therefore we used the total population of the island (449 persons; according to a local health staff), since most houses on the island were within 1 km from another participant's houses. The shortest straight-line distance from the study participants' house to the lake shore was obtained using Quantum GIS [41]. Spearman's rank correlation was used to test associations between prevalence and intensity of S. mansoni infection and distance to the lake. All statistical analyses were carried out using R version 3.0.1[42] and P-values less than 0.05 were considered significant. The glmmADMB package was applied for analysis of an over-dispersed continuous variable (infection intensity).
To examine whether the intensity of S. mansoni infection is spatially clustered, a spatial scan statistical treatment was applied to point data on household location using SaTScan software (version 9.1.1.) [43]. As models for over-dispersed count data are not available in SaTScan, we applied normal model to the log (N+1) transformed egg count. A purely spatial model was applied and a scan for areas with high values was performed. The maximum size of high-risk clusters was set to 50% of the total number of subjects. To evaluate statistical significance, 999 Monte-Carlo replications were conducted. To examine whether any spatial clusters could be explained by individual risk factors, a scan was also performed with the residual values of a linear regression model of log (N+1) with independent variables which were significantly associated by the egg count in the NB-GLM. The spatial clusters of the residuals can be interpreted as those adjusted for independent risk factors.
Demographic and socio-economic characteristics of the study participants are shown in Table 1. The study involved 310 fourth grade schoolchildren from eight schools, 138 (44.5%) of the children were male, and 172 female (55.5%). Their ages ranged from 9 to 19 years and the median age was 12 years for both sexes. The majority (81.0%) of the children lived in traditional houses with an average of 2–3 rooms. Most of the children (95.3%) lived in overcrowded houses, 95.8% had no electricity and 53.9% had no latrine. Over three quarters (76.4%) of families owned land and 84.8% of mothers/female guardians completed 4th grade or further education. The majority of households (84.8%) used the lake as the main source of drinking water. Apart from one child, the rest of the children (99.7%) had routine lake water contact an average of 2–3 times per week mainly through bathing and domestic washing purposes. Individual treatment history of schistosomiasis was also confirmed by questionnaire and none of participants was treated at least one year before the study.
The overall prevalence of schistosomes, STHs and P. falciparum in each school is summarized in Table 2. More than three quarters (76.8%) of the students were infected with S. mansoni, while seven (2.3%) were infected with S. haematobium. All the children infected with S. haematobium were co-infected with S. mansoni and had previously stayed in areas endemic for S. haematobium, further inland from Lake Victoria. At least 12.6% of the schoolchildren were infected with one or more species of STHs. Prevalence of hookworm, Trichuris trichiura and Ascaris lumbricoides was 6.1%, 5.2% and 2.3%, respectively. Thirty-seven schoolchildren (12.0%) were infected with P. falciparum. A total number of 248 (80%) were infected with at least one of the examined parasites. Co-infection of S. mansoni with P. falciparum 13.1% (31/236) was the most common in the study area. In addition, multiple infections with more than three species were found in a few cases. There was no difference in males and females in co-infection of S. mansoni with P. falciparum (P = 0.52). All possible combinations for parasitic infections were found nearly in the expected numbers (data not shown), indicating neither synergistic nor antagonistic effects of polyparasitism.
The prevalence of S. mansoni differed significantly between schools, ranging from 31.7 to 98.3 percent, (Pearson chi-square test P<0.001). Age was not associated with prevalence of S. mansoni in this study. There was no significant difference between males and females for prevalence of any examined parasitic infections. Table 3 shows the intensity of S. mansoni infection in each school. Among those who were positive for S. mansoni eggs, the geometric mean number of eggs excreted per gramme of feces (epg) varied from 2.0 to 303.5 epg between schools. The overall mean intensity of S. mansoni infection was 207 epg with inter quartile range of 8 to 214 epg. The intensity of S. mansoni infection was categorized according to the WHO guidelines [33], children with light, moderate and heavy infections were 110 (35.5%), 78 (25.2%) and 50 (16.1%) respectively.
At the school level, the intensity of S. mansoni infections was strongly correlated with its prevalence (Spearman's rank correlation, rho = 0.98, P<0.001). The four schools with high prevalence and intensity of S. mansoni infection (Wasaria, Wakondo, Kamasengre, and Kombe) were aggregated around the bay in the west side of Mbita causeway (Figure 2). The mean of log (N+1) transformed egg counts were significantly different between inside and outside of clusters as 5.23 and 2.55, respectively (common estimate for standard deviation, 1.94; P = 0.001).
Table 4 shows the results of a bivariate analysis on the association between the intensity of S. mansoni infection and the potential risk factors with no consideration of school effects (NB-GLM). This indicated that males were more intensely infected than females (marginally significant). Several household-based factors also showed a significant association with high infection risk; houses in areas with higher population density, permanent houses and houses with latrine.
Spatial scan statistics was performed for the residuals of the regression model of log (N+1) transformed egg count. The variables that were found to be significant and/or marginally significant in the NB-GLM (sex, population density, house structure and houses with latrine; orange dotted circle in Figure 2) were included in the calculation. A significant high-risk cluster occurred in a similar location to the unadjusted cluster (red circle in Figure 2) although the size of the cluster was smaller (radius of 4,006 meters; 53 children were included). The adjusted cluster included all the children of Wakondo and some of the children of Wasaria and Kamasengre but did not include any of the children living in Kombe. The mean of the residuals was significantly higher inside (1.50) than outside the cluster (−0.31), common estimate of standard deviation, 1.92; P = 0.001.
When the school areas were included as random factor in a NB-GLMM, the effects of sex, house structure and latrine became non-significant (Table 5). Local population density was the only statistically significant factor for S. mansoni infection in the NB-GLMM (Table 5; P = 0.011). The association between population density and intensity of S. mansoni infection is further depicted in the map and scatter plot as shown in Figure 2 and 3. Evidently, population density is an independent factor influencing risk and intensity of S. mansoni infection, while evaluated socio-economic factors appeared not to affect the risk and intensity of S. mansoni infection in this study area.
This study goes some way towards elucidating the risk factors associated with S. mansoni infection among schoolchildren in Mbita district, western Kenya. The prevalence of S. mansoni was high in almost all the schools sampled and more than three quarters (76.8%) of children were infected with S. mansoni. This result is consistent with the previous reports [32], [44], in which they showed an increased risk of S. mansoni infection compared to the early 1970s [45], [46], at which time the prevalence of S. mansoni was less than 50% among schoolchildren along the shores and islands of Lake Victoria in Mbita.
In a bivariate NB-GLM analysis, children living in permanent houses with latrine were infected with larger numbers of S. mansoni eggs. However, the statistical significance of these effects did not remain in the model when considering school effects as a random factor (NB-GLMM). This was considerable based on previous studies which showed higher infection risk associated with lower socioeconomic status [47]. We attribute this to the fact that families living in permanent houses with latrine tended to be found in densely populated areas where infection risk was high, since the local habit of defecation along lake shore is common even though the most of families living around town centre have latrines [28]. In the result, the residential location was closely associated with S. mansoni infection risk. This finding corroborates previous research by Booth and colleagues, which clearly indicated that environmental living circumstances were tightly connected with infection status and disease burden. In short, environmental exposure due to residential location rather than some fixed characteristics of an individual determines risk of infection [48]. Several studies have reported high risks of S. mansoni infection among people living close to a permanent water body [25], [29]–[31]. However, the effect of the distance to the lake was not significant in the present study. This could be the result of a small range of proximity to the lake, as the schools surveyed were all located within 1.0 km, and the children lived within 2 km, of Lake Victoria.
Population density was the single most important factor associated with S. mansoni infection risk on the shores and islands of Lake Victoria. Theoretically, the basic reproductive number (R0) of schistosomiasis linearly increases with human density. This is due to the fact that the rate of infection among snails depends on the absolute number, not the prevalence of infected hosts [49]. Thus higher infection risk in densely populated areas can be explained purely by numerical dynamics of transmission. In addition, higher nutritional load via domestic waste water from densely populated areas might enhance population growth of the snails [50]. We can therefore strongly suggest that the increase in population density in recent decades may partially explain the increase in S. mansoni prevalence in this area.
Notably, our study revealed the highest prevalence and intensity of S. mansoni infection was around the Mbita causeway. There was a tendency that infection risk decreased towards the eastern part of the mainland (Figure 2). Additionally, infection risk was very high in the three schools on Rusinga Island but not on Ngodhe Island. Our results suggest that a simple dichotomy like island-mainland comparison may obscure micro-geographical heterogeneity in S. mansoni transmission. This calls for additional ecological and environmental survey to understand the distribution and population dynamics of snail intermediate host which directly relates with the transmission of schistosomes. Spatial analysis indicated a high-risk cluster that includes the town center, the causeway and nearby villages in Rusinga. The high risk of infection in these areas could be partially explained by local population density. However, a significant high-risk cluster remained in a similar location even after adjustment for the effects of local population and other potential risk factors. Therefore, an aggregated risk factor that was not measured in the present study may exist in the west side of Mbita causeway. The construction of Mbita causeway in the 1980s has likely impacted the ecosystem surrounding Rusinga and the mainland, by promoting population activities, restricting water circulation and free movement of aquatic biota through blockage of the natural channels [51]. Such a change in ecological conditions may be one of the reasons why S. mansoni prevalence has drastically increased compared with 1970s, before the causeway construction [52],[53].
To conclude, increased risk of S. mansoni infection was observed in Mbita along the shores and islands of Lake Victoria. Moreover, the infection risk of S. mansoni was associated with high population density and was concentrated around the Mbita causeway. Urgent intervention efforts should be considered in order to reduce morbidity and mortality due to S. mansoni infection, taking into consideration region-specific risk factors for disease transmission.
|
10.1371/journal.pbio.1001474 | Strigolactone Can Promote or Inhibit Shoot Branching by Triggering Rapid Depletion of the Auxin Efflux Protein PIN1 from the Plasma Membrane | Plants continuously extend their root and shoot systems through the action of meristems at their growing tips. By regulating which meristems are active, plants adjust their body plans to suit local environmental conditions. The transport network of the phytohormone auxin has been proposed to mediate this systemic growth coordination, due to its self-organising, environmentally sensitive properties. In particular, a positive feedback mechanism termed auxin transport canalization, which establishes auxin flow from active shoot meristems (auxin sources) to the roots (auxin sinks), has been proposed to mediate competition between shoot meristems and to balance shoot and root growth. Here we provide strong support for this hypothesis by demonstrating that a second hormone, strigolactone, regulates growth redistribution in the shoot by rapidly modulating auxin transport. A computational model in which strigolactone action is represented as an increase in the rate of removal of the auxin export protein, PIN1, from the plasma membrane can reproduce both the auxin transport and shoot branching phenotypes observed in various mutant combinations and strigolactone treatments, including the counterintuitive ability of strigolactones either to promote or inhibit shoot branching, depending on the auxin transport status of the plant. Consistent with this predicted mode of action, strigolactone signalling was found to trigger PIN1 depletion from the plasma membrane of xylem parenchyma cells in the stem. This effect could be detected within 10 minutes of strigolactone treatment and was independent of protein synthesis but dependent on clathrin-mediated membrane trafficking. Together these results support the hypothesis that growth across the plant shoot system is balanced by competition between shoot apices for a common auxin transport path to the root and that strigolactones regulate shoot branching by modulating this competition.
| Plants can adapt their form to suit the environment in which they are growing. For example, genetically identical plants can develop as a single unbranched stem or as a highly ramified bush. This broad developmental potential is possible because the shoot system is produced continuously by growing tips, known as shoot meristems. Meristems produce the stem and leaves of a shoot, and at the base of each leaf, a new meristem is formed. This meristem can remain dormant as a small bud or activate to produce a branch. Thus, the shoot system is a community of shoot meristems, the combined activity and inactivity of which shape shoot form. Here we provide evidence that growth is balanced across the Arabidopsis shoot system by competition between the shoot meristems. This competition is likely mediated by the requirement of meristems to export the plant hormone auxin in order to activate bud outgrowth. In our model, auxin in the main stem, exported from active branches, can prevent auxin export by dormant buds, thus preventing their activation. Our findings show that a second hormone, strigolactone, increases the level of competition between branches by making auxin export harder to establish. Together, these hormones balance growth across the shoot system, adjusting it according to the environmental conditions in which a plant is growing.
| Plants can alter their body plan to adapt to the environment in which they are growing (reviewed in Leyser 2009 [1]). This is possible because plant development is continuous, with postembryonic development being dependent on the activity of meristems. For example, the primary shoot apical meristem is laid down during embryogenesis at the apical embryonic pole, and after germination, the meristem gives rise to the adult shoot system through the production of a series of phytomers consisting of a leaf, a segment of stem, and a new shoot apical meristem, established in the axil of each leaf. These axillary meristems can remain dormant, or they can activate to produce a new shoot axis, with the same developmental potential as the primary shoot. Thus the mature shoot system can range from a solitary stem to a highly ramified bush, depending on the activity of the axillary meristems. The large number of meristems in the shoot system allows the plant to recover quickly from damage and to adjust its growth according to spatially heterogeneous environmental inputs such as unilateral shading, and to systemic inputs such as the nutrient status of the plant. Thus multiple inputs are integrated to balance growth across the shoot system.
Axillary meristem activity is controlled by a network of systemically moving endogenous signals, among which auxin plays a pivotal role. Auxin, synthesized principally in the young expanding leaves of growing shoot tips, moves rootward in the stem through the polar auxin transport stream (PATS). The direction of the PATS is determined by the polar localisation of PIN-FORMED (PIN) plasma membrane (PM) auxin efflux carriers [2],[3]. In the stem, efficient rootward auxin flow requires PIN1 [4], which is basally localised in the PM of xylem parenchyma cells [5]. Auxin in the PATS inhibits the outgrowth of axillary buds. Pharmacological inhibition of the PATS or removal of the primary shoot apex triggers outgrowth of axillary buds, and application of auxin to the decapitated stump prevents this outgrowth [6],[7]. However, direct application of auxin to axillary buds does not prevent their outgrowth [8], and apically applied auxin is not transported into the axillary buds [9], suggesting that auxin in the PATS inhibits shoot branching indirectly.
Two nonexclusive mechanisms have been proposed to account for the indirect action of auxin. Firstly, it has been proposed that auxin regulates the synthesis of one or more second messengers, which move up into the axillary buds to regulate their activity. Two classes of phytohormone, cytokinins and strigolactones, are good candidates for these signals. Cytokinins can move up the plant in the transpiration steam in the xylem. Direct application of cytokinin to axillary buds can induce outgrowth, even in an intact plant [10]. Decapitation elevates but auxin application reduces endogenous cytokinin levels in xylem sap [11] and in the stem of nodal explants [12]. Together these data suggest that auxin inhibits bud outgrowth in part by reducing systemic and local cytokinin levels, and thus cytokinin supply to buds. A similar dataset exists for strigolactones. They can also be transported up the plant in the xylem [13]. Their direct application to buds can inhibit outgrowth on intact and decapitated plants [14], and decapitation reduces but auxin application elevates the transcription of strigolactone biosynthetic genes [15],[16]. These data suggest that auxin inhibits bud activity in part by increasing systemic and local strigolactone synthesis and thus strigolactone levels in buds.
However, strigolactones only inhibit shoot branching in the presence of a competing auxin source, such that supply to a solitary bud has little or no effect and supply to an explant carrying two buds inhibits only one of the buds, which can be either the more apical or more basal bud [17],[18]. Furthermore, strigolactone addition results in a reduction in PIN1 levels in xylem parenchyma cells within 6 h, accompanied by a reduction in polar auxin transport [17]. Thus in strigolactone biosynthetic mutants, high levels of branching correlate with high levels of PIN1 and polar auxin transport and high auxin concentration in the main stem [19],[20]. These observations have led to an alternative model both for strigolactone action and for the indirect mode of inhibition of axillary bud growth by auxin in the PATS in the main stem.
This alternative model derives from considerations of the auxin transport canalization hypothesis, originally proposed to explain vascular pattern formation. The central tenet of the canalization hypothesis is positive feedback between auxin flux and auxin flux capacity [21]. Restated in terms of PIN proteins, an initial passive flux of auxin from an auxin source to an auxin sink results in the up-regulation and polarisation of PINs in the direction of the flux. This results in files of cells with high levels of PINs polarised in the direction of the sink, some of which may differentiate into vascular strands. The emergence of such files between an auxin source and sink has been directly observed [22],[23].
Given that active axillary buds are sources of auxin [23],[24], and the main stem can act as an auxin sink, by carrying auxin away to the root, auxin transport canalization can act to connect the bud to the stem, transporting auxin away from the bud apex and establishing vascular connectivity between the bud and the rest of the plant. However, high auxin levels in the main stem can prevent canalization of auxin transport out of the bud by reducing stem sink strength for auxin, limiting the initial flux of auxin out of the bud, thereby preventing escalation of the positive feedback at the heart of the canalization process [20],[23]. If auxin transport canalization out of the bud is required for bud activity, then this could explain the indirect inhibition of buds by auxin in the main stem, without the need for a second messenger relaying the auxin signal into the bud. Instead, buds and the main shoot apex compete for access to a common auxin transport pathway down to the root. Computational simulations of this model demonstrate its plausibility [20]. Moreover the model can explain the association of high branching with high PIN1, auxin transport, and auxin levels observed in strigolactone mutants, by postulating that the mode of action of strigolactone is to reduce the accumulation of PIN1 on the PM, thus making canalization more difficult to achieve, requiring a higher initial flux of auxin to drive the positive feedback loop. The model also explains the requirement for a competing auxin source for strigolactone-mediated bud inhibition [17],[18].
One attractive feature of this model is that it establishes a regulatory framework underpinning the ability of plants to balance growth across the shoot system, integrating local (e.g., light quality) and systemic (e.g., nutrient limitation) information, through bud–bud competition. However, the validity of this model remains controversial because of the substantial body of evidence consistent with the hypothesis that strigolactones act locally and specifically in buds to inhibit their growth, by up-regulating the transcription of the TCP family transcription factor, BRC1, which is known to be required for stable bud inhibition [14],[25],[26].
In this article, we use computational modelling to generate predictions that allow these alternative hypotheses for strigolactone action to be distinguished. Our results strongly support the auxin transport canalization model for shoot branching control. Specifically, we demonstrate that strigolactone treatment can either inhibit or promote shoot branching, depending on the auxin transport status of the treated plants. This is difficult to reconcile with direct bud inhibition by strigolactone. In contrast, these responses can be explained if strigolactones act by regulating PIN1 removal from the PM of cells in the shoot. Consistent with this mode of action, we show that a rapid primary response to strigolactone is clathrin-dependent PIN1 depletion from the PM.
The auxin transport canalization-based model for shoot branching control places the auxin transport network as a central component of systemic growth co-ordination in plants. To test this idea further, we investigated the relationship between auxin transport, PIN1 accumulation, and shoot branching in Arabidopsis mutants. In roots, many PM proteins, including PIN1, cycle between the PM and endocytic compartments (reviewed in [27]). This process involves GNOM (GN), a Brefeldin A (BFA)–sensitive ARF–GEF that mediates exocytosis [28], and TRANSPORT INHIBITOR RESISTANT3 TIR3/BIG [29],[30], a putative E3 ligase [31] that co-localises in detergent-resistant membrane fractions with PIN1 and the 1-N-naphthylphthalamic acid-binding auxin transporter, ABCB19 [32]. Mutation in GN or TIR3 causes increased shoot branching soon after floral transition [20],[33], whereas mutation in ABCB19 and its homologue ABCB1 have no discernible effect at this stage [34].
To investigate the relationship between increased shoot branching, strigolactone action, and PM PIN1 accumulation, we compared gn, tir3, and the strigolactone-signalling mutant max2 in single and double mutant combinations for these phenotypes (Figure 1). PM accumulation of PIN1 was assessed in hand sections through main inflorescence stems of 6-wk-old plants harbouring a PIN1:PIN1-GFP transgene. All the mutants tested show increased overall fluorescence in xylem parenchyma cells (Figure 1A,B). For gn and tir3, reduced PM PIN1 was associated with reduced auxin transport and increased shoot branching (Figure 1A–D). The reduction in auxin transport observed in gn and tir3 is of a similar magnitude to that reported by Okada et al. [4] for the pin1 mutant and confirmed in our conditions (Figure S1). In contrast, max2 had increased shoot branching, with increased PM PIN1 and increased auxin transport (Figure 1A–D), consistent with previous reports [19]. Double mutants between these two classes had at least partially additive phenotypes (Figure 1A–D), with higher shoot branching than the single mutants, and intermediate levels of auxin transport and PM PIN1, except in the max2 tir3 double mutant, where PM PIN1 levels were similar to max2. These results suggest that while GN, TIR3, and strigolactones are all involved in PM PIN1 accumulation, their modes of action are at least partially independent.
Many biological behaviours are the outcome of interlinked feedback regulation acting recursively. Consequently, these behaviours are difficult to understand by intuitive interpretation of biological observations. Formalisation of these systems through mathematical modelling and computer simulation can link mechanistic hypotheses for their action to emergent higher order behaviour and thereby increase understanding of the underlying mechanisms. We previously presented a computational model for shoot branching control, based on the auxin transport canalization hypothesis described above [20]. This model can account for the phenotypes of gn or tir3 mutation, and strigolactone treatment, if their actions are to reduce insertion or enhance removal of PIN1 from the PM [20]. The heart of the model is Equation 1, which encapsulates the positive feedback of auxin transport canalization. PIN1 levels in the membrane depend on both insertion, captured by a rate (ρ) proportional to the flux of auxin across the membrane, and removal, captured by a rate (μ) (for full details, see [20]):(1)
To dissect further which parameters in this model might be affected by max, gn, and tir3 mutation, we set “wild-type” values of the parameters and ran simulations with individual input values for each parameter in turn, changed around the wild-type value. The simulation outputs are summarised for shoot branching levels, polar auxin transport levels, and PIN protein levels in Table 1. Of the 14 parameters, 13 were able to capture branchy phenotypes with some input values. Of these, only three captured both branchy phenotypes and altered levels of polar auxin transport. These were ρ (the PIN insertion constant), μ (the PIN removal constant), and T (the polar transport coefficient—the efficiency with which each PIN protein transports auxin). To match the biological data, GN and TIR3 activity should be explained by a parameter whose reduction can elevate branch numbers, reduce polar auxin transport, and reduce PIN1 accumulation (Figure 1). Only ρ (the PIN insertion constant) satisfies these criteria (Table 1). Similarly, strigolactone/MAX activity should be explained by a parameter whose reduction can increase shoot branching, polar auxin transport, and PIN1 accumulation (Figure 1). Only μ (the PIN removal constant) satisfies these criteria (Table 1).
To understand better the relationship between the parameters and simulation outputs, we plotted two 3-dimensional graphs that show PAT (Figure 2A) and shoot branching (Figure 2B) levels as heights on the μ–ρ plane. The relationship between polar auxin transport levels and μ–ρ was relatively simple: as PIN removal (μ) decreased and PIN insertion (ρ) increased, the polar auxin transport level gradually increased, resulting in a smooth slope (Figure 2A,C,D). In contrast, the relationship between shoot branching level and μ–ρ was more complex: as PIN removal (μ) decreased, the shoot branching level increased, creating a plateau of high branching at low μ values. However, as PIN insertion (ρ) decreased the branching level increased, even when PIN removal (μ) was quite high, resulting in a ridge of high branching (Figure 2B). High branching on the low μ (low PIN removal) plateau is caused by easy establishment of canalization of auxin transport from bud to stem, with low initial auxin fluxes able to establish canalization through positive feedback, making buds difficult to inhibit. High branching along the low ρ (low PIN insertion) ridge is caused by low auxin efflux from active shoot apices, such that a larger number of active apices are needed to supply sufficient auxin to the main stem to prevent activation of further buds. The profiles for branch number at any one μ or ρ value made much more abrupt transitions than for auxin transport levels (Figure 2C,D), with mostly high or low branch numbers, and only narrow regions of parameter space giving intermediate branch numbers. This is because branch activation in the model is triggered by canalization of auxin transport out of the simulated bud and the positive feedback inherent in the canalization process produces switch-like behaviour [20].
To capture the behaviour of strigolactone biosynthesis mutants such as max4 or strigolactone-signalling mutants such as max2, we assigned a low value to μ, conditioning slow PIN removal. This resulted in higher levels of both polar auxin transport and branching compared with those of the defined wild-type (Figure 2A,B), consistent with biological results (Figure 1 and [17],[19]). Similarly we simulated the gn or tir3 mutations as a low ρ value, conditioning low PIN insertion, resulting in a lower level of polar auxin transport and a higher level of branching (Figure 2A,B), as observed in biological experiments (Figure 1 and [20],[33],[35]). To simulate addition of the synthetic strigolactone, GR24, we increased the value of μ (increasing PIN removal), which gave slightly lower polar auxin transport and shoot branching levels compared to the defined wild-type (Figure 2A,B), consistent with published biological data [17]. When the low μ value of max and the low ρ value of gn or tir3 were simultaneously applied, the model predicts moderate polar auxin transport levels and high branching, consistent with biological results (Figure 1 and [20]). Thus, single parameter changes in the model capture the phenotypes of wild-type, single and double mutants, and where known, their responses to GR24. Furthermore, the relative magnitude of the responses to GR24 in different genetic backgrounds and with respect to branching versus auxin transport is also captured.
This analysis led to an interesting and counterintuitive prediction. The dose-response curve of max4 tir3 branch number to GR24 is predicted to have two peaks, which lie on the low PIN removal (μ) plateau and low PIN insertion (ρ) ridge (Figure 2B). To test this prediction, we grew wild-type, max4, tir3, and max4 tir3 plants for 8 wk on agar-solidified medium supplemented with GR24 ranging from 10 nM to 1 µM (Figure 3A). As previously shown [17], in both wild-type and max4, GR24 reduced branching levels monotonically, although this effect was not statistically significant in the wild-type. In contrast, in tir3, GR24 significantly elevated branching levels at 10 nM, and reduced branching at higher concentrations, with 1 µM resulting in very poor growth. In max4 tir3, 10 nM GR24 reduced branching levels, but 50 nM GR24 restored branching levels to those of untreated plants and higher concentrations reduced them again. This latter part of the curve was shifted compared to the tir3 alone, with branched plants produced at 1 µM, a concentration that severely inhibits growth in tir3 mutants. Therefore, GR24 did not simply inhibit but also promoted shoot branching depending on the concentration and the genetic background of the treated plant. These results validate the predictions of the model with the minor modification that the effects of tir3 mutation on PIN insertion (ρ) suggest that it is placed on the low μ slope of the low ρ ridge, rather than at its summit, as proposed in Figure 2.
As well as the unusual dose–response relationships, the parameter space exploration predicts no branching at high PIN removal (μ) and low PIN insertion (ρ), caused by insufficient auxin transport to support bud growth. In the dose–response experiments described above, 1 µM GR24 severely affected the growth of the tir3 mutant. To explore the response of tir3 and gn mutants to high levels of GR24 in more detail, we grew wild-type, max4, gn, and tir3 plants for 8 wk on agar-solidified medium containing 5 µM GR24, or an equivalent volume of solvent. GR24 affected the overall vigor of gn and tir3 plants, such that their total dry weights were significantly reduced compared to untreated controls (Figure 3B,C). This effect was particularly noticeable in tir3 plants (Figure 3B,C), which often did not survive to maturity in the presence of 5 µM GR24. GR24 had no effect on dry weight in wild type or max4 (Figure 3C). Thus gn and tir3 shoots are hypersensitive to GR24.
These data strongly support the hypothesis that strigolactones increase the removal of PIN1 from the PM, and indeed we have previously shown that GR24 treatment reduces PIN1 abundance in xylem parenchyma cells within 6 h in a MAX2-dependent manner [17]. To investigate the dynamics of this process in more detail, we prepared hand sections of stems of different genotypes harbouring the PIN1–GFP fusion, as described above, and recorded basal PM PIN1 levels every 10 min over a 90-min period. PIN1 was significantly reduced by the addition of 5 µM GR24 within 40 min in wild-type plants and within 30 min for max1 plants (Figure 4A). As expected, no significant difference was observed in max2 mutants (Figure 4A). We also examined wild-type sections treated with 50 µM cycloheximide for 30 min before a 60-min incubation with 5 µM GR24 or the vehicle control. GR24-induced depletion of PM PIN1 level was unaffected by cycloheximide treatment (Figure 4B), suggesting that this process is independent of new protein synthesis.
Depleted PM PIN1 could in principle result from either increased removal or reduced insertion of PIN1. In roots, there is good evidence that many membrane proteins, including PIN1, cycle rapidly between the PM and endomembrane compartments [28]. The removal of these proteins from the PM is mediated by clathrin-dependent endocytosis, which is often assessed by quantifying the accumulation of BFA-induced endomembrane compartments [36],[37]. BFA inhibits the activity of ARF–GEFs such as GN, preventing recycling of proteins back to the PM, resulting in their depletion from the PM and accumulation in endomembrane compartments. We treated stem segments with 50 µM BFA for 3 h, but we observed that this treatment had no significant effect on the amount of PIN1 on the basal PM (Figure 5A), and relatively few PIN1 containing compartments were identified. Only 9 of 29 cells examined with optical sectioning throughout the z-axis contained a compartment. This contrasts to results previously described for PIN1 in roots, where after 90 min treatment with 25 µM BFA, the mean number of BFA compartments per cell was more than 1 [37]. These results suggest that either PIN1 endocytosis in stems is BFA-sensitive (for tissue-dependent BFA effects, see [38]) and/or PIN1 cycles only slowly in stem segments. To assess the rate of PIN1 allocation to the PM, we used fluorescence recovery after photobleaching. In root cells, after photobleaching total PIN1 signal from a cell, nonpolar PM PIN1 was detected after 100 min [39]. We bleached only the basal PM PIN1 of xylem parenchyma cells, and no significant fluorescence recovery was detected 90 min after bleaching, with little visible effect even after 3 h (Figure 5B), suggesting low insertion rates for PIN1 from either intracellular stores or de novo synthesis. This suggests that at steady state, either cycling rates in stems are low or the fraction of PIN1 in intracellular compartments is very low.
To test whether GR24-triggered PIN1 depletion is clathrin-dependent, we determined the effect of the clathrin inhibitor, A23 [40]. A23 treatment alone had no effect on PIN1 levels, providing further evidence for a low rate of insertion of PIN1 or a low intracellular fraction. However, in the presence of A23, including a 30-min pretreatment, the ability of GR24 to deplete PM PIN1 was abolished (Figure 5C), whereas when treated with the structurally related but inactive control A51, GR24 triggered PIN1 depletion from the PM as previously observed (Figure 5C). These results suggest that a rapid, nontranscriptional mode of action of strigolactone is to promote a clathrin-mediated step in PIN1 depletion. Indeed in this experiment, statistically significant depletion of basally localised PIN1 was observed within 10 min.
As described above, in roots there is rapid constitutive cycling of PIN1 between the PM and the endomembrane system. However, this cycling is not specific but rather reflects general cycling of many proteins. Treatments that affect PIN1 levels at the PM, such as auxin, BFA, and A23 treatment, also affect many other membrane proteins, such as water channel proteins of the PIP1 and PIP2 families [28],[36],[37]. Thus in the root, a major contributor to PIN1 behaviour is general trafficking activity. We therefore tested the specificity of the effects of GR24 on PIN1 PM levels in shoots by assessing its effects on PIP1 [41]. PIP1 levels on the PM were less stable over time than PIN1 levels and were halved after 90 min, regardless of the presence or absence of GR24 (Figure 5D), indicating that GR24 has no effect on PIP1 levels. These results suggest that the effect of strigolactone on PIN1 PM levels in stems is more specific than known mechanisms regulating PM PIN1 levels in roots.
The apparent specificity of strigolactone effects on PM PIN1 in shoots raises interesting questions concerning the effect of strigolactones on roots. Various aspects of root development, such as primary root length, lateral root development, and root hair elongation, have recently been shown to be modulated by strigolactones [42]–[44]. The role of auxin transport in these phenotypes is unclear, but there is some evidence to suggest that they are at least in part mediated by differences in auxin transport, either locally in the root or systemically from the shoot.
To investigate the relationship between auxin transport and the effects of GR24 on roots, we grew Arabidopsis seedlings for 3 d on agar medium without exogenous hormones, preincubated them for 24 h with various concentrations of GR24, and then observed their elongation over the next 24 h. In the wild-type, two responses were found: agravitropic root growth and root growth inhibition (Figure 6; also see [42]). With respect to effective doses, agravitropic root growth required very high concentrations of GR24 (greater than 10 µM, Figure 6A), and there was no significant difference between wild-type and max2 mutants in response to 100 µM GR24 (Figure 6B). The very high levels of GR24 needed for this effect and lack of dependence on MAX2 led us to conclude that it is of limited physiological relevance. In contrast, as previously shown [42], root elongation was inhibited by more physiologically relevant levels of GR24, with GR24 levels between 3 and 30 µM having a significantly weaker effect on max2 than on wild-type, although dose-dependent inhibition was observed in both these genotypes (Figure 6C). Thus, root growth inhibition by GR24 is partially MAX2-dependent.
To assess the involvement of auxin in GR24-induced root growth inhibition, we measured root growth in seedlings treated with 0.1 µM 2,4-D, a synthetic auxin, or 5 µM GR24, comparing wild-type, max2, gn, tir3, and the auxin signalling mutants axr1 and tir1 (Figure 6D) [45],[46]. The max2 mutant responded to 2,4-D as wild-type and showed resistance to GR24; axr1 and tir1 showed resistance to 2,4-D and responded normally to GR24; gn and tir3 responded normally to 2,4-D but showed mild hypersensitivity to GR24. Thus, as in the shoot, there is an interaction between GR24 and GN/TIR3.
To test the effects of GR24-treatment and gn and tir3 mutation on PIN1 protein levels in roots, we observed the root tips of 4-d-old Arabidopsis seedlings harbouring a PIN1:PIN1–GFP transgene in either the wild-type, gn, or tir3 genetic backgrounds after a 12-h incubation with or without 10 µM GR24. Neither GR24-treatment, gn, nor tir3 altered total signal levels or obvious subcellular localisation of PIN1 protein in the root tip (Figure 6E,F). Even after a 48-h incubation, 10 µM GR24 did not alter total signal levels or obvious subcellular localisation in wild-type (Figure 6G,H). These results are consistent with different PIN1 trafficking dynamics in roots compared to shoots, such that relatively modest increases in strigolactone-triggered PIN1 PM depletion have a much more dramatic effect in the shoot compared to the root.
In the 1930s Thimann and Skoog established that auxin synthesized in active shoot apices is transported down the main stem and inhibits the activity of axillary shoot apices in subtending leaf axils [6],[7]. However, it was rapidly discovered that auxin acts indirectly to inhibit axillary bud growth, and furthermore there was a fundamental paradox in auxin behaviour. On the one hand, auxin inhibited the activity of axillary buds, but on the other, its synthesis and export from active apices protected them from inhibition by other auxin sources [47]. These classical observations are explicable by the auxin transport canalization based model for shoot branching control. According to this idea, all the meristems in a shoot compete for access to a common auxin transport path down the main stem to the root. Rootward auxin transport from each shoot apex is established by the positive feedback process of auxin transport canalization, the dynamics of which are critically dependent on the strength of the bud as an auxin source, the strength of the stem as an auxin sink, and the dynamics of the positive feedback loop at the centre of the canalization process that connects them. Thus, the auxin transport system in the shoot forms a self-organising network through which all shoot apices communicate by contributing auxin into the system, thereby influencing the ability of other apices to export auxin.
This mechanism for shoot branching control is attractive because it explains the classical observations mentioned above and readily supports the integration of both local and systemic factors in balancing growth distribution across the shoot. However, the idea remains controversial, largely due to different ideas about the mechanism of action of another branch-regulating hormone, strigolactone. One hypothesis, generally referred to as the second messenger hypothesis, posits that auxin in the main stem up-regulates the production strigolactone, which moves into the axillary buds and inhibits their growth by locally up-regulating transcription of the TCP family transcription factor BRC1, which is known to be required for stable bud inactivation [14],[25],[26]. A second hypothesis assumes that axillary bud activity is regulated by the auxin transport canalization-based mechanism described above and that strigolactone acts by modulating auxin transporter accumulation, thereby modulating the ease with which axillary buds can establish active auxin transport into the main stem (Figure 7) [17],[19],[20],[23]. Thus the mechanism of strigolactone action and the mechanism of auxin-mediated bud inhibition are tightly intertwined, representing two different scenarios for the systemic coordination of growth across the shoot system.
The results presented here strongly support the second hypothesis. A particularly striking illustration of this is the ability of strigolactone to promote shoot branching in the tir3 mutant background (Figure 3), which is difficult to explain if strigolactones act as direct inhibitors of bud growth but is a prediction of the model in which strigolactones act to modulate auxin transport (Figure 2). It should be noted that the two models can easily be reconciled. For example, the primary mode of action for strigolactone could be on PIN1 accumulation, and the resulting effects on auxin transport could in turn influence BRC1 transcript levels. Up-regulation of BRC1 by strigolactone addition to pea buds has been shown to be independent of new translation, but so far it has only been measured after 6 h [26],[48], and no such up-regulation was detected in a similar experiment in rice after 3 h of treatment [49]. In contrast, in Arabidopsis stems, an effect on PIN1 accumulation was observed within 10–40 min of strigolactone application (Figures 4 and 5), and this effect is also independent of new protein synthesis. It is therefore possible that BRC1 transcript changes are downstream of changes in PIN1 accumulation, and the role of BRC1 could be to stabilise bud inactivation caused by low auxin export. Some stabilizing system to maintain bud inactivity seems intuitively important, because bud activation by the positive feedback inherent in canalization is highly likely to be triggered by stochastic variation in the system.
These two models differ in that in the canalization model, strigolactones act systemically on the auxin transport network, including in the bud (Figure 7), whereas in the second messenger model, they act locally and specifically in buds. The systemic expression of MAX2 in xylem-associated cells and the effect of strigolactone on PIN1 accumulation in the main stem are consistent with systemic action. This mode of action allows strigolactones to modulate bud–bud competition systemically, for example in response to nutrient deprivation [13]. In this context, systemic strigolactone levels determine how many buds can activate, but they do not determine which buds activate. This can be regulated by local factors such as light levels [50]. Thus, both local and systemic modifications to the auxin transport network can integrate different environmental inputs to direct resource allocation across the plant body. A more direct mode of action for strigolactone locally in buds does not have this interesting property. However, the two models, and indeed others, are mutually compatible and could operate in parallel with either species-specific and/or environment-specific variation in their relative importance.
Little is known about the molecular mechanism of strigolactone action. Only two genes have been implicated in strigolactone signalling. These are MAX2, which encodes an F-box protein presumed to be required for the strigolactone-regulated ubiquitination of one or more specific target proteins, and D14, which encodes an α/β hydrolase protein that binds GR24, confers signal specificity to the pathway [51]–[53], and could either act as a receptor or could process strigolactones to form a final bioactive product. The immediate downstream effectors of the pathway are unknown, but the largely nuclear localization of MAX2 [54] and the rapid changes in transcription induced by many F-box-protein–mediated plant hormone signalling pathways [55] have led to an assumption that the primary targets for the strigolactone pathway are also transcriptional. The evidence to support this mode of action is currently quite weak. Few reliable transcriptional readouts for strigolactone response have been identified. These tend to have slow induction kinetics, in the order of several hours, and relatively small fold inductions [26], suggesting that they may be secondary responses or limited to a small proportion of cells. Microarray analysis of Arabidopsis seedlings treated with or without 1 µM GR24 for 90 min shows that 76% of all the GR24-repressible genes are categorised as auxin-inducible [56], and thus these transcriptional effects may be indirectly mediated via changes in auxin distribution.
Consistent with this idea, we have shown that a rapid translation-independent response to stigolactone addition is changes in PM PIN1 accumulation (Figure 4). Thus, at least one immediate early target downstream of MAX2 in the stem is not transcriptional but involves PIN1 depletion from the PM by an A23-sensitive mechanism, such as clathrin-mediated endocytosis [40]. The mechanism by which the substantially nuclear MAX2 influences PM PIN1 is not known. However, our data suggest that it is both quantitatively and qualitatively different from the major PIN1-regulatory systems operating in the root. Several lines of evidence support this conclusion. First, in stems, strigolactone response is independent of TIR3 activity, which has been reported to be required for auxin-induced inhibition of clathrin-mediated endocytosis in roots [30]. Second, the effect of strigolactone on PIN1 depletion from the PM in stems appears to be more specific than the systems operating in roots, since it does not affect the PM levels of PIP1, although we have not excluded targets beyond PIN1. Third, the MAX2-dependent effects of strigolactones on root phenotype are generally less dramatic than those observed in shoots, both with respect to cell biological and whole organ-level phenotypes.
Although more modest than the effects on shoots, long-term effects of GR24 treatment on PIN1 accumulation in the root tip have been detected following 6 d of growth in the presence of 5 µM GR24 [44]. These effects have been correlated with reduced shoot-to-root auxin transport, suggesting that they represent a transcriptional response to low auxin rather than the protein trafficking mechanism we propose here. Consistent with this idea, the accumulation of multiple PIN proteins is affected in these root tips, including in cell layers where MAX2 is not expressed at detectable levels [44],[54]. However, although we found only weak MAX2-dependent root growth inhibition by GR24, this occurred with equal effect in the auxin signalling mutants, axr1 and tir1, suggesting that GR24 reduces root growth at least to some extent independently of auxin concentration-mediated effects. Similarly, in the trafficking mutants, gn and tir3, GR24 reduced root growth more severely than in wild-type, suggesting some overlap in the mechanism underlying the control of shoot branching by strigolactone and its effects on root growth.
Computer simulations of shoot growth using our canalization-based model consistently reproduce biological results when strigolactone action is ascribed to a linear process of PIN removal from the PM, independent of PIN insertion and auxin flux. Consistent with this idea, bioimaging of PIN1 protein in inflorescence stems revealed a substantial increase in PIN1 protein in the basal PM in strigolactone mutants. Furthermore, GR24 promoted rapid, translation-independent, MAX2-dependent depletion of PIN1 from the PM through a mechanism sensitive to A23, an inhibitor of clathrin-mediated membrane trafficking. These results are consistent with the hypothesis that strigolactone functions to promote endocytosis of PIN1 from the PM.
It is interesting that the phenotypes affected by the max mutations and by strigolactone treatment are generally those where auxin transport canalization has been implicated. In the root tip, canalization is not usually considered to play an important role in PIN accumulation, although auxin-induced changes in the lateralisation of PIN1 in the root endodermis have been described and compared to canalization processes [22]. If the effects of strigolactones on auxin transport are specifically to modulate canalization, then they provide an opportunity to understand better this enigmatic and poorly understood process, which nonetheless provides powerful explanations for complex patterning events in plants and for their impressive developmental plasticity.
All lines are in the Col-0 background. Experiments involving max2, used max2-3 [57], max4, max4-1 [15], axr1, axr1-3 [58], tir1, tir1-1 [46], gn, gnomB/E [59], and tir3, tir3-101 [60]. Because we found that the tir3-101 line from a public stock had an additional glabrous mutation besides a C-to-T nonsense mutation at the 3,095th codon of TIR3, a tir3-101 line free from the additional mutation was made and used. For bioimaging, each line homozygous for the PIN1:PIN1–GFP transgene cassette [61] was used. For the PIP1 experiments, the UBQ10:PIP1–YFP (Wave138Y) fusion line was used [41]. On-soil and axenic growth conditions were as described previously [17].
For quantifying root growth inhibition and agravitropic root growth, axenic seedlings grown vertically for 3 d on hormone-free agar medium were preincubated for 24 h on vertically placed agar medium containing either only the vehicle, GR24, or 2,4-D. For evaluating the root growth inhibition, the root tip position was recorded just after the preincubation and 24 h after; thus, the length of the primary root grown for the 24 h was obtained. For evaluating the agravitropic root growth, preincubated seedlings were placed horizontally; the root tip position was recorded just after the gravistimulation and every hour up to the next 24 h; thus, the index Curvature, which we defined as the change in root tip angle per the length of grown root within a range between 1 and 3 mm, was calculated. Other physiological experiments were as described previously [17].
All simulations were according to the model of Prusinkiewicz et al. (2009) [20]. For simulating the auxin transport assay (Figure S2), the two most basal metamers in the main stem of the whole plant simulated for 2,000 time steps were used; of these two metamers, the top one provided an initial value of the PIN concentration at the basal face, and the bottom one provided an initial value of the PIN concentration at the apical face. Auxin concentration in the top metamer was assumed to be 10 and constant over time; auxin concentration in the bottom metamer was assumed to be zero initially and change over time according to the Equations 1 and 2 of Prusinkiewicz et al. (2009) [20]. The auxin concentration of the bottom metamer at time step 10 was calculated from those two initial values of the PIN concentrations and converted to the percentage of wild-type. This percentage is shown as the simulated polar auxin transport level.
For imaging PIN1–GFP in inflorescence stems, the most basal part of the primary inflorescence stem of 6-wk-old soil-grown plants was longitudinally halved by hand with a razor blade. The cut surface was immediately observed using light microscopy and a Zeiss LSM 710 confocal microscope to identify xylem parenchyma cells according to both the relative position to xylem vessels and the morphology of the cell. With excitation at 488 nm, images containing emission spectra from 490 to 655 nm were then acquired within a single dynamic range. Reference spectra of GFP and chloroplast autofluorescence were obtained using a PIN1:PIN1–GFP line in the wild-type background and were used for linear unmixing of the images. For its quantitative analysis, only xylem parenchyma cells that appeared intact and were exposed to the cut surface were taken into account, and the intensity of their unmixed GFP signal was measured in a region of the basal PM that was manually traced. Data were obtained in the same way for real-time monitoring experiments, except that sections were observed with Zeiss LSM 780 confocal microscope. With excitation at 488 nm, images containing emission spectra from 507–550 nm and 593–719 nm were acquired simultaneously in separate channels. Data were obtained in the same way for the PIP1 experiment, except excitation was at 514 nm, and emission spectra were acquired from 518–621 nm and 647–721 nm. For photobleaching experiments, a region of interest (basal PM of xylem parenchyma cell) was selected and bleached using the 488 nm laser at 50% power for 75 iterations. In all experiments, cells from three or more plants were included for each genotype/treatment, and the results presented are typical of at least two independent experiments.
For imaging PIN1–GFP in the root tip, 3- to 5-d-old seedlings incubated for 12 or 48 h on agar medium containing the vehicle or 10 µM GR24 were immersed in 10 µg/ml propidium iodide for 10 min. The primary root was then observed with Zeiss LSM 510 Meta confocal microscope. The GFP signal excited with a 488 nm laser and the propidium iodide signal excited with a 543 nm laser were collected with a 505–550 nm bandpass filter. For its quantitative analysis, the average intensity of the GFP signal was measured in the stele region of each root.
Based on the assumption that the root angle after gravistimulation and the number of branches do not always follow the normal distribution, nonparametric methods of Wilcoxon, Steel–Dwass, and Shirley–Williams were used. Otherwise parametric methods of Student, Tukey, Dunnett, and Williams were used. Unless otherwise stated, statistical results of two-tailed tests are shown in graphs in the conventional manner. In Steel–Dwass' and Tukey's tests, different letters denote significant differences at p<0.05. In other tests, no marks or n.s. indicate not significant, and significant differences are indicated by asterisks as follows: p>0.05; * p<0.05; ** p<0.01; *** p<0.001.
|
10.1371/journal.pntd.0001459 | Murine Dendritic Cells Transcriptional Modulation upon Paracoccidioides brasiliensis Infection | Limited information is available regarding the modulation of genes involved in the innate host response to Paracoccidioides brasiliensis, the etiologic agent of paracoccidioidomycosis. Therefore, we sought to characterize, for the first time, the transcriptional profile of murine bone marrow-derived dendritic cells (DCs) at an early stage following their initial interaction with P. brasiliensis. DCs connect innate and adaptive immunity by recognizing invading pathogens and determining the type of effector T-cell that mediates an immune response. Gene expression profiles were analyzed using microarray and validated using real-time RT-PCR and protein secretion studies. A total of 299 genes were differentially expressed, many of which are involved in immunity, signal transduction, transcription and apoptosis. Genes encoding the cytokines IL-12 and TNF-α, along with the chemokines CCL22, CCL27 and CXCL10, were up-regulated, suggesting that P. brasiliensis induces a potent proinflammatory response in DCs. In contrast, pattern recognition receptor (PRR)-encoding genes, particularly those related to Toll-like receptors, were down-regulated or unchanged. This result prompted us to evaluate the expression profiles of dectin-1 and mannose receptor, two other important fungal PRRs that were not included in the microarray target cDNA sequences. Unlike the mannose receptor, the dectin-1 receptor gene was significantly induced, suggesting that this β-glucan receptor participates in the recognition of P. brasiliensis. We also used a receptor inhibition assay to evaluate the roles of these receptors in coordinating the expression of several immune-related genes in DCs upon fungal exposure. Altogether, our results provide an initial characterization of early host responses to P. brasiliensis and a basis for better understanding the infectious process of this important neglected pathogen.
| Paracoccidioidomycosis is a systemic disease that has an important mortality and morbidity impact in Latin America, mainly affecting rural workers of Argentina, Colombia, Venezuela and Brazil. Upon host infection, one of the most important aspects contributing to disease outcome is the initial encounter of the Paracoccidioides brasiliensis fungus with dendritic cells. This phagocytic cell is specialized in decoding microbial information and triggering specific immune responses. Thus, using a molecular biology technique to examine the response of thousand of genes, we aimed to identify the ways in which murine dendritic cells interact with P. brasiliensis during an early time point following infection. This approach allowed us to recognize diverse modulated genes, in particular those associated with a proinflamatory response and fungal recognition. Our work provides an initial molecular characterization of early infection process and should promote further investigations into the innate host response to this important fungal pathogen.
| The thermodimorphic fungus Paracoccidioides brasiliensis is the causative agent of paracoccidioidomycosis (PCM), a systemic human disease that is geographically confined to Latin America. PCM is mainly endemic in Argentina, Colombia, Venezuela and especially in Brazil, where it is the most prevalent cause of death among systemic mycoses not associated with AIDS [1].
P.brasiliensis infection is acquired upon the inhalation of airborne propagules derived from the saprophytic mycelium form of the fungus. Once in the lungs, P. brasiliensis converts to its parasitic yeast form and interacts with resident macrophages and dendritic cells (DCs) [2], [3]. DCs are the most powerful antigen-presenting cells and are uniquely able to recognize pathogen-associated molecules and activate qualitatively different adaptive T-helper (Th) cell responses [4]. Protective immunity against P. brasiliensis has been credited to a Th1 type response, whereas the anti-inflammatory cytokine IL-10 is generally correlated with deleterious effects in murine and human PCM [5]–[7]. Recent experiments have shown that P. brasiliensis infection activates DCs to migrate to regional lymph nodes and trigger a Th response [8]. The direct activation of DCs occurs via the recognition of specific microbial compounds, known as pathogen-associated molecular patterns (PAMPs), by germline-encoded pattern recognition receptors (PRRs). In particular, the Toll-like receptors (TLRs) and C-type lectin receptors (CLRs) are the most important PRRs for the recognition of fungal molecules [9], [10].
During the activation process, DCs are subject to profound changes due to the differential expression of a variety of immune-related genes, which regulate the efficiency of the DC response to pathogens [11]. From this perspective, the use of microarrays to evaluate the gene expression profiles of DCs has served as an important tool to investigate how these cells respond to infection and modulate the immune system upon interaction with different microorganisms [12]. Because little data are available about the regulation of DC genes upon P. brasiliensis infection, we sought to examine the transcriptional profile of murine bone marrow-derived DCs at an early time of interaction with yeast cells. Gene expression profiles were analyzed using microarray and validated using real-time RT-PCR. Cytokine secretion was also monitored. We identified 299 genes that were differentially expressed upon infection, including many genes that are involved in immunity (e.g., inflammatory cytokines, chemokines and PRRs), signal transduction, transcription and apoptosis. Additionally, we used inhibition assays to evaluate the role of the CLRs dectin-1 and mannose receptor (MR) in coordinating the expression of several immune-related genes upon exposure to P. brasiliensis. Taken together, our results provide a foundational description of early host gene expression changes in response to P. brasiliensis, which may allow a better understanding of the infectious process of this important, but neglected, fungal pathogen.
All work was conducted with the approval of the Committee on the Ethics of Animal Experiments of the University of Sao Paulo (CEUA/FCF permit number: 2921) according to the National Council on Animal Experiments and Control (CONCEA-MCT-Brazil) guidelines.
Male BALB/c mice were obtained from the animal laboratory of the University of São Paulo and used in experiments at 8 to 12 weeks of age. This strain has been shown to have intermediate resistance to P. brasiliensis infection [13].
The yeast form of the highly virulent P. brasiliensis isolate 18 was grown on Sabouraud agar and used for in vitro infection assays. Viability, as determined with Janus Green B vital dye (Merck), was always greater than 80%.
Bone marrow-derived DCs were generated from BALB/c mice according to the protocol described by Inaba et al. [14] with slight modifications. Briefly, mouse femurs and tibias were flushed with 2 ml phosphate buffered saline (PBS) containing 1% bovine serum albumin (BSA). Bone marrow cells were differentiated into DCs by culturing in RPMI 1640 tissue culture medium supplemented with 10% fetal calf serum (FCS), 10 mg/ml gentamicin and 50 ng/ml recombinant granulocyte-macrophage colony stimulating factor (GM-CSF) for 7 days at 37°C in a humidified atmosphere containing 5% CO2. On days 3 and 5, nonadherent cells (granulocytes and lymphocytes) were removed, and fresh medium supplemented with GM-CSF was added. On day 7, the non-adherent cells were removed and analyzed by flow cytometry using DC cell surface markers. Phenotypically, 80% of these cells express MHC class II, CD80, CD40, CD11b and CD11c being characterized as bone marrow-derived DCs (data not shown). Following DC generation, 107 cells were infected with P. brasiliensis at a yeast-to-cell ratio of 1∶1 at 37°C. Using this ratio of infection, an average of 70% of DCs is engaged in phagocytosis of at least one yeast cell (data not shown). At 6 h after infection, extracellular and weakly adherent fungi were removed by washing with pre-warmed RPMI. DCs were then lysed, and total RNA was extracted with the Trizol reagent (Invitrogen) according to the manufacturer's instructions. Total RNA from control (uninfected) DCs was also extracted with Trizol.
The transcriptional response of murine DCs to infection with P. brasiliensis was assessed using cDNA microarrays prepared on silane-coated UltraGAPS slides (# 40015, Corning). The microarrays contained a total of 4,500 target tissue-restricted antigen cDNA sequences, representing most murine tissues and organs. The cDNA clones on the microarrays were isolated from the Soares thymus 2NbMT normalized library prepared from the thymus of C57BL/6J 4-week-old male mice, which is available at the IMAGE Consortium (http://image.hudsonalpha.org/). The microarrays were prepared based on published protocols [15] using a Generation III Array Spotter (Amersham Molecular Dynamics) according to the manufacturer's instructions and cross-linked using an ultraviolet cross-linker. The cDNA complex probes were prepared by reverse transcription using 10 µg of total DC RNA followed by labeling of the resulting cDNAs with Cy3 or Cy5 fluorochromes using the CyScribe post-labeling kit (Amersham Biosciences) and oligo dT12–18 as a primer. The cDNA complex probes derived from total RNA obtained from P. brasiliensis-infected and non-infected control DCs were labeled with Cy5 using the CyScribe post-labeling kit (GE Healthcare). As a control for the hybridization procedure, we used equimolar quantities of Cy3-labeled cDNA generated from total RNA isolated from different mouse organs (thymus, spleen and lung). This approach allowed for the amount of cDNA targeted in each microarray spot to be estimated. The 15-h period of hybridization, followed by washing, was performed in an automatic slide processor system (ASP; Amersham Biosciences), and the microarrays were scanned with a Generation III laser scanner (Amersham Biosciences).
Microarray image quantification was performed using Spotfinder (http://www.tm4.org/spot.nder.html). The normalization process was carried out using the R platform (http://www.r-project.org), and statistical analyses were conducted with Multiexperiment Viewer (MeV) (version 3.1; http://www.tm4.org/mev.html). After normalization, SAM (Significance Analysis of Microarrays) was used to identify statistically significant differences in gene expression between the experimental and control conditions [16]. SAM computes a statistic for each gene, measuring the strength of the relationship between gene expression and the response variable (P. brasiliensis infected and non-infected DCs groups). It uses repeated permutations of the data to determine if the expression of any genes is significantly related to the response. The threshold for significance is determined by a tuning parameter delta based on the false-positive rate or false discovery rate (FDR). A high stringent FDR 0.5% (Delta = 1.017) and q-value≤0.05 were selected. The q-value (p-value adapted to multiple-testing conditions) for each gene is the lowest FDR at which that gene modulation is called significant. SAM analysis also provides the optional fold change parameter, to ensure that significant modulated genes change by at least a pre-specified amount. In the present work, a fold-change cutoff set to 1.2, simulating 20% cutoff, was used. Microarray data were deposited according to MIAME (Minimum Information About a Microarray Experiment) guidelines in the ArrayExpress databank (http://www.ebi.ac.uk/arrayexpress/) under accession number A-MEXP-2009.
Bone marrow-derived DCs, generated as described above, were plated at a concentration of 1×106 cells/ml in 24-well cell-culture-treated plates and pre-incubated for 30 min at 37°C with 100 µg/ml α-mannan obtained from Saccharomyces cerevisae (Sigma-Aldrich) or 200 µg/ml laminarin (Sigma-Aldrich) to block the mannose or dectin-1 receptors, respectively. The DCs were then infected with P. brasiliensis yeast cells, following the protocol for the in vitro infection assay described above for the microarray experiments. IL-12, IL-10 and TNF-α secretion were measured from culture supernatants using commercially available ELISAs according to the manufacturer's recommendations (BD Pharmingen).
qRT-PCR was used to validate the differential modulation of DC genes revealed by the microarray experiment and for the analysis of DC gene expression following mannose and dectin-1 receptor inhibition. To remove any genomic DNA contamination, total RNA extracted from cells from both experimental conditions was treated with RNase-free DNaseI (Promega) and precipitated with ethanol. These DNA-free RNA samples were then used for qRT-PCR. Equal amounts of RNA (1 µg) were reverse transcribed (Superscript III, Invitrogen) using an oligo(dT)12–18 primer and submitted to real time PCR. Amplification assays were carried out on a 7500 Fast Real-Time PCR System with SDS software (Applied Biosystems) in 10 µl reactions containing 0.2 µM of each primer, 5 µl SYBR Green PCR master mix (2×) and 0.2 µl cDNA. After initial denaturation at 95°C for 20 s, amplifications were carried out for 40 cycles at 95°C for 3 s and 60°C for 20 s. To confirm the amplification specificity, the PCR products were subjected to a melting curve analysis. The comparative CT (crossing threshold) method [17], using the constitutively expressed murine 40S ribosomal protein S9 (RPS9) gene as a control, was used to evaluate the expression (fold change) of each gene of interest. All primers used for qRT-PCR (Table S1) were based on sequences obtained from the mouse transcriptome database (http://www.informatics.jax.org) and designed with Primer3, which is available online (http://www-genome.wi.mit.edu).
GraphPad Prism 5.0 (GraphPad Software) was used for statistical analyses. The paired two-tailed Student's t test was used, and a P value≤0.05 was considered significant. In addition, multiple group comparisons were conducted by one-way ANOVA followed by Bonferroni tests, as appropriate.
The pattern of gene expression in murine bone marrow-derived DCs infected with P. brasiliensis yeast cells was assessed using microarray. Previous studies have shown that a 4- to 6-h co-cultivation of P. brasiliensis yeast cells with murine bone marrow-derived DCs stimulated significant phagocytosis [18]. Based on these results, we selected a 6-h infection period because it represented an early time point of fungal internalization by DCs. For each condition (P. brasiliensis-infected or non-infected control DCs), two independent cDNA microarray experiments were performed with 4,500 cDNA clones on each microarray. The analysis of DC gene-expression data using SAM revealed significant changes in the expression profiles of 299 genes (81 up-regulated and 218 down-regulated) in response to infection with P. brasiliensis (Table S2). Based on the findings of previous fungal-phagocyte interaction studies, we selected modulated genes and clustered them into different functional categories shown in Tables 1 and 2.
Genes encoding cytokines such as tumor necrosis factor alpha (TNF-α) and interleukin 12 (IL-12) were up-regulated. The induction of TNF-α was confirmed by qRT-PCR (Table 3). In addition, protein levels, as assayed by ELISA, were increased, consistent with the increased accumulation of TNF-α and IL-12 mRNA in DCs exposed to P. brasiliensis. In contrast, no significant IL-10 secretion was observed (Figure 1). P. brasiliensis infection also modulated the expression of genes encoding chemokines, which are critical chemotactic factors in the immune system. As shown in Tables 1 and 2, the genes encoding the chemokines CCL22, CCL27 and CXCL10 were up-regulated, whereas CCL25 transcription was decreased. Moreover, a ten-fold increase in CCL22 transcript levels was observed using qRT-PCR, confirming the microarray data (Table 3).
The expression levels of some DC membrane receptor genes that are associated with immune responses were also significantly modulated. Microarray results revealed that the genes encoding the CLR receptor DC-SIGN (CD209a), the IgG receptor FcγR1 and TLR4 were down-regulated at an early time point after yeast infection, with fold-change values (FC) of −1.37, −1.33 and −1.26, respectively (Table 2). Although the decrease in TLR4 mRNA levels at 6 h after infection was confirmed by qRT-PCR (Table 3), no significant modulation was found at a later time point, 24 h after infection (data not shown). Microarray and qRT-PCR data showed that TLR2 gene expression did not appear to be influenced by the presence of fungal cells at 6 h (Table 3). Likewise, the expression levels of two other TLRs family members (TLR6 and TLR9, data not shown) and the universal adaptor molecule of the TLR signaling pathway, MyD88, were not modulated. The unchanged expression of MyD88 was also validated using qRT-PCR (Table 3). In this manner, genes that encode negative regulators of TLR-mediated signaling, such as TOLLIP (Toll-interacting protein) and the lymphocyte antigen 86 (LY86) known as MD1, were up-regulated 6 h after infection (Table 1).
Integrins are a family of proteins whose members are involved in a variety of cell-matrix and cell-cell adhesion processes and signaling events that are central to immunologic and inflammatory processes [19]. These proteins are heterodimeric transmembrane glycoproteins that consist of a series of related α and β subunits. As shown in Table 2, down-regulation of the DC integrin genes ITGAM (CD11b), Itg2b (CD18) and ITGA6 was observed in response to P. brasiliensis infection. The α subunit of CD11b bound to the β subunit CD18 is known as integrin CD11b/CD18 or Complement Receptor 3 (CR3). In addition to its ability to promote the phagocytosis of iC3b-opsonised particles, CR3 recognizes exogenous ligands, such as β-glucan, and has been implicated in DC responses to fungi [10], [20].
Infected DCs also showed altered expression of interferon-inducible genes and genes encoding transcription factors. IRG1, or immunoresponsive gene 1, was up-regulated in response to P. brasiliensis infection (Table 1). Interestingly, Degrandi et al. [21] showed that the mRNA level of this gene increased following TNF-α and IFN-γ treatment. Another interferon-inducible gene, IFI203, was shown to be up-regulated, with an FC of 2.27. As shown in Tables 1 and 3, after 6 h of in vitro infection with P. brasiliensis yeast cells, NFκB was up-regulated in DCs. Its product is the major transcription factor that induces the expression of pro-inflammatory genes. Moreover, NκRF was down-regulated. NκRF encodes a transcriptional repressor that binds to specific negative regulatory elements (NREs) to counteract NFκB activity at certain gene promoters [22]. Other down-regulated genes involved in transcriptional regulation included STAT2, STAT6, IER5 and ILF2 (Table 2).
Another important group of genes with altered expression consisted of those related to apoptosis. The pro-apoptotic genes CASP2, BCLAF1 and DEDD were up-regulated (Table 1), while BTG2 and RTN4 were down-regulated (Table 2). Moreover, the expression of the anti-apoptotic gene CFLAR was inhibited, with an FC of 1.75 (Table 2).
As described above, the expression of several DC receptor-encoding genes, including opsonin-dependent receptors (CR3 and FcγR1) and non-opsonin dependent PRRs (TLR2, TLR4, TLR6, TLR9 and DC-SIGN), were analyzed using microarray. The genes encoding TLR4, CR3, FcγR1 and DC-SIGN were down-regulated, while TLR2, TLR6 and TLR9 were not differentially regulated. To obtain a more comprehensive analysis of immune receptor expression, we evaluated the expression profiles of dectin-1 and MR, two major fungal CLRs [23] that were not included in the microarray target cDNA library. qRT-PCR analysis revealed that the dectin-1 receptor gene was up-regulated by more than ten fold after 6 h of P. brasiliensis infection, while MR was down-regulated (Figure 2). These results prompted us to investigate the possible effects of dectin-1 and MR on the transcriptional profiles of six immune-related genes (CCL22, TNF-α, NFκB, TLR2, TLR4 and MYD88) that were selected based on the microarray results and importance to the fungi-host interaction. DCs were incubated with laminarin or mannan to block dectin-1 and MR, respectively. After 30 minutes, DCs were infected with P. brasiliensis for 6 h, total RNA was extracted, and the expression profiles of the selected immune-related genes were determined by qRT-PCR. The comparative CT method [17], using the constitutively expressed murine RPS9 gene as a control, was used to evaluate the expression (fold change) of each interest gene. In this analysis we have considered each gene expression ratio obtained between P. brasiliensis infected and uninfected DCs with respect to host cells treatment or not with mannan or laminarin. In the presence of mannan, no differences in transcript accumulation were observed. In contrast, DCs treated with laminarin before infection showed a significant up-regulation of all six genes investigated, except CCL22 (Figure 3). These data suggest that dectin-1 plays a prominent role in the coordination of gene expression during the initial phase of P. brasiliensis infection.
The secretion of IL-10, IL-12 and TNF-α was evaluated after inhibitor assays (Figure 4). Treatment with laminarin prior to infection did not alter the secretion of any of these three cytokines when DCs were exposed to P. brasiliensis, whereas MR inhibition by mannan significantly reduced IL-12 secretion. Treatment with laminarin or mannan alone (i.e., without subsequent infection) had no effect on basal the secretion of any of these three cytokines by DCs (data not shown).
To our knowledge, this is the first study that has investigated the gene expression profile of mouse DCs in response to a primary pathogenic fungus. Previous studies have evaluated gene expression in human monocyte- and mouse splenic-derived DCs in response to the opportunistic fungi Candida albicans and Cryptococcus neoformans, respectively [12], [24]. Our microarray data identified several genes whose expression is modulated at an early time point after bone marrow-derived DCs are exposed to P. brasiliensis yeast cells. In particular, genes related to immune responses (mainly inflammation-associated genes), signal transduction, transcriptional regulation and apoptosis were altered.
DCs connect innate and adaptive immunity by recognizing pathogen-associated molecules and producing cytokines that subsequently drive qualitatively different adaptive Th responses. IL-12 produced by DCs is the key cytokine that stimulates a Th1-type cell-mediated response, the major source of immunity against systemic fungal infections [10]. Here, we demonstrate the increased production of IL-12 mRNA and protein by BALB/c DCs infected with P. brasiliensis for 6 hours. Considering that this mouse lineage has intermediate resistance to P. brasiliensis infection [13], we could draw a parallel with previous observations demonstrating that DCs derived from resistant mice (A/J) stimulate a Th1 response in vitro more efficiently than DCs derived from susceptible mice (B10.A) when pulsed with the immunodominant P. brasiliensis antigen, gp43 [25]. Diverse antigen-recognition and -processing mechanisms in DCs from resistant and susceptible mice, which would give rise to differential production of IL-12, could be involved in determining susceptibility to this fungus. In fact, gp43-pulsed DCs from resistant mice were later reported to secrete higher levels of IL-12 than those from susceptible mice, but this increase was not statistically significant [26]. Based on this observation, the authors speculated that IL-12 is not the key factor for promoting a Th1-type response in A/J mice. Other mechanisms, such as the high expression of the costimulatory molecule CD80, concomitant with low production of IL-4 and IL-6, could contribute to the induction of Th1 cells. In contrast to these results, a recent study described a role for IL-12 in determining resistance to experimental PCM. Mice that are deficient for the IL-12p40 subunit produce no detectable IFN-γ and high levels of IL-10 protein, and this phenotype is associated with uncontrolled fungal proliferation and dissemination [27]. Moreover, Moraes-Vasconcellos et al. [28] described a patient with disseminated PCM who harbored a primary immunodeficiency in the beta 1 subunit of the IL-12/IL-23 receptor.
Our microarray analysis showed that, in addition to IL-12, other proinflammatory cytokine- and chemokine-encoding genes (TNF-α, CCL22, CCL27 and CXCL10) were induced by P. brasiliensis exposure. Interestingly, microarray studies published by Lupo et al. [24] demonstrated the up-regulation of all these genes (except CCL27) in murine DCs exposed to acapsular C. neoformans relative to DCs exposed to encapsulated strains. These results are consistent with the role of the polysaccharide capsule as the main virulence factor for this important opportunistic fungus and the ability of the capsule to act as a shield from immune recognition and activation. Notably, IL-12, CCL22 and CXCL10 are part of a cluster of murine DCs signature expressed genes that discriminate very accurately between inflammatory and non-inflammatory stimuli, such as lipopolysaccharide and dexamethasone, respectively [29]. Another signature inflammatory DC gene identified in our study was NFκB1, which encodes the p105 subunit of the NFκB protein. NFκB is a master regulator of gene transcription during development and inflammatory processes and plays a critical role in the activation of innate and adaptive immunity [30]. Among the known targets of NFκB, we were able to show that DCs up-regulated IL-12, TNF-α, CCL22 and CXCL10 in response to P. brasiliensis. It is important to note that levels of NκRF (a nuclear inhibitor of NFκB activity) mRNA were reduced concomitant with NFκB induction.
TNF-α is a cytokine that is critical for the successful control of fungal infections and the development of a Th1-dependent response. This cytokine augments the cytotoxic activity of activated macrophages, induces chemokine production and, along with IFN-γ, regulates granuloma formation [31]. P. brasiliensis-infected mice lacking the p55 subunit of TNF-α receptor were reported to develop severe PCM associated with non-organized granulomas [32]. In experimental pulmonary cryptococcosis, the transient depletion of TNF-α production during the early innate immune response permanently impaired the long-term control of fungal growth. This effect was coupled with a temporary decrease in neutrophil lung influx, reduced IL-12 production and recruitment of DCs to draining lymph nodes [33]. In this context, the early TNF-α gene expression and protein production observed in the current study is likely to be associated with the induction of a protective response. As discussed above, and consistent with a proinflammatory scenario, we observed the up-regulation of several chemokines (CCL22, CXCL10 and CCL27) in our study. Chemokines play a major role in mediating the extravasation and accumulation of specific leukocytes at sites of infection, which is crucial for the local control of fungal invasion [34]. Both CXCL10 and CCL22 are mainly chemotactic for monocytes and T-lymphocytes. CCL22 is also chemotactic for DCs, the major cell source for this chemokine in vivo and in vitro [35]. Notably, our group has previously shown the induction of CCL22 gene, among other chemokine-encoding genes, in peritoneal murine macrophages that were infected with P. brasiliensis [36]. In addition to its role as a chemoattractant, CCL22 enhances the microbicidal activity of macrophages by stimulating a strong respiratory burst and lysosomal enzyme release [37]. Similarly, CXCL10 not only induces leukocyte migration but also up-regulates the production of Th1 cytokines (mainly IFN-γ) and down-regulates the production of Th2 cytokines upon interaction with its receptor on T- cells [38]. In humans, single nucleotide polymorphisms in the CXCL10 gene lead to reduced chemokine production by DCs exposed to Aspergillus fumigatus, causing invasive aspergillosis [39].
Of particular interest in our study was the assessment of transcriptional modulation of genes encoding PRRs, which specifically interact with pathogen PAMPs and thus regulate the production of various immune-related molecules. A great number of PRRs have been identified; of these, the TLRs and CLRs families are of major interest because they appear to have critical roles in fungal immunity [40]–[41]. Regarding TLRs, we observed no appreciable modulation of the expression of TLR2, whereas TLR4 was down-regulated. In accordance, the expression of MYD88 was unaltered in P. brasiliensis-infected DCs, as shown using microarray and qRT-PCR. MYD88 is a universal adaptor molecule in the TLR signaling pathway that ultimately activates NFκB and thus affects subsequent cytokine and chemokine production [42]. Interestingly, the importance of MYD88 signaling in the experimental murine model of PCM is controversial. Gonzalez et al [43] showed that this adaptor protein is not necessary for the effective control of blood-borne disseminated P. brasiliensis infection. In contrast, a recent study demonstrated that MyD88-dependent signaling participates in the induction of protective immune host defense against pulmonary PCM [44]. Different fungal strains and routes of infection may have been responsible for this divergence. In our study model, the limited participation of TLR-mediated signal transduction in response to P. brasiliensis is further supported by the fact that two genes (TOLLIP and MD1) that encode negative regulators of this signaling pathway were induced. TOLLIP is an adaptor molecule that can associate with TLR2 and TLR4 to inhibit MyD88 binding and activation [45]. Indeed, the overexpression of TOLLIP precludes NFκB activation in response to TLR2 and TLR4 agonists [46]. Likewise, MD1 is a helper molecule for RP105, a TLR homolog that acts as a physiological negative regulator of TLR4 responses [47]. These results suggest that TLRs may have only a minor role in the host responses elicited by P. brasiliensis. In fact, TLR2 and TLR4 deficient mice infected with this fungus demonstrated equivalent mortality rates compared with wild-type littermates [48]–[49]. Interestingly, similar results were obtained in TLR2 and TLR4 knockout mice infected with C. neoformans [50].
Our microarray data suggesting limited role for TLR-mediated signaling in response to P. brasiliensis, coupled with significant production of IL-12 and TNF-α but not IL-10, prompted us to search for a TLR2-4/MyD88-independent mechanism that could explain the induction of these proinflammatory cytokines in DCs. In accordance, bone marrow-derived DCs from mice deficient in the TLR2 and TLR4 genes and infected with C. neoformans had no significant reduction in IL-12 and TNF-α protein levels [50]. Because receptors of the C-type lectin family, particularly dectin-1 and MR, have been reported to be critical for the recognition of fungi and the activation of macrophages and DCs [40], we sought to evaluate their expression using qRT-PCR analysis (these genes were not represented in our microarray). The dectin-1 receptor gene, whose product recognizes β-(1,3)-glucans on the cell walls of fungi, was up-regulated by ten fold, whereas the MR gene transcription was diminished. These results suggest that dectin-1 participates in the induction of TNF-α and the production of IL-12. Unexpectedly, the blockade of this receptor with laminarin did not significantly reduce the production of these cytokines by DCs. This apparent contradiction may be because dectin-1 inhibition results in significant induction of the TLR2, TLR4, MYD88, NFκB, and TNF-α genes, which could be involved in the sustained IL-12 and TNF-α production, probably via TLR-mediated signaling. Thus, dectin-1 may act as negative regulator of the TLR signaling pathway in our model. These results could be associated with the intermediate PCM resistance pattern of the mouse strain used in our study. DCs from mice that are susceptible to this fungus, but not resistant mice, secrete significant amounts of IL-10 in a TLR2/dectin-1 collaborative signaling-dependent manner. Furthermore, TLR2 gene expression was only induced by P. brasiliensis in susceptible mice and its deletion suppress IL-10 production [51]. Altogether, these results suggest that dectin-1 signaling down-regulates the expression of TLR-associated genes, leading to a Th1-like response with little anti-inflammatory IL-10 production (i.e., no collaborative dectin-1 and TLR signaling). MR blockade did not alter TNF-α or IL-10 secretion, as observed in DCs treated with laminarin. However, despite the downregulation of the MR-encoding gene, a significant reduction in IL-12 expression was observed after treatment with mannan. MR recognizes P. brasiliensis, C. neoformans and C. albicans, and this receptor has been implicated in mediating the production of pro-inflammatory cytokines, including IL-6, TNF-α, IL-1β and IL-12 [40]. These data could indicate a lack of correlation between the MR mRNA level and the expression of the functional receptor. Alternatively, because we used fungal mannan in our inhibition assay and this mannose polymer is also recognized by other PRRs, such as DC-SIGN, SCARF1 and mincle, we could not rule out the possibility that one or more of those receptors promote the induction of IL-12 secretion.
In summary, our findings demonstrate that P. brasiliensis triggers the accumulation of mRNAs for genes that encode proinflammatory cytokines and chemokines as well as other molecules involved in the early response of DCs to this fungus. These results provide a better understanding of the molecular pathogenesis of PCM and should promote future investigations into the innate host response to this fungus, including in vivo analysis. Of particular interest were the results regarding receptor inhibitor assay, because the mechanisms by which dectin-1 negatively regulates TLR-associated gene expression remain to be determined.
The information concerning genes and proteins accession numbers mentioned in the manuscript (text and Tables 1 and 2) is described below, following the criteria: Gene name - Gene Accession number/Protein Accession number Il12b (Il12p40) - NM_008352/NP_032378; Cxcl10 - AK146144/NP_067249; Ccl22 - AF052505/NP_033163; Ccl27a - AK146066/NP_035466; Cd27 - L24495/NP_001036029; Cd53 - NM_007651/NP_031677; C1qb - AK152764/NP_033907; Ifi203 - AK172243/NP_032354; Irg1 - NM_008392/NP_032418; Gbp2 - BC032882/NP_034390; Ly86 - AK172197/NP_034875; Tollip - BC062139/NP_076253; Nr3c1 - NM_008173/NP_032199; Pias1 - AK075708/NP_062637; Nfkb1 - AK036827/NP_032715; Taf4a - NM_001081092/NP_001074561; Casp2 - BC034262/NP_031636; Ccl25 - AK154211/NP_033164; Cd209a - AY049062/NP_573501; Fcgr1 - AK033874/NP_034316; Itgb2 - AK136502/NP_032430; Itgam - NM_001082960/NP_032427; Mapk3 - BF579077/NP_036082; Map3k10 - NM_001081292/NP_001074761; Ier5 - NM_010500/NP_034630; Stat2 - AF206162/NP_064347; Stat6 - NM_009284/NP_033310; Otud7b - BC141397/NP_001020785; Hspa8 - BC006722/NP_112442; Hsp90b1 - AK160827/NP_035761; Nadk - NM_001159637/NP_619612; Rtn4 - NM_194054/Q99P72.
|
10.1371/journal.ppat.1000840 | Suppression of mRNAs Encoding Tegument Tetraspanins from Schistosoma mansoni Results in Impaired Tegument Turnover | Schistosomes express a family of integral membrane proteins, called tetraspanins (TSPs), in the outer surface membranes of the tegument. Two of these tetraspanins, Sm-TSP-1 and Sm-TSP-2, confer protection as vaccines in mice, and individuals who are naturally resistant to S. mansoni infection mount a strong IgG response to Sm-TSP-2. To determine their functions in the tegument of S. mansoni we used RNA interference to silence expression of Sm-tsp-1 and Sm-tsp-2 mRNAs. Soaking of parasites in Sm-tsp dsRNAs resulted in 61% (p = 0.009) and 74% (p = 0.009) reductions in Sm-tsp-1 and Sm-tsp-2 transcription levels, respectively, in adult worms, and 67%–75% (p = 0.011) and 69%–89% (p = 0.004) reductions in Sm-tsp-1 and Sm-tsp-2 transcription levels, respectively, in schistosomula compared to worms treated with irrelevant control (luciferase) dsRNA. Ultrastructural morphology of adult worms treated in vitro with Sm-tsp-2 dsRNA displayed a distinctly vacuolated and thinner tegument compared with controls. Schistosomula exposed in vitro to Sm-tsp-2 dsRNA had a significantly thinner and more vacuolated tegument, and morphology consistent with a failure of tegumentary invaginations to close. Injection of mice with schistosomula that had been electroporated with Sm-tsp-1 and Sm-tsp-2 dsRNAs resulted in 61% (p = 0.005) and 83% (p = 0.002) reductions in the numbers of parasites recovered from the mesenteries four weeks later when compared to dsRNA-treated controls. These results imply that tetraspanins play important structural roles impacting tegument development, maturation or stability.
| Schistosomes, or blood flukes, reside in the blood vessels surrounding the liver and bowel of their human hosts. They infect 200 million people and kill many thousands each year in developing countries. The parasites cover themselves in a unique series of cell membranes called the tegument. Molecules in the tegument membranes are a major target for the development of new drugs and vaccines against the parasite. Here we show that at least one member of a family of tegument membrane proteins called tetraspanins, Sm-TSP-2, is integral to the proper formation of the tegument and subsequent survival of the parasite in its human host, providing a potential mechanism by which a vaccine based on Sm-TSP-2 protects immunized hosts.
| Schistosomes are parasitic trematodes that cause chronic infection in over 207 million people in 76 developing tropical countries. Schistosomiasis is generally associated with poverty, poor water supply and inadequate sanitation [1]. Infection rates and intensities are high in early childhood, peak around 8 to 15 years and decrease in adulthood [2]. Despite effective and inexpensive widespread treatment with the anthelmintic drug praziquantel for over 20 years, this parasitic disease still causes more than 250,000 deaths per year and accounts for 1.7 to 4.5 million disability-adjusted life years (DALYs) lost annually [3].
Humans become infected with schistosomes when they are exposed to free-living cercariae in fresh water. Cercariae penetrate the skin, shed their tails and transform into schistosomula, which reside in the dermis of the skin before entering the blood capillaries to migrate through the vasculature to the portal venous system where they mature into adult worms [4]. The outer surface of schistosomula and adult worms, the tegument, is a multinucleated syncitium that contains tegumental cell bodies situated below the muscular layers. During transformation from cercaria to schistosomula, the outer surface of the tegument (the interface with the host) is remodeled from a single membrane with a prominent glycocalyx into an unusual double membrane (or “heptalaminate”) structure [5]. This double membrane is widely believed to play an essential role in the ability of schistosomes to evade the host immune system, a characteristic that allows them to live for years within their hosts [6]. The outer of the two surface membranes also has the ability to adsorb host blood molecules, masking its non-self status thereby contributing to immune evasion and prolonged survival [7]. We believe that tegumental proteins are ideal targets for immunological and pharmacological intervention [8]. The generation of a large number of S. mansoni expressed sequence tags [9] and the recently completed genome sequence [10], in combination with advances in characterizing the tegument proteome has led to the discovery of many tegument specific proteins [11]. Among them are a group of membrane proteins called tetraspanins, which are highly expressed in the outer tegument membrane of adult schistosomes [12], [13]. To date, five tetraspanin cDNAs have been described from S. mansoni, namely Sm-23 [12], Sm-tsp-1 and Sm-tsp-2 [13], Sm-tetraspanin-B and Sm-tetraspanin-C [14].
Tetraspanins are a large superfamily of surface-associated membrane proteins characterized by the conserved structure of four hydrophobic transmembrane domains, a small and large extracellular loop, an interconnecting intracellular loop, and cytoplasmic amino- and carboxyl- termini [15]. Tetraspanins undergo post-translational modification in which palmitate is bound to the membrane proximal cysteine residues and associates with cholesterol-rich domains [16]. This process enables tetraspanins to play key roles in molecular organization of cell membranes, interacting with one another and also specific partner proteins such as integrins, MHC and co-stimulatory molecules to form large signal transducing complexes termed tetraspanin-enriched microdomains (TEMs) [17]. Tetraspanins are widely distributed in many cell types but their physiological roles are mostly unknown. Several lines of evidence have implicated tetraspanins in the regulation of cell adhesion, differentiation, motility, aggregation, cell signaling and sperm-egg fusion [18], [19], [20], [21]. They have been linked to various pathological processes including lymphocyte activation [19], cancer [22], fertilization [23], [24], and interactions between pathogens and host cells such as HIV [25], HCV [26] and Plasmodium [27].
We previously identified two cDNAs, Sm-tsp-1 (Sm01494) and Sm-tsp-2 (Sm12366), in adult S. mansoni using signal sequence trapping [13], and showed that both of these tetraspanins were expressed in the tegument of the adult parasite [28]. Other authors confirmed the surface expression of these tetraspanins using various mass spectrometric approaches to characterize the schistosome surface [11], [29], [30]. We expressed the large extracellular loop of Sm-TSP-1 and Sm-TSP-2 in E. coli and used the soluble recombinant proteins to immunize mice and then challenged them with cercariae. Mice vaccinated with recombinant Sm-TSP-1 and Sm-TSP-2 had significantly reduced adult worm, liver egg and fecal egg burdens [28]. Moreover, strong IgG1 and IgG3 antibody responses against Sm-TSP-2 were detected in sera of individuals deemed putatively resistant (PR) to S. mansoni in comparison to sera from chronically infected individuals [28].
Despite their promise as vaccines against schistosomiasis, the functions of Sm-TSP-1 and Sm-TSP-2 have not yet been elucidated. We therefore employed RNA interference (RNAi) to explore the roles of Sm-tsp-1 and tsp-2 in larval and adult S. mansoni. RNAi has been utilized with S. mansoni to suppress endogenous gene expression in schistosomula [31], adult worms [32], eggs [33] and sporocysts [34]. Here, we show that RNAi results in reductions in expression of Sm-tsp-1 and tsp-2 mRNAs in schistosomula and adult worms, and malformation of the tegument in worms cultured in vitro. Moreover, silencing of tsp-1 and tsp-2 expression in schistosomula results in up to 90% fewer worms maturing to adulthood when introduced into mice compared with parasites exposed to control dsRNAs, highlighting their essential roles in tegument biogenesis and maintenance and further supporting the development of novel therapies targeting these genes and their protein products.
Expression of Sm-tsp-1 and Sm-tsp-2 mRNAs in different stages of the S. mansoni life cycle was determined relative to control Sm-α-tubilin mRNA using qRT-PCR. Sm-tsp-1 and Sm-tsp-2 mRNAs were detected in all stages of the schistosome life cycle with higher levels identified in eggs, miracidia and cercariae than in 5-day old schistosomula, males and female worms for tsp-1; a similar expression profile was observed for tsp-2 but gene expression was notably reduced in cercariae (Figure 1). Interestingly, the highest level of Sm-tsp-1 expression was detected in cercariae whereas Sm-tsp-2 expression was lowest in cercariae.
We previously demonstrated that Sm-TSP-1 and Sm-TSP-2 are expressed on the tegument surface membrane of adult worms [28]. The tegument is fully formed by 3h after cercarial transformation [35], so to determine whether these TSPs are expressed in the tegument at this early stage after host entry and whether they are accessible to antibodies on live parasites, we probed live newly transformed schistosomula with antibodies against both proteins. Both Sm-TSP-1 and Sm-TSP-2 were detected over the entire surface tegument of live schistosomula when probed with mouse anti-TSP-1 or -TSP-2 sera followed by FITC-labelled anti-mouse IgG (Figure 2).
Adult worms soaked for 7 days in Sm-tsp-1 dsRNA had a 61% (p = 0.009) reduction in Sm-tsp-1 mRNA expression compared to parasites soaked in control dsRNA (Figure 3A). A 74% (p = 0.009) reduction in Sm-tsp-2 mRNA levels was detected in worms that were cultured in media containing Sm-tsp-2 dsRNA compared to parasites soaked in luciferase dsRNA (Figure 3B). Parasites were visually monitored for motility on a daily basis but no differences were detected between groups (not shown).
Soaking of 3 h old schistosomula in Sm-tsp-1 dsRNA for 7, 14 and 21 days caused 75% (p<0.001), 67% (p = 0.019) and 69% (p = 0.021) decreases in Sm-tsp-1 mRNA expression in comparison to the control group (Figure 4A). Larval parasites incubated with Sm-tsp-2 dsRNA for 7 days exhibited an 88% (p<0.001) decrease in Sm-tsp-2 transcript levels compared to luciferase dsRNA treated schistosomula (Figure 4B). RNAi knockdown was maintained with reductions of 82% (p = 0.004) and 69% (p = 0.021) at days 14 and 21, respectively, compared to the control group. As observed in adult worms, suppression of Sm-tsp RNAs resulted in no obvious phenotypic differences compared to the luciferase dsRNA-treated control group when examined by light microscopy. Cultures were visually inspected using a light microscope on a daily basis and no differences in early growth and development of schistosomula (development of intestinal ceca or size of schistosomula) [36] were apparent between test and control dsRNA treated groups.
To determine whether knockdown of Sm-tsp-2 RNA was evident at the protein level, we performed Western blot analysis on dsRNA treated adult (Figure 5A) and larval (Figure 5B) parasites. Parasites were treated with Sm-tsp-2 or luciferase dsRNAs, lysed in 1% Triton X-100 and immunoblotted with anti-Sm-TSP-2 or anti-Sm-Pmy antibodies which target a sub-tegumental muscle protein, paramyosin [37]. Sm-TSP-2 protein expression was decreased in adult worms treated with Sm-tsp-2 dsRNA compared to worms treated with luciferase dsRNA for the four concentrations (2.0, 1.0, 0.5 and 0.25 µg) tested. In contrast, the Sm-Pmy protein expression levels did not change in both test and control groups. The experiment was repeated three times with similar results and a representative image is shown (Figure 5A). Densitometry analysis was performed on each band and the ratio of Sm-TSP-2 to Sm-Pmy at each concentration was calculated. Analysis of whole worm lysates (0.25 µg) by densitometry (not shown) revealed an average of 61% (p = 0.027) reduction in Sm-TSP-2 expression in adult worms treated with Sm-tsp-2 dsRNA compared to the control luciferase group. For RNAi treated schistosomula, the amount of Sm-TSP-2 protein expressed by schistosomula after 7 days in culture with Sm-tsp-2 dsRNA was reduced compared to parasites soaked in luciferase dsRNA (Figure 5B). Densitometry analysis of lysates (2 µg, 1 µg and 0.5 µg) showed an average decline of 36% (data not shown). This decrease was lower than expected since suppression of Sm-tsp-2 mRNA was more pronounced in schistosomula than in adult parasites. Adult and larval parasites soaked in Sm-tsp-1 dsRNA demonstrated no obvious differences in protein expression to luciferase dsRNA control worms by Western blotting analysis (data not shown).
Adult parasites and schistosomula treated with Sm-tsp-2 dsRNA in vitro displayed modified tegument structure when visualized with transmission electron microscopy (TEM) compared with luciferase dsRNA treated controls (Figure 6). The tegument of adult worms incubated in vitro in Sm-tsp-2 dsRNA (Figure 6C,E) was more highly vacuolated than luciferase dsRNA controls (Figure 6A), with extensive and enlarged vacuoles throughout the surface layer. The tegument of these parasites had less apparent cytoplasm and hence fewer cytoplasmic inclusions and was frequently much thinner than that of controls (Figure 6C,E). Schistosomula transformed and cultured in vitro presented a tegument that resembled that of larvae from natural or experimental infection (Figure 6B) [38]. The tegument in Sm-tsp-2 dsRNA treated schistosomula (Figure 6D,F) was consistently thinner than those of luciferase controls (P<0.001), measuring on average 0.3784±0.016 µm compared with 0.5842±0.323 µm for luciferase controls (Figure 6G). Volume density measures for invaginations and clear vesicular compartments of the tegument showed higher volumes for these compartments in Sm-tsp-2 treated schistosomula (p = 0.014; Figure 6F). The morphology of the schistosomula tegument was consistent with a failure to close invaginations of the surface (Figure 6D,F). Adult worms and schistosomula soaked in Sm-tsp-1 dsRNA showed no obvious differences to luciferase dsRNA control worms when examined by transmission electron microscopy (data not shown).
In the mammalian host, larval schistosomes migrate from the skin through the lungs to the liver and then mature in the mesenteric veins [4]. In an effort to mimic in vivo conditions, 3 h schistosomula were electroporated with 100 µg/ml of Sm-tsp-1, Sm-tsp-2 or luciferase dsRNA and then injected intramuscularly into female C57BL/6 mice. Four weeks later mice were perfused to determine the number of parasites that reached maturity in the mesenteries. Significantly fewer parasites were recovered from the mesenteric veins compared to the luciferase control group (see Figure 7A for results of three experiments). Mice injected with schistosomula that were electroporated with Sm-tsp-1 dsRNA yielded 48% (p = 0.045), 60% (p = 0.009) and 67% (p = 0.019) reduction in the number of parasites recovered for Experiments 1, 2 and 3, respectively in comparison to the luciferase control group. Schistosomula pretreated with Sm-tsp-2 dsRNA and then injected into mice resulted in 70% (p = 0.039), 91% (p = 0.009) and 78% (p = 0.018) decreases in parasite survival for Experiments 1, 2 and 3, respectively when compared to the luciferase dsRNA group. The numbers of mature worms harvested from the luciferase control group were very low, with recovery ranging from 0.5-1.5%, however the data was consistent between three experiments, with a reproducible and significant reduction in worm recovery rates between tsp and luciferase dsRNA treated parasites.
RNA was extracted from surviving worms that were perfused from mice and transcript levels were analyzed by qRT-PCR. Sm-tsp-1 expression was only slightly lower (17%) in worms recovered from mice that were infected with Sm-tsp-1 dsRNA-treated schistosomula compared to the control group. Likewise, Sm-tsp-2 expression was slightly reduced (15%) in worms recovered from mice that were infected with Sm-tsp-2 dsRNA treated worms compared to the luciferase control group (Figure 7B). However, when the same batch of dsRNA electroporated schistosomula were cultured in vitro for the same period of time (4 weeks), as opposed to being injected into mice, significant knockdown of Sm-tsp-1 and Sm-tsp-2 transcripts by 58% and 87%, respectively (Figure 7C), was observed. These results illustrate that silencing of Sm-tsp-1 and Sm-tsp-2 by either soaking or electroporation leads to suppression of tetraspanin genes in schistosomes, and suppression is maintained for at least 4 weeks in culture. The data also implies one of three possible outcomes for Sm-tsp dsRNA treated schistosomula that survived to adulthood after being transferred into mice; (1) RNAi was not as effective in those individual schistosomula that survived in mice as opposed to those that perished; (2) some of the RNAi treated parasites received (or took up) less dsRNA, and therefore the efficacy of gene suppression was variable between individuals in a single electroporated batch; (3) it is also possible that host developmental cues stimulate transcription.
Schistosomes express a family of tetraspanins in their tegument. Sm23 was the first tetraspanin identified in S. mansoni [12], and is of interest as a DNA vaccine antigen against schistosomiasis [39]. Its orthologue from S. japonicum, Sj23, protects water buffaloes against challenge infection when administered as a DNA vaccine [39]. We identified two additional tetraspanins, Sm-tsp-1 and Sm-tsp-2, which showed high levels of protection when administered to mice as recombinant protein vaccines against S. mansoni [13], [28]. However, despite the protective efficacy that these tetraspanins afford, their functions in the parasite are unknown. To understand the roles that these proteins play in the schistosome tegument, we herein explored the effects of silencing the expression of Sm-tsp-1 and Sm-tsp-2 mRNAs in adult and larval S. mansoni.
RNAi has been used to suppress a number of schistosome genes in an effort to determine their functions [40], [41]. Soaking of S. mansoni with dsRNA encoding the intestinal protease cathepsin B (SmCB1), resulted in greater than 10-fold decrease in SmCB1 mRNA levels and significant growth inhibition compared to parasites treated with control dsRNA [42]. Suppression of the mRNA encoding another intestinal protease, S. mansoni cathepsin D (SmCD), in schistosomula by electroporation with dsRNA led to reduction in RNA transcript levels, growth retardation in vitro and in vivo, and decreased cathepsin D enzymatic activity [43]. Silencing of the SmAQP gene encoding a water channel protein by electroporating schistosomula with short interfering RNAs suppressed mRNA and protein expression in the tegument, and treated parasites cultured in vitro exhibited stunted growth and lower viability [44]. RNAi has been used to determine the functional importance of tetraspanins in other organisms [45]. Suppression of tetraspanin-15 mRNA by feeding C. elegans with dsRNA resulted in dissociation of the cuticle and degeneration of the hypodermis, compromising epidermal integrity [46]. RNAi has also been used to determine the function of human tetraspanins in various cell types [45]. For example, the CD151 tetraspanin interacts with membrane proteins including the laminin-binding integrin α3β1; when lung adenocarcinoma cells were cultured on laminin-511 and then treated with CD151 siRNA, abnormal membrane protrusions on laminin-511 were apparent and tyrosine phosphorylation dependent signalling was reduced [47]. These findings indicate a role for tetraspanins in the maintenance of cell membrane biogenesis and structural integrity, and support our observations on the compromised tegument membrane formation in S. mansoni when tsp mRNA expression is suppressed.
Numerous reports have documented molecular interactions between tetraspanins and MHC, and involvement of human tetraspanins in regulating T cell co-stimulation and peptide/MHC presentation [48], [49], [50], indicating additional, non-structural roles. Schistosomes acquire host MHC onto their surfaces [51], presenting the intriguing possibility that they function as a receptor for host MHC. However, the majority of mammalian tetraspanin binding partners identified to date are membrane proteins rather than extracellular ligands [45]; moreover, our data presented here implies that schistosome tetraspanins are pivotal for proper tegument formation, even during in vitro culture in the absence of immune cells, supporting a structural role in the establishment and maintenance of the tegument. Indeed, the tetraspanin CD9 complexes with numerous proteins including Ig-containing proteins [52], a family of proteins which are also present in the S. mansoni tegument membrane [30]. Various authors have described the contribution of tetraspanins, such as CD9 and CD151, with members of the integrin family in promoting cell-cell interactions and migration [53], [54], [55]. Mass spectrometric analysis of the S. mansoni tegument revealed a β-integrin subunit in the sub-tegumental layer [29]. Suppression of tetraspanin mRNA expression in schistosomes may affect lateral interactions with integrins in the tegument, and the parasite's ability to migrate through the lungs to the liver and mesenteries where they would mature. The binding partner(s) associated with Sm-TSP-1 or Sm-TSP-2, or any of the other three S. mansoni tegument tetraspanins, have yet to be identified. We have produced monoclonal antibodies to Sm-TSP-2 and these antibodies are being used to immunoprecipitate Sm-TSP-2 and its binding partners in an effort to unravel the tegumental tetraspanin web.
To assess the viability of dsRNA treated parasites in vivo, we injected tsp or luciferase dsRNA treated parasites into mice via the intramuscular route [56]. Recovery of adult worms from the mesenteries 4 weeks later was very low but was in agreement with other reports where newly transformed schistosomula were electroporated with dsRNAs prior to intramuscular injection into mice and subsequent recovery of adult worms from the mesenteries [41]. The natural route of S. mansoni infection is through percutaneous penetration of cercariae; exposure of laboratory mice to cercariae is generally performed via the abdomen or tail. Intramuscular injection of mice with schistosomula is not the natural infection route and consequently may have contributed to the low recovery rates. Despite the low recovery of adult parasites, we consistently over three experiments recovered significantly fewer worms from the mice injected with tsp dsRNA treated parasites. Moreover, tsp mRNA levels in those parasites that were recovered from mice were higher than levels in parasites cultured in vitro for the same time period after electroporation with dsRNAs, indicating that the parasites that survived in vivo had not succumbed to the effects of RNAi.
We envisage that interruption of Sm-TSP-1 and TSP-2 protein expression in the tegument of maturing schistosomula results in impaired turnover of the tegument apical membrane complex. Our observations from adults and schistosomula treated with Sm-tsp-2 dsRNA would indicate that a likely role for Sm-tsp-2 is in invagination and internalization of the surface membrane, and perhaps the closure and internalization of surface invaginations. This postulate is consistent with the suggestion that TSP-2 binds other parasite sub-surface and surface molecules in the tegument. The vaccine efficacy of TSP-2 may thus result from impairment of the surface recycling mechanisms in developing and adult schistosomes. While this impaired surface turnover was not deleterious to in vitro cultivated adult worms and schistosomula, the effect was particularly marked in treated schistosomula transferred into the host. In addition, schistosomes have the capacity to adsorb host blood molecules that mask antigenic epitopes from the host's immune system [7]. By affecting surface tegument development and turnover, suppression of tsp expression (and potential disruption of TEMs) may render the organism susceptible to immune recognition and clearance.
All animals were maintained in accordance with the guidelines of the Animal Ethics Committee (AEC) of Queensland Institute of Medical Research and the Institutional Animal Care and Use Committee (IACUC) of The University of Pennsylvania. All studies and procedures were reviewed and approved by the AEC and IACUC of Queensland Institute of Medical Research and The University of Pennsylvania respectively.
The Puerto Rican strain of S. mansoni and Biomphalaria glabrata snails were provided by the National Institutes of Allergy and Infectious Diseases Schistosomiasis Resource Centre at the Biomedical Research Institute (Rockville, Maryland, USA). B. glabrata infected with miracidia were exposed to incandescent light for 1h to obtain cercariae which were used to percutaneously infect 6–8 week old C57BL/6 female mice (www.jax.org). After 8 weeks, adult parasites were recovered by hepatic-portal perfusion and then washed three times with wash medium containing RPMI 1640, 1% antibiotic/antimycotic and 10 mM Hepes (www.invitrogen.com) before experimentation.
To obtain schistosomula, cercariae were passed through a 22-gauge emulsifying needle 25 times to mechanically shear the cercarial tails from the bodies [57]. The resulting schistosomula were isolated from free tails by centrifugation through a 60% percoll gradient [58], washed three times with washing medium and incubated at 37°C under 5% CO2 atmosphere before experimentation.
Three hour schistosomula (n = 500) were blocked in blocking buffer containing 1% goat serum in Dulbecco's Phosphate Buffered Saline (DPBS) containing MgCl2 and CaCl2 (www.invitrogen.com). Schistosomula were labelled with sera against recombinant Sm-TSP-1, Sm-TSP-2 or control pre-vaccination sera [28] diluted to 1∶50 in blocking buffer for 1 h. Secondary goat anti-mouse Ig-FITC (www.chemicon.com) was then introduced at 1∶100 dilution in blocking buffer for 1 h followed by 4% paraformaldehyde to fix the parasites. Incubations were carried out at 4°C and parasites were washed in DPBS between incubations. Approximately 200 schistosomula were examined using a Leica MRIRB microscope and DC500 camera (www.leica.com).
dsRNAs were prepared from DNA templates that were amplified by PCR from S. mansoni paired adult worm cDNA using primers flanked with T7 RNA polymerase promoter sequence (underlined) at the 5′ ends. A 523 bp fragment of the Sm-tsp-1 coding DNA was generated using primers (forward: 5′-TAATACGACTCACTATAGGGACTTGCTTCGGGACAACAAC-3′, reverse: 5′-TAATACGACTCACTATAGGGTTCGAAAGCTGCAATAGAAACA-3′) and a 565 bp fragment of the Sm-tsp-2 coding DNA was produced using primers (forward: 5′-TAATACGACTCACTATAGGGTGATTGTGGTTGGTGCACTT-3′, reverse: 5′-TAATACGACTCACTATAGGGGACCAATGCGAACAGAAACA-3′). The GenBank accession numbers for Sm-tsp-1 and Sm-tsp-2 are AF521093 and AF521091, respectively. The PCR products were then utilized as templates for synthesis of dsRNAs using the T7 Megascript kit (www.ambion.com), following the manufacturer's instructions. An irrelevant negative control, firefly luciferase dsRNA derived from pGL3-basic (www.promega.com), was prepared as described previously [31].
Adult schistosomes were cultured in vitro in Medium 199 (www.invitrogen.com) supplemented with 10% fetal calf serum (www.gembio.com), 1% antibiotic/antimycotic and 10 mM Hepes at 37°C under 5% CO2 atmosphere. Five pairs of adult worms were soaked in the presence of Sm-tsp-1, Sm-tsp-2 or luciferase dsRNAs at 1 µg/ml for 7 days with changes of media and dsRNAs every second day. Schistosomula were maintained at 37°C with 5% CO2 in Medium 169 [36] supplemented with 10% human AB serum (www.gembio.com) and mouse whole blood. Larval parasites (3 h old) were soaked in 1 µg/ml of Sm-tsp-1, Sm-tsp-2 or luciferase dsRNAs and cultured in vitro at 37°C under 5% CO2 atmosphere for 7, 14 and 21 days, with fresh changes of media, blood and dsRNAs every second day. Adult and larval parasites were washed in wash medium prior to RNA or protein extraction.
Newly transformed schistosomula were incubated in wash medium at 37°C with 5% CO2 for 3 h. Parasites were then resuspended in 50 µl of wash medium with 100 µg/ml of Sm-tsp-1, Sm-tsp-2 or luciferase dsRNAs and electroporated in a 4 mm cuvette at 125 V for 20 ms using a square-wave BTX ECM 830 electroporator (www.btxonline.com). After three washes in wash medium, schistosomula were counted and 2000 were injected intramuscularly into each C57BL/6 female mouse (3 mice per group) using a 23-gauge needle. Adult worms were perfused 28 days later to assess the number of worms that had matured and reached the mesenteries.
RNA was isolated from parasites using RNeasy Mini kit (www.qiagen.com) and then treated with Turbo DNA-free endonuclease (www.ambion.com) to remove contaminating genomic DNA. The quantity of RNA was measured on a Nanodrop Spectrophotometer (www.nanodrop.com) and 250 ng of total RNA, SuperScript II reverse transcriptase (www.invitrogen.com) and oligo dT15 primer (www.promega.com) were used to synthesize first strand cDNA.
The following primers were designed for real-time qRT-PCR; Sm-TSP-1 (forward: 5′-TGGTTGTGCTTATTGGGTTG-3′ and reverse: 5′-TGATGTCTTGTGCCTCTGGT-3′); Sm-TSP-2 (forward: 5′-CGAAATTGAACCCCCACTAC-3′ and revere: 5′-CATGCTCCAAACATCCCTAAA-3′); Sm-Paramyosin (forward: 5′-CGTGAAGGTCGTCGTATGGT-3′ and reverse 5′-GACGTTCAAATTTACGTGCTTG-3′) and Sm-α-tubilin (forward: 5′-CCAGCAAAATCAGATGGTGAA-3′ and reverse: 5′-TTGACATCCTTGGGGACAAC-3′). qRT-PCR was conducted in triplicate and each reaction underwent 40 amplification cycles using an Applied Biosystems 7500 real-time PCR system (www.appliedbiosystems.com) with cDNA equivalent to 20 ng of total RNA, 50 nM of primers and SYBR green PCR Master Mix (www.appliedbiosystems.com). Dissociation curves were generated for each sample to verify the amplification of a single PCR product. Sm-tsp transcript levels were calculated relative to Sm-paramyosin in test and irrelevant dsRNA treated parasites using the 2−ΔΔCt method [59], and data was expressed as percent differences. For relative endogenous expression of tsp mRNAs in schistosome life cycle stages, Sm-α-tubulin was used as the endogenous standard. Sm-paramyosin was used as the housekeeping gene for analyzing Sm-tsp expression in RNAi experiments.
RNAi-treated adult parasites and schistosomula were harvested after 7 days and then lysed with 1% Triton X-100 in Tris buffered saline supplemented with complete protease inhibitor cocktail EASYpacks (www.roche.com). Protein concentrations of lysates were determined using a BCA protein assay kit (www.pierce.com), and lysates were electrophoresed in 12% SDS-PAGE gels at concentrations of 2, 1, 0.5 and 0.25 µg total protein per well. Proteins were transferred to nitrocellulose membrane (Hybond-ECL, www.gehealthcare.com) and then probed with either anti-Sm-TSP-2 (3H5/2) monoclonal antibody supernatants (L. Cooper, M. Tran and A. Loukas, unpublished) diluted 1∶1,000 followed by anti-mouse Ig-HRP (www.chemicon.com) diluted 1∶5,000. Reactive proteins were detected by ECL (www.gehealthcare.com) as per the manufacturer's instructions. To assess equal protein loading, nitrocellulose membranes were stripped after reacting with anti-TSP-2 antibodies and then re-probed with anti-paramyosin (Sm4B1) monoclonal antibody supernatants [60] diluted at 1∶1,000 followed by anti-mouse Ig-HRP. Experiments were repeated three times and protein quantities in gel bands were determined using Syngene Tools and software (www.syngene.com).
Adult parasites and schistosomula were soaked in 1 µg/ml of Sm-tsp or luciferase dsRNAs for 7 days at 37°C under 5% CO2 atmosphere, washed three times in wash medium and then fixed in 3% glutaradehyde in 0.1M phosphate buffer at pH 7.4, followed by fixation in potassium ferricyanide-reduced osmium tetroxide. After fixation, parasites were dehydrated in acetone and embedded in Epon Resin (ProSciTech). Ultrathin sections were mounted onto copper grids, contrasted in uranyl acetate and lead citrate and examined and photographed using a JEM 1011 transmission electron microscope operated at 80 kV and equipped with a digital camera.
A morphometric approach was employed to quantify possible changes to tegument structure in schistosomula treated with Sm-tsp-2 relative to those treated with luciferase dsRNA. Point counting stereology [61], [62] was used to measure the volume of tegument occupied by vacuolar compartments or tegument invaginations in the tegument. Such regions were evident as clear spaces in TEM sections. Twenty individual schistosomula were selected at low magnification in the TEM. For each parasite, the first region of tegument observed that fulfilled the two criteria below was photographed at ×10,000 magnification. Criteria for selection were, firstly, that the region photographed was from the lateral aspect of a parasite that was clearly longer than wide and in which internal organs were present, and secondly, that the region was not excessively spinous. Volume density of vacuolar compartments of tegument were estimated using grids generated by Image J analysis software (NIH Besthesda), and were calculated as the number of points on the grid intersecting a vacuolar space divided by the number of points intersecting the tegument. This was measured across the entire profile of the tegument in each electron micrograph, so that only one measure was obtained for each schistosomulum. In addition to the volume density measure, the thickness of the tegument was measured at 10 different points using the line tool in Image J. For each measure, a line was drawn digitally on each micrograph from the basal membrane of the tegument to the apical membrane. Regions where the tegument was excessively invaginated, and those containing isolated spines and sensory receptors were not measured. The 10 thickness measurements were averaged for each schistosomulum.
All data are presented as the mean±standard error. Differences between groups were assessed for statistical significance using Student t-test (GraphPad Prism Software, www.graphpad.com). A statistically significant difference for a particular comparison was defined as p<0.050.
|
10.1371/journal.pbio.3000136 | A mutagenesis screen for essential plastid biogenesis genes in human malaria parasites | Endosymbiosis has driven major molecular and cellular innovations. Plasmodium spp. parasites that cause malaria contain an essential, non-photosynthetic plastid—the apicoplast—which originated from a secondary (eukaryote–eukaryote) endosymbiosis. To discover organellar pathways with evolutionary and biomedical significance, we performed a mutagenesis screen for essential genes required for apicoplast biogenesis in Plasmodium falciparum. Apicoplast(−) mutants were isolated using a chemical rescue that permits conditional disruption of the apicoplast and a new fluorescent reporter for organelle loss. Five candidate genes were validated (out of 12 identified), including a triosephosphate isomerase (TIM)-barrel protein that likely derived from a core metabolic enzyme but evolved a new activity. Our results demonstrate, to our knowledge, the first forward genetic screen to assign essential cellular functions to unannotated P. falciparum genes. A putative TIM-barrel enzyme and other newly identified apicoplast biogenesis proteins open opportunities to discover new mechanisms of organelle biogenesis, molecular evolution underlying eukaryotic diversity, and drug targets against multiple parasitic diseases.
| Plasmodium parasites, which cause malaria, and related apicomplexan parasites evolved from photosynthetic algae that acquired their chloroplast through two successive endosymbioses. Although no longer photosynthetic, the apicomplexan plastid—or apicoplast—was retained in these pathogens and provides critical metabolites during host cell infection. The apicoplast is of major interest for its unique biology and potential to yield new antimalarial drug targets. Here, we focused on the critical genes required to grow, divide, and inherit new apicoplasts during parasite replication. Given the apicoplast’s divergent evolution, most of these cannot be recognized by their homology to genes with known functions. Instead, we overcame significant technical challenges in the Plasmodium experimental system to perform an unbiased screen to search for these critical genes. Our screen has uncovered new genes with intriguing evolution and function that open up opportunities to understand and ultimately exploit apicoplast biology. Finally, assigning new, essential gene functions in Plasmodium parasites remains a daunting task. The successful identification of essential gene functions using an unbiased approach in this study provides a viable route for expansion of this screen or developing screens for other novel Plasmodium pathways in the future.
| Plasmodium spp., which cause malaria, and related apicomplexan parasites are important human and veterinary pathogens. These disease-causing protozoa are highly divergent from well-studied model organisms that are the textbook examples of eukaryotic biology, such that parasite biology often reveals striking eukaryotic innovations. The apicoplast, a nonphotosynthetic plastid found in apicomplexa, is one such “invention” [1, 2]. These intracellular parasites evolved from photosynthetic algae that acquired their plastids through secondary endosymbiosis [3]. During secondary endosymbiosis, a chloroplast-containing alga, itself the product of primary endosymbiosis, was taken up by another eukaryote to form a secondary plastid [4]. Despite the loss of photosynthesis in apicomplexans, the apicoplast contains several prokaryotic metabolic pathways, is essential for parasite replication during human infection, and is a target of antiparasitic drugs [5–8].
There are many traces of the apicoplast’s quirky evolution in its present-day cell biology, in particular the pathways for its biogenesis. Like other endosymbiotic organelles, the single apicoplast cannot be formed de novo and must be inherited by its growth, division, and segregation into daughter cells. The few molecular details we have about apicoplast biogenesis hint at the major innovations that have occurred in the process of adopting and retaining this secondary plastid. The apicoplast is bound by four membranes acquired through successive endosymbioses, such that apicoplast proteins transit through the endoplasmic reticulum (ER) and use a symbiont-derived ER-associated degradation (ERAD)-like machinery (SELMA) to cross the two new outer membranes [9]. Curiously, autophagy-related protein 8 (Atg8), a highly conserved eukaryotic protein and key marker of autophagosomes, localizes to the apicoplast and is required for its inheritance in Plasmodium and the related apicomplexan parasite Toxoplasma gondii [10–12]. While SELMA and PfAtg8 are clear examples of molecular evolution in action, other novel or repurposed proteins required for apicoplast biogenesis remain undiscovered.
So far, new apicoplast biogenesis proteins have primarily been discovered through candidate approaches. SELMA was first identified as homologs of the ERAD machinery encoded in the nucleomorph, the remnant nucleus of the eukaryotic symbiont found in some algal secondary plastids [13]. Nuclear-encoded versions of SELMA containing apicoplast-targeting sequences were then detected in the genomes of apicomplexan parasites (which lack a nucleomorph) [14]. Because the apicoplast proteome is enriched in proteins likely to perform biogenesis functions such as protein import or genome replication, several apicoplast-targeted proteins of unknown function have also been shown to be required for its biogenesis [15]. ATG8’s novel apicoplast function was discovered serendipitously by its unexpected localization on the apicoplast instead of autophagosomes [10]. Though candidate approaches have yielded new molecular insight [16–18], in general they are indirect and may bias against novel pathways.
In blood-stage Plasmodium falciparum, a method to chemically rescue parasites that have lost the apicoplast has paved the way for functional screens [19–21]. Addition of isopentenyl diphosphate (IPP) to the growth media is sufficient to reverse growth inhibition caused by apicoplast loss because it is the only essential metabolic product of the apicoplast in the blood stage. Taking advantage of the apicoplast chemical rescue, we recently took the first unbiased approach to discover a new apicoplast biogenesis protein [22]. We first screened for small-molecule inhibitors that specifically disrupt apicoplast biogenesis in P. falciparum. Subsequent target identification led us to a membrane metalloprotease, the P. falciparum ATP-dependent zinc metalloprotease 1 (PfFtsH1), with an unexpected role in apicoplast biogenesis. This chemical genetic screen has the advantage of unbiased sampling of druggable targets in apicoplast biogenesis pathways. Unfortunately, it lacks throughput given the painstaking process of mapping inhibitors to their molecular targets.
Forward genetic screens are widely performed to uncover novel cellular pathways, such as those required for apicoplast biogenesis. Recently, genome-scale deletion screens performed in several apicomplexan parasites have uncovered a plethora of essential genes of unknown function [23–25]. Several challenges impede large-scale functional analysis of these essential genes. First, targeted gene modifications are still slow and labor intensive in P. falciparum, the most deadly of the human malaria species, despite an available in vitro blood culture system [26, 27]. Second, efficient methods for generating conditional mutants, such as RNA interference (RNAi) or Clustered Regularly Interspaced Palindromic Repeats interference (CRISPRi) systems, are lacking in all apicomplexan organisms [28]. Finally, high-throughput, single-cell phenotyping for important functions need to be developed [29]. Overcoming these significant limitations, we designed a forward genetic screen using chemical mutagenesis, apicoplast chemical rescue, and a fluorescent reporter for apicoplast loss to identify essential apicoplast biogenesis genes in blood-stage P. falciparum. The screen identified known and novel genes required for apicoplast biogenesis and is, to our knowledge, the first forward genetic screen to assign essential cellular functions in P. falciparum.
The apicoplast must be propagated during parasite replication, such that biogenesis defects result in newly replicated daughter parasites that do not contain an apicoplast. Because no clearance mechanism is known to eliminate the apicoplast in asexual blood-stage parasites, defective organelle biogenesis resulting in loss of inheritance is likely the major cause of apicoplast loss. To isolate rare P. falciparum mutants that have lost their apicoplast due to biogenesis defects, we set out to design a live-cell reporter for apicoplast loss. The apicoplast contains a prokaryotic caseinolytic protease (Clp) system composed of a Clp chaperone (ClpC) that recognizes and unfolds substrates and a Clp protease (ClpP) that degrades the recognized substrates [30–32]. We hypothesized that (1) in the presence of a functional apicoplast, ClpCP could be co-opted to degrade and turn “off” a fluorescent reporter, whereas (2) loss of the apicoplast would result in loss of ClpCP activity and therefore turn “on” the reporter (Fig 1A).
Clp substrates are typically recognized by unstructured degron sequences, the best studied of which is a transfer-messenger RNA (ssrA) that appends a short peptide to the C-terminus of translationally stalled proteins [33, 34]. However, the substrate specificity of the apicoplast ClpCP system has yet to be defined. Reasoning that the PfClpCP homolog might recognize similar degrons as bacterial or algal Clp systems, we tested two degrons recognized by Escherichia coli ClpXP—the E. coli ssrA peptide (EcssrA) and X7—and the predicted ssrA peptide in the red alga Cyanidium caldarium, CcssrA (S1 Table) [35–37]. To assess their functionality in P. falciparum, the C-terminus of an apicoplast-targeted green fluorescent protein (acyl-carrier protein leader peptide [ACPL-GFP]) was modified with each of the degrons (Fig 1B). A cytosolic mCherry marker was also expressed on the same promoter as ACPL-GFP via a T2A “skip” peptide to normalize apicoplast GFP levels to stage-specific promoter expression [38]. Each construct was then integrated into an ectopic attB locus in Dd2attB parasites to generate the reporter strains [39].
For each reporter strain, the ratio of GFP:mCherry fluorescence (as detected by flow cytometry) was assessed in untreated parasites containing an intact apicoplast, designated apicoplast(+), and in parasites treated with actinonin (which causes apicoplast loss via inhibition of FtsH1) and rescued with IPP rendering them apicoplast(−). In the absence of a degron, the GFP:mCherry ratio decreased modestly in apicoplast(−) parasites compared to apicoplast(+) (Fig 1C, S1 Fig and S1 Data). Though GFP was present in both populations, live fluorescence microscopy confirmed its localization to a branched structure characteristic of the apicoplast in apicoplast(+) parasites and to dispersed punctae as previously reported in apicoplast(−) parasites (S1 Fig) [19]. Addition of each of the three degrons to the C-terminus of ACPL-GFP caused a 60%–84% decrease in the GFP:mCherry ratio in apicoplast(+) parasites, consistent with specific degradation of GFP (Fig 1C). In ACPL-GFP-EcssrA and ACPL-GFP-CcssrA populations, the GFP:mCherry ratio recovered to 88% and 44% relative to the ACPL-GFP population in apicoplast(−) parasites, respectively (Fig 1C and 1E, S1 Fig and S1 Data). Unexpectedly, ACPL-GFP-X7 populations showed further reduction of GFP:mCherry in apicoplast(−) parasites (Fig 1C, S1 Fig and S1 Data). Notably, cytoplasmic mCherry levels were similar between apicoplast(+) and apicoplast(−) parasites in all reporter strains, suggesting that the differences in GFP levels were not due to altered expression levels (S1 Fig and S1 Data). Of the 3 degrons tested, both the EcssrA and CcssrA peptide caused apicoplast-specific GFP degradation as designed.
We further characterized ACPL-GFP-EcssrA because the greatest recovery of GFP fluorescence following apicoplast loss was observed with this reporter (Fig 1C). Consistent with degradation of GFP in the apicoplast being dependent on EcssrA peptide, ACPL-GFP-EcssrA protein was detected at a significant level only in apicoplast(−) parasites, while unmodified ACPL-GFP was detected in both apicoplast(+) and (−) parasites (Fig 1D). Of note, cleavage of the apicoplast-targeting ACPL sequence does not occur in apicoplast(−) parasites, resulting in an ACPL-GFP protein of higher molecular weight compared to apicoplast(+) parasites [19]. Flow cytometry and live fluorescence microscopy confirmed that apicoplast(+) parasites displayed only cytosolic mCherry fluorescence, whereas apicoplast(−) parasites displayed both cytosolic mCherry and dispersed, punctate GFP fluorescence (Fig 1E and 1F). As expected, addition of IPP alone did not result in significant formation of apicoplast(−) parasites or recovery of ACPL-GFP-EcssrA (S1 Fig and S1 Data). Taken together, these results demonstrate that ACPL-GFP-EcssrA serves as a specific reporter for apicoplast loss in P. falciparum.
Next, we used the ACPL-GFP-EcssrA reporter strain to perform a phenotypic screen for apicoplast(−) mutants (Fig 2A). Ring-stage parasites were mutagenized with the alkylating agents ethyl methanesulfonate (EMS) or N-ethyl-N-nitrosourea (ENU) with the expectation that this would generate a more diverse population of parasites, some of which harbor mutations in apicoplast biogenesis genes rendering them apicoplast(−) [21, 22]. To rescue the lethality of apicoplast loss, mutagenized parasites were supplemented with 200 μM IPP in the growth media. A control group of non-mutagenized parasites was also cultured with IPP to assess whether apicoplast(−) mutants naturally occurred over time in the absence of selective pressure to maintain the organelle.
After two replication cycles to allow for initial apicoplast loss, mCherry-expressing parasites displaying GFP fluorescence greater than about 70% percentile were isolated by fluorescence-activated cell sorting (FACS). Selected mCherry(+) and GFP(+) parasites were then allowed to propagate to a detectable parasitemia before being subjected to another round of FACS. In one exemplary ENU-mutagenized population, a distinct population of GFP-positive parasites began to emerge after just two rounds of FACS (Fig 2B). Other populations showed enrichment beginning after three rounds of FACS. No significant enrichment of GFP signal was observed when parasites were (1) grown with IPP over time without FACS or (2) grown without IPP and subjected to FACS, suggesting that we specifically enriched for apicoplast(−) mutants. Parasites from the final enriched apicoplast(−) populations were individually sorted to generate apicoplast(−) clones derived from single parasites. Each clonal population was then re-checked for IPP-dependent growth and GFP fluorescence.
To determine the genetic basis of apicoplast loss, we sequenced the genomes of 51 apicoplast(−) clones (mutagenized = 40; non-mutagenized = 11) and three parent populations. Sequenced reads were mapped to the genome of P. falciparum 3D7 (version 35) with an average read depth of 26 for all samples. Notably, while the apicoplast genome was sequenced at an average depth of 16 reads in the parent populations, it was only detected with an average of 0.02 reads in apicoplast(−) clones (Fig 2C and S2 Data). Because the organellar genome is a marker for the apicoplast, the absence of the apicoplast genome confirmed the loss of the apicoplast in the sequenced clones [19, 22].
We next performed single nucleotide variant (SNV) analysis. A raw variant list was generated for each sample by comparison to the reference 3D7 genome and included SNVs found in parent populations at ≥5% allele frequency (minimum one read), and in apicoplast(−) clones at ≥90% (minimum five reads). Any SNV identified in apicoplast(−) clones that was also identified in any of the three parent populations was removed. We also filtered out SNVs detected in noncoding regions or resulting in synonymous amino acid changes in coding regions. Finally, SNVs identified in hypervariable regions of the genome (including the rifin, stevor, and EMP gene families) and/or previously annotated in the PlasmoDB single nucleotide polymorphism (SNP) database were excluded. After these filtering steps, 23 apicoplast(−) clones had at least one but no more than three SNVs that differed from the parent populations (Fig 2D; S2 Table).
Because genes required for apicoplast biogenesis ought to be essential, we used essentiality data from literature or whole-genome deletion screens performed in blood-stage P. falciparum and P. berghei to prioritize gene candidates [24, 25]. Of 18 unique SNVs identified, 12 were in genes categorized as “essential” in blood-stage P. falciparum and/or P. berghei (Table 1 and S2 Table). Although PfFtsH1 (Pf3D7_1239700) is categorized as “dispensable” in the P. falciparum deletion screen, it has been shown experimentally to be essential [25]. Overall, a mutation in one of these 12 essential gene candidates was identified in each of the 23 apicoplast(−) clones, consistent with the mutation causing apicoplast loss. Potentially disruptive mutations included a I437S variant in the known apicoplast biogenesis protein (PfFtsH1), truncation of Atg7 (PfAtg7) likely required for a cytoplasmic pathway for apicoplast biogenesis, and truncations of three proteins of unknown function (Pf3D7_0518100, Pf3D7_1305100, and Pf3D7_1363700) (Table 1). The remaining candidates contained point mutations and had no known prior function in apicoplast biogenesis or localization to the organelle.
PfFtsH1 is an apicoplast membrane metalloprotease that was previously identified in a chemical genetic screen as the target of actinonin, an inhibitor that disrupts apicoplast biogenesis. Subsequent knockdown of PfFtsH1demonstrated that it is required for apicoplast biogenesis [22]. We hypothesized that the I437S variant identified in our screen disrupted PfFtsH1 function, leading to apicoplast loss. PfFtsH1 contains both an ATPase and peptidase domain. To test the effect of I437S on enzyme activity, we compared the activity of I437S with that of wild-type (WT) enzyme, an ATPase-inactive E249Q variant, and a peptidase-inactive D493A variant (Fig 3A). All enzymes were purified without the transmembrane domain as previously described (S3 Fig) [22].
We first measured peptidase activity on a fluorescent casein substrate. Neither the I437S nor the D493A variant had detectable peptidase activity, whereas WT enzyme turned over substrate at 0.4 min−1 (Fig 3B, S3 Table and S3 Data). Similarly, the I437S and E249Q variants showed no detectable ATP hydrolysis activity, in contrast to both WT and the D493A variants (Fig 3C, S3 Table and S3 Data). Taken together, these results validate the identified missense mutation leading to expression of inactive PfFtsH1 I437S variant as causative for apicoplast loss.
Though non-apicoplast proteins are expected to play a role in apicoplast biogenesis, all apicoplast biogenesis proteins validated so far have localized to the apicoplast because this criterion is most often used to select candidates. A significant advantage of our forward genetic screen is that it can uncover non-apicoplast pathways required for apicoplast biogenesis, which are biased against by other approaches. A cytoplasmic protein strongly identified in our screen was PfAtg7 (Pf3D7_1126100), which contained a nonsense mutation causing a protein truncation at position Q719. The premature stop codon was upstream of the E1-like activating domain, consistent with PfAtg7 loss of function (Fig 4A). In yeast and mammalian cells, Atg7 is required for conjugation of Atg8 to the autophagosome membrane [41]. PfAtg7 has not specifically been implicated in apicoplast biogenesis; however, PfAtg8 has been shown to localize to the apicoplast membrane and be required for apicoplast biogenesis. In analogy with its role in autophagy, PfAtg7 is likely required for conjugating PfAtg8 to the apicoplast membrane. Therefore, we suspected the loss-of-function mutant we identified caused apicoplast loss via loss of membrane-conjugated PfAtg8 [12].
To confirm PfAtg7’s role in apicoplast biogenesis, we generated a P. falciparum strain in which it is conditionally expressed. The endogenous PfAtg7 locus in a NF54 strain harboring a CRISPR cassette was modified with a C-terminal triple hemagglutinin (HA) tag and a 3′ untranslated region (UTR) RNA aptamer sequence that binds a tetracycline repressor (TetR) and development of zygote inhibited (DOZI) fusion protein to generate PfAtg7-TetR/DOZI [42, 43]. In the presence of anhydrotetracycline (ATc), the 3′UTR aptamer is unbound, and PfAtg7-3×HA protein was detectable, albeit at low levels consistent with its low expression in published transcriptome data (PlasmoDB.org) (S2 Fig). Removal of ATc causes binding of the TetR/DOZI repressor, which resulted in undetectable PfAtg7 protein levels by western blot within one replication cycle (S2 Fig). Immunofluorescence of HA-tagged PfAtg7 showed diffuse cytoplasmic localization in the presence of ATc and overall reduction of fluorescence signal upon removal of ATc (S4 Fig). Parasitemia also decreased over the course of several replication cycles, consistent with a previous study showing that PfAtg7 is essential (Fig 4B and S4 Data) [40].
We tested whether the requirement for PfAtg7 was due to its role in apicoplast biogenesis. Growth inhibition caused by PfAtg7 knockdown was partially rescued by addition of IPP (Fig 4B and S4 Data). Furthermore, in −ATc/+IPP parasites, the apicoplast marker acyl-carrier protein (ACP) mislocalized from a discrete organellar localization to multiple cytoplasmic puncta, a hallmark of apicoplast loss (Fig 4C) [19]. Loss of transit peptide cleavage of the apicoplast protein ClpP also confirmed apicoplast loss, because mislocalized apicoplast proteins do not undergo removal of their targeting sequences (Fig 4D) [15, 44]. IPP rescue of PfAtg7 knockdown parasites was incomplete, raising the possibility that PfAtg7 is also required for a non-apicoplast function. Alternatively, PfAtg7 knockdown may cause stalling of apicoplast morphologic development leading to general cellular toxicity that cannot be fully rescued with IPP until apicoplast loss is complete. To test these models, instead of PfAtg7 knockdown followed by apicoplast loss, we first induced apicoplast loss with actinonin and then assessed the effects of PfAtg7 knockdown. PfAtg7 knockdown in apicoplast(−) parasites did not cause any additional growth defect and was fully rescued by IPP, suggesting that the partial rescue observed in apicoplast(+) parasites was due to the order of disruption (S5 Fig and S4 Data). These results confirmed that PfAtg7 is required for apicoplast biogenesis and likely is its only essential function.
Finally, to determine whether PfAtg7’s role in apicoplast biogenesis was via PfAtg8 membrane conjugation, we transfected PfAtg7-TetR/DOZI with a transgene encoding GFP-PfAtg8. In this strain, GFP-PfAtg8 primarily localized to a branched structure in schizonts, consistent with its previously described apicoplast localization (Fig 4E) [10, 45]. Upon PfAtg7 knockdown, GFP-PfAtg8 localization to this discrete structure was lost, and accumulation in the cytoplasm was observed within a single replication cycle prior to apicoplast loss (Fig 4E). This result suggests that, like yeast and mammalian Atg7 homologs, PfAtg7 has a conserved function in conjugating Atg8 to lipids. Altogether, PfAtg7 stood out as a cytoplasmic protein required for apicoplast biogenesis identified in our screen.
The real power of forward genetics is the ability to uncover novel pathways without any a priori knowledge. Therefore, we next turned our attention to the nearly 50% of the Plasmodium genome annotated as “conserved Plasmodium protein, unknown function.” Three candidate genes (Pf3D7_0518100, Pf3D7_1305100, and Pf3D7_1363700) encoding proteins of unknown function were identified by nonsense mutations that caused protein truncation. The position of the premature stop codon near the 5′ end (Pf3D7_0518100, Pf3D7_1305100) or in the middle (Pf3D7_1363700) of the genes suggested that these were loss-of-function mutations (Fig 5A). Incidentally, all were also identified in a recently published proteomic dataset of apicoplast proteins, and immunofluorescence colocalization with the apicoplast marker ACP confirmed that Pf3D7_0518100 and Pf3D7_1305100 are localized to the apicoplast (S4 Fig) [15].
Therefore, we assessed whether knockdown of these genes disrupted apicoplast biogenesis. Similar to the experiments performed to validate PfAtg7, ATc-regulated knockdown strains for each of the candidate genes were generated. Upon ATc removal, protein levels for each gene decreased within 24 hours as detected by western blot (S2 Fig). Significant growth inhibition was also observed for all candidate genes tested, with varying kinetics of growth inhibition observed for each candidate (Fig 5B, 5E and 5H and S4 Data). Of note, the gene essentiality demonstrated here for Pf3D7_1305100 and Pf3D7_1363700 confirmed whole-genome essentiality data reported in P. berghei and/or P. falciparum. However, the essentiality of Pf3D7_0518100 did not agree with its “dispensable” annotation in the P. falciparum dataset. IPP supplementation reversed growth inhibition for all the candidates, demonstrating that their essentiality was due to an apicoplast-specific function (Fig 5B, 5E and 5H and S4 Data). Finally, mislocalization of ACP and loss of transit peptide cleavage of ClpP confirmed apicoplast loss for all candidates (Fig 5C, 5D, 5F, 5G, 5I and 5J). Because these genes have so far lacked any functional annotation and given their shared knockdown phenotype, we designated them “apicoplast-minus, IPP-rescued” (amr) genes: amr1 (Pf3D7_1363700), amr2 (Pf3D7_0518100), and amr3 (Pf3D7_1305100). Taken together, we successfully identified three novel proteins of unknown function required for apicoplast biogenesis, prioritizing these amr genes for functional studies.
To set up future functional studies, we noted that PfAMR1 contained a TIM-barrel domain with closest homology to indole-3-glycerol phosphate synthase (IGPS), a highly conserved enzyme in the tryptophan (trp) biosynthesis pathway [46–48]. This was surprising because Plasmodium and the related apicomplexan parasite Toxoplasma are trp auxotrophs [49–51]. Indeed, analysis of >30 apicomplexan genomes did not detect any of the other six trp biosynthetic enzymes, except the terminal enzyme tryptophan synthase (TS)-β, which was horizontally transferred into Cryptosporidium spp. [52]. Therefore, we suspected that PfAMR1 may have a function unrelated to trp biosynthesis.
To test the conservation of active-site residues, we aligned the sequences of several known IGPSs with IGPS homologs identified from P. falciparum, T. gondii, and Vitrella brassicaformis (S6 Fig) [53–55]. V. brassicaformis, a member of the Chromerids, is the closest free-living, photosynthetic relative to apicomplexan parasites. It contains a secondary plastid with the same origin as the apicoplast but, as a free-living alga [56], is also expected to have intact trp biosynthesis. Known catalytic and substrate binding residues based on enzyme structure-function studies performed in bacteria were first identified [57]. For known IGPSs and one of the V. brassicaformis IGPS homologs (Vbra_4894), the key catalytic and substrate binding residues were all conserved, despite the vast evolutionary distance between bacteria, metazoans, and Chromerids/apicomplexans (Fig 6A). However, in PfAMR1 and the other two V. brassicaformis homologs, key functional residues were not conserved. Based on the conservation of functional residues, we separated these sequences into two groups: “canonical IGPS proteins” (which have been shown, or are likely, to encode for IGPS activity) and “IGPS-like proteins” (e.g., PfAMR1), which we suggest have functionally diverged.
We next looked at the pattern of canonical IGPS versus IGPS-like proteins through two biological transitions: loss of trp biosynthesis (Vitrella versus Plasmodium spp.) and loss of the apicoplast (Plasmodium versus Cryptosporidium spp.) (Fig 6B). As expected for a role in trp biosynthesis, canonical IGPS proteins were retained until the emergence of parasitism. In addition, genes encoding the remaining set of enzymes for trp biosynthesis were identified in the V. brassicaformis genome [58]. Unlike canonical IGPSs, however, IGPS-like proteins were retained in parasites that have lost trp biosynthesis. Instead, loss of IGPS-like proteins is associated only with loss of the apicoplast in Cryptosporidium spp. This pattern of acquisition and loss of IGPS-like proteins suggests an apicoplast-specific function separate from trp biosynthesis.
Finally, we performed functional complementation to test the biochemical activity of canonical IGPS and IGPS-like genes from V. brassicaformis and P. falciparum. An E. coli strain (trpC9800) containing an inactivating mutation in trpC, the E. coli IGPS homolog, was grown on minimal agar (M9) [59]. As expected, trpC9800 was dependent on trp supplementation for growth (Fig 6C). Complementation with the Vbra_4894 homolog restored trpC9800 growth on M9, comparable to that of an E. coli strain with intact trp biosynthesis, suggesting that Vbra_4894 is an IGPS protein (Fig 6D). In contrast, neither PfAMR1 nor any of the IGPS-like genes were able to functionally replace trpC, supporting an alternative biochemical function. Because glutathione S-transferase (GST)-tagged complemented protein could not be detected in any strain, we cannot rule out that the lack of functional complementation was due to differential protein expression or solubility below the detection limit of western blot. However, isopropyl β-D-1-thiogalactopyranoside (IPTG) induction of protein expression was toxic for all complemented strains, suggesting all constructs supported protein expression. Overall, we propose that AMR1 has evolved a new biochemical function required for apicoplast biogenesis.
Apicoplast biogenesis provides a fascinating window into molecular evolution, including examples of proteins that have been reused (e.g., translocon on the inner chloroplast membrane/translocon on the outer chloroplast membrane [TIC/TOC] complexes) [17, 60, 61], repurposed (e.g., Atg8, symbiont-derived ERAD-like machinery [SELMA]) [12, 14], or newly invented in the process of serial endosymbioses. Overcoming significant technical challenges in the Plasmodium experimental system, we designed a forward genetic screen to identify essential apicoplast biogenesis pathways. This singular screen opens up opportunities to discover evolutionary innovations obscured by candidate-based approaches, including cytoplasmic pathways and genes lacking any functional annotations. In addition to confirming the role of PfFtsH1 in apicoplast biogenesis and identification of PfAtg8 conjugation machinery, we identified several proteins of unknown function required for apicoplast biogenesis that have so far gone undetected. Because our reporter specifically looked for apicoplast loss, we cannot rule out the possibility that some identified genes may be involved in maintenance of already formed apicoplasts. However, because no clearance mechanism is known for defective apicoplasts, we are not aware of any pathway by which defective apicoplasts would lead to organelle loss independent of parasite replication.
One surprising gene we identified was PfAMR1, which encodes a TIM-barrel domain found in diverse enzymes catalyzing small-molecule reactions [46]. PfAMR1 may have evolved from gene duplication of IGPS, an enzyme in the trp biosynthetic pathway [47]. However, the evolutionary pattern of retention in apicomplexan parasites lacking tryptophan biosynthesis and loss in Cryptosporidium spp., concomitant with plastid loss, supports a critical function of PfAMR1 in the apicoplast independent of tryptophan biosynthesis. We hypothesize that PfAMR1 may be involved in biosynthesis of a specialized lipid or signaling molecule required specifically for building new plastids in this lineage. Multiple new amr genes identified in this study provide striking examples of the unexpected findings enabled by unbiased screens.
Uncovering novel apicoplast biogenesis pathways also has important biomedical applications. While targeting the metabolic function of the apicoplast has been a major strategy for antimalarial drug discovery [62], it has become apparent that apicoplast biogenesis is equally as, or likely more, valuable as a therapeutic target [22]. These distinct pathogen pathways are nonetheless required in every proliferative stage of the Plasmodium life cycle and conserved among apicomplexan parasites. Targeting apicoplast biogenesis has the benefit of efficacy against multiple Plasmodium life stages and multiple pathogens. Consistent with this broad utility, antibiotics that inhibit translation in the apicoplast and disrupt its biogenesis are used clinically for malaria prophylaxis, acute malaria treatment, and treatment of babesiosis and toxoplasmosis [7, 8, 63].
Until now, a forward genetic screen for essential pathways in blood-stage Plasmodium has not been achieved. Previous screens in murine P. berghei and the human malaria parasite P. falciparum identified nonessential genes required for gametocyte formation [64, 65], the developmental stage required for mosquito transmission. Functional screens for essential pathways have been impeded by several technical challenges, including the low transfection efficiency of P. falciparum, in vivo growth requirement of P. berghei, and absence of efficient methods for generating conditional mutants in both organisms [28]. Nonetheless, genome-scale deletion screens in P. berghei and P. falciparum using a homologous recombination-targeted deletion library or saturating transposon-based mutagenesis, respectively, have revealed a plethora of essential genes [24, 25]. Functional assignment of these essential genes is a priority. In this context, the apicoplast biogenesis screen presented here is a major milestone towards unbiased functional identification of novel, essential genes.
A top priority for “version 2.0” is to expand the screen to genome scale, maximizing our ability to uncover novel pathways. Apicoplast biogenesis is a complex process encompassing a multitude of functions and is expected to require hundreds of gene products. The identification of 12 candidate genes and our sparse sampling of known genes suggest that the current screen is far from saturating. The most significant bottleneck is the dependence of this screen on chemical mutagenesis. Mapping mutations by whole-genome sequencing limits the number of mutants that can be analyzed. Even for sequenced clones, less than half had a detectable point mutation in a coding region. The remaining apicoplast(−) clones may have contained mutations in noncoding regions or other types of mutations that are more difficult to detect (insertion, deletions, or copy number variations). Particularly for apicoplast(−) clones selected from non-mutagenized conditions, we considered the possibility that some parasites might spontaneously fail to inherit the apicoplast due to mechanical defects during parasite replication; these daughter cells resulting from erroneous apicoplast division and segregation usually would not survive but are rescued by IPP. Finally, specific mutations identified in candidate genes also need to be validated one by one. In this study, four nonsense mutations were validated by conditional knockdown, while a missense mutation in PfFtsH1 was validated using an available activity assay. Although we were able to demonstrate loss of function as result of the PfFtsH1 I437S variant, other missense mutations identified in genes of unknown function will be challenging to follow up with available genetic tools.
Given these limitations, an alternative mutagenesis method will increase the screen throughput. Options include adaptation of the piggyBac transposon developed for the P. falciparum deletion screen [25] or development of large-scale targeted mutagenesis. Switching to more genetically tractable apicomplexan organisms, such as P. berghei or Toxoplasma, would provide ready options for large-scale targeted gene disruptions [23, 24]. However, these would have to be performed under conditional regulation because chemical rescue of the apicoplast is not feasible in these organisms. We anticipate that continued advances in genetic methods in apicomplexan organisms will open up opportunities to expand this screen in the future.
Human erythrocytes were purchased from the Stanford Blood Center (Stanford, California) to support in vitro P. falciparum cultures. Because erythrocytes were collected from anonymized donors with no access to identifying information, IRB approval was not required. All consent to participate in research was collected by the Stanford Blood Center.
P. falciparum Dd2attB (MRA-843) were obtained from MR4. NF54Cas9+T7 Polymerase parasites were kindly provided by Jacquin Niles. Parasites were grown in human erythrocytes at 2% hematocrit (Stanford Blood Center) in RPMI 1640 media (Gibco) and supplemented with 0.25% Albumax II (Gibco), 2 g/L sodium bicarbonate, 0.1 mM hypoxanthine (Sigma), 25 mM HEPES (pH 7.4; Sigma), and 50 μg/L gentamicin (Gold Biotechnology) at 37°C, 5% O2, and 5% CO2.
For transfections, 50 μg of plasmid DNA was added to 200 μL packed red blood cells (RBCs), adjusted to 50% hematocrit in RPMI 1640, and electroporated as previously described [66]. On day 4 post transfection, parasites were selected for with 2.5. mg/L blasticidin S (RPI Research Products International). TetR/DOZI strains were cultured with 500 nM ATc for the entire duration of transfection. For TetR/DOZI strains expressing ACPL-GFP or GFP-PfAtg8, parasites were additionally selected for with 500 μg/mL G418 sulfate (Corning) during transfection.
Oligonucleotides were purchased from the Stanford Protein and Nucleic Acid facility or IDT. gBlocks were ordered from IDT. Molecular cloning was performed using In-Fusion Cloning (Clontech) or Gibson Assembly (NEB). Primer and gBlock sequences for all cloning are available in S3 Table.
To generate the plasmid pRL2-mCherry-T2A-ACPL-GFP, T2A-ACPL-GFP was first amplified from the pRL2-ACPL-GFP vector. mCherry was amplified from pTKO2-mCherry vector (kind gift from J. Boothroyd) and inserted in front of T2A-ACPL-GFP in the pRL2 backbone using the In-Fusion Cloning kit (Takara). To generate the pL2-mCherry-T2A-ACPL-GFP-degron plasmids, T2A-ACPL-GFP-degron was amplified from pRL2-mCherry-T2A-ACPL-GFP.
For CRISPR-Cas9–based editing of endogenous Pf3D7_0518100, Pf3D7_1126100, Pf3D7_1305100, and Pf3D7_1363700 loci, sgRNAs were designed using the eukaryotic CRISPR guide RNA/DNA design tool (http://chopchop.cbu.uib.no/). To generate a linear plasmid for CRISPR-Cas9–based editing, left and right homology regions were first amplified for each gene. For each gene, a gBlock containing the recoded coding sequence C-terminal of the CRISPR cut site and a triple HA tag was synthesized with appropriate overhangs for Gibson Assembly. This fragment along with the left homology region was simultaneously cloned into the FseI/ApaI sites of the linear plasmid pSN054-V5. Next, the appropriate right homology region and a gBlock containing the sgRNA expression cassette were simultaneously cloned into the AscI/I-SceI sites of the resultant vector to generate the plasmids.
To generate plasmid for expression of GFP-PfAtg8, GFP with a GlyAlaGlyAla linker was amplified from pRL2-ACPL-GFP. PfAtg8 was amplified from P. falciparum gDNA. Both fragments were inserted into pfYC110 vector [38] using the In-Fusion Cloning kit.
V. brassicaformis RNA from strain CCMP3346 was purchased from the National Center for Marine Algae and Microbiota and was subsequently used to generate cDNA using Superscript III cDNA Kit (Life Technologies). For Plasmodium PF3D7_1363700 cloning, codon optimized gBlocks were used to construct the Plasmodium construct. Constructs were cloned into the pGEXT vector using the In-Fusion Cloning kit.
Ring-stage mCherry-T2A-ACPL-GFP-degron parasites were treated with 10 μM actinonin (Sigma) and 200 μM IPP (Isoprenoids, LLC) to disrupt the apicoplast. Both treated and nontreated parasites were analyzed two cycles post treatment at the schizont stage on a BD Accuri C6 flow cytometer. For each condition, 100,000 to 500,000 events were recorded. Uninfected RBCs were first removed from the population by setting a gate for mCherry fluorescence. For each strain, the average GFP and mCherry fluorescence intensities were then calculated for the infected cell population using FlowJo. For example, if 10,000 infected cells were counted, then the GFP and mCherry fluorescence for each cell was measured by the flow cytometer, and the average fluorescence was determined for the whole population. To calculate the GFP:mCherry ratios for comparative analysis of degron efficiency, the GFP to mCherry fluorescence ratio for each individual infected cell was first obtained. The fluorescence ratios of all infected cells were then averaged to determine the overall population GFP:mCherry ratio.
Ring-stage mCherry-T2A-ACPL-GFP-EcssrA (EcssrA) parasites were seeded onto a 96-well plate at a volume of 200 μL, 2% hematocrit, and 0.5%–1% parasitemia. Parasites were either untreated or treated with 1 mM EMS or 100 μM ENU for 2 hours, and then washed three times afterwards to remove the mutagen from the growth media. Parasites were cultured in growth media + 200 μM IPP for the duration of the screen.
At 120 hours post treatment, mutants were isolated on a Sony SH800S Cell Sorter. Uninfected RBCs were first analyzed to set the gate for overall fluorescence. Untreated EcssrA parasites were analyzed to gate for positive and negative mCherry and GFP expression, respectively. Actinonin-treated EcssrA parasites were analyzed to gate for positive GFP expression. Non-mutagenized and mutagenized parasites displaying both positive mCherry and GFP expression were FACS’d into a new 96-well plate. Enriched parasites were allowed to propagate to a detectable parasitemia before being subjected to subsequent FACS rounds.
Mutants were enriched until mCherry and GFP fluorescence approached actinonin-treated levels. In the final enrichment, up to 52 mutants were single-cell cloned. Mutants that survived single-cell sorting were split into growth media containing either 200 μM IPP or no IPP. Mutants displaying growth only in IPP were expanded to a 10 mL culture at approximately 10% parasitemia and ring-stage synchronized; 5 mL of culture was harvested for DNA extraction, and 5 mL culture was frozen at −80°C.
Ring-stage parasites were isolated from RBC in 0.1% saponin and washed three times in PBS. gDNA from mutants and the parental strain was isolated using the Quick-DNA Universal Kit (Zymo Research). Paired-end gDNA libraries were generated and barcoded for each mutant and the parental using the Nextera Library Prep Kit, modified for 8 PCR cycles (Stanford Functional Genomics Facility). Up to 26 pooled libraries were analyzed per lane of an Illumina HiSeq 4000 using 2 × 75 bp, paired-end sequencing (Stanford Functional Genomics Facility).
Fastq files were checked for overall quality using FastQC. Ten and 15 bp were trimmed from the 5′ and 3′ ends of all 75 bp sequence reads, respectively, to remove low-quality reads; 20 and 30 bp were trimmed 5′ and 3′ ends of 150 bp sequence reads, respectively, from an additional parental Dd2 strain sequenced by the DeRisi lab (https://www.ncbi.nlm.nih.gov/sra/SRX326518). The resulting paired-end sequencing reads were mapped using Bowtie2 against the P. falciparum 3D7 (version 35) reference genome. One mismatch per read was allowed, and only unique reads were aligned (reads were removed if they aligned to more than one region of the genome). PCR duplicates were removed using Samtools rmdup, and raw SNVs were called using Samtools mpileup. Indels were not analyzed.
Bcftools was used to generate the raw variant list for parental (allele frequency ≥ 0.05, depth ≥ 1) and mutant (allele frequency ≥ 0.9, depth ≥ 5) strains. Variants found in the parental strain were excluded from the mutant variant list. Variants were filtered to only include protein-coding mutations (missense and nonsense). Mutations found in hypervariable gene families were excluded. Remaining variants that were previously annotated in PlasmoDB were excluded to generate the final variant list. The reported variants were confirmed to meet the filtering requirements using Samtools tview. Mutants containing nonsense mutations were Sanger sequenced to confirm the presence of the mutations prior to genetic validation. The custom script and parameters used for analysis are available at https://github.com/yehlabstanford/biogenesis_screen.
Parasites were separated from RBCs by lysis in 0.1% saponin and were washed in PBS. Parasite pellets were resuspended in PBS containing 2× NuPAGE LDS sample buffer and boiled at 95°C for 10 minutes before separation on NuPAGE gels. Gels were transferred onto nitrocellulose membranes using the Trans-Blot Turbo Transfer System (Bio-Rad). Membranes were blocked in 0.1% Hammarsten casein (Affymetrix) in 0.2× PBS with 0.01% sodium azide. Antibody incubations were performed in a 1:1 mixture of blocking buffer and Tris-buffered saline with Tween (TBST)-20 (10 mM Tris [pH 8.0], 150 mM NaCl, 0.25 mM EDTA, 0.05% Tween 20). Blots were incubated with primary antibody overnight at 4°C at the following dilutions: 1:20,000 mouse-α-GFP JL-8 (Clontech 632381), 1:20,000 rabbit-α-Plasmodium aldolase (Abcam ab207494), 1:1,000 mouse-α-HA 2–2.2.14 (Thermo Fisher 26183), 1:1,000 guinea pig-α-ATG8 (Josman LLC), and 1:1,000 rabbit-α-ClpP (kind gift from W. Houry). Blots were washed three times in TBST and were incubated for 1 hour at room temperature in a 1:10,000 dilution of the appropriate fluorescent secondary antibodies (LI-COR Biosciences). Blots were washed twice in TBST and once in PBS before imaging on a LI-COR Odyssey imager.
Parasites expressing GFP-PfAtg8 were grown in the presence or absence of ATc for 24 hours; 10 ml cultures were lysed with 0.1% saponin and washed 3 times with PBS. Parasite pellets were resuspended in ice-cold lysis buffer (1× PBS, 1% Triton X-114 [Thermo Scientific 28332], 2 mM EDTA, 1× protease inhibitors [Pierce A32955]) and incubated on ice for 30 minutes. Cell debris were removed by 10-minute centrifugation at 16,000 × g, 4°C. Supernatant was transferred to a fresh Eppendorf tube, incubated 2 minutes at 37°C to allow phase separation, and centrifuged 5 minutes at 16,000 × g at room temperature. The top (aqueous) layer was transferred to another tube. The interphase was removed to avoid cross-contamination between the layers. The bottom (detergent) layer was resuspended in 1× PBS, 0.2 mM EDTA to equalize the volumes of the two fractions. Both fractions were subjected to methanol-chloroform precipitation, resuspended in PBS containing 2× NuPAGE LDS sample buffer, boiled for 5 minutes at 95°C, and analyzed by western blot as described above.
For live imaging, parasites were settled onto glass-bottom microwell dishes Lab-Tek II chambered coverglass (Thermo Fisher 155409) in PBS containing 0.4% glucose and 2 μg/mL Hoechst 33342 stain (Thermo Fisher H3570). Cells were imaged with a 100×, 1.4 NA objective on an Olympus IX70 microscope with a DeltaVision system (Applied Precision) controlled with SoftWorx version 4.1.0 and equipped with a CoolSnap-HQ CCD camera (Photometrics). Images were captured as a series of z-stacks separated by 0.2 μm intervals, deconvolved (except for mCherry images), and displayed as maximum intensity projections. Brightness and contrast were adjusted equally in SoftWorx or Fiji (ImageJ) for display purposes.
For immunofluorescence, fixed-cell imaging, parasites were first fixed with 4% paraformaldehyde (Electron Microscopy Science 15710) and 0.0075% glutaraldehyde in PBS (Electron Microscopy Sciences 16019) for 20 minutes. Cells were washed once in PBS and allowed to settle onto poly-L-lysine-coated coverslips (Corning) for 60 minutes. Coverslips were then washed once with PBS, permeabilized in 0.1% Triton X-100/PBS for 10 minutes, and washed twice more in PBS. Cells were treated with 0.1 mg/mL NaBH4/PBS for 10 minutes, washed once in PBS, and blocked in 5% BSA/PBS. Primary antibodies were diluted in 5% BSA/PBS at the following concentrations: 1:500 rabbit-α-PfACP (kind gift from S. Prigge) and 1:100 rat-α-HA 3F10 (Sigma 11867423001). Coverslips were washed three times in PBS, incubated with secondary antibodies goat-α-rat 488 (Thermo Fisher A-11006) and donkey-α-rabbit (Thermo Fisher A10042) at 1:3,000 dilution, and washed three times in PBS prior to mounting in ProLong Gold antifade reagent with DAPI (Thermo Fisher).
Ring-stage TetR/DOZI strain parasites were washed two times in growth media to remove ATc. Parasites were divided into three cultures supplemented with 500 nM ATc, no ATc, or no ATc + 200 μM IPP. Samples were collected at the schizont stage in each growth cycle for flow cytometry analysis and western blot. Parasites in each condition were diluted equally every growth cycle for up to six growth cycles.
For parasitemia measurements, parasite-infected or uninfected RBCs were incubated with the live-cell DNA stain dihydroethidium (Thermo Fisher D23107) for 30 minutes at a dilution of 1:300 (5 mM stock solution). Parasites were analyzed on a BD Accuri C6 flow cytometer, and up to 100,000 events were recorded.
The parent His6-SUMO-PfFtsH191-612-GST plasmid as well as E249Q and D493A mutants were obtained from laboratory stocks. An I437S mutant was constructed by site-directed mutagenesis. Recombinant proteins were expressed from these plasmids and purified as described [22].
Rates of ATP hydrolysis by PfFtsH1 were measured using a coupled spectrophotometric assay [67] in protein degradation (PD) buffer (25 mM HEPES [pH 7.5], 200 mM NaCl, 5 mM MgSO4, 10 μM ZnSO4, 10% glycerol) with 3% dimethyl sulfoxide (DMSO) or 50 μM actinonin in 3% DMSO at 37°C. PD rates were measured by incubating PfFtsH1 (1 μM) with FITC-labeled (2 μM, Sigma C0528) and unlabeled casein (8 μM) in PD buffer plus 3% DMSO or 50 μM actinonin in 3% DMSO. Reactions were started by adding ATP (4 mM) or buffer with a regeneration system (16 mM creatine phosphate and 75 μg/mL creatine kinase), and degradation was followed by measuring the fluorescence intensity (excitation 485 nm; emission 528 nm) at 37°C.
IGPS and IGPS-like proteins from V. brassicaformis were identified by BLAST through using CryptoDB. First, secondary structure was predicted using PSI-PRED in the XtalPred suite [53, 55]. Only the sequences containing the TIM barrels of each sequence were used for alignment because there are large N- and C-terminal extensions in the noncanonical proteins. PROMALS3D was subsequently used to perform a multiple sequence alignment based on secondary structure and homology to proteins with determined 3D structures [54].
W3110trpC9800 E. coli strain was purchased from the Yale University Coli Genetic Stock Center and were made chemically competent using calcium chloride. BL21 Star (DE3) competent cells (Thermo Fisher) were used for the WT condition. The competent W3110trpC9800 cells were transformed with the pGEXT vectors containing the different Vitrella or Plasmodium IGPS and IGPS-like genes and were plated on LB agar plates containing carbenicillin. For each construct, a colony was picked and washed in M9 minimal media (M9) (22 mM potassium phosphate monobasic, 22 mM sodium phosphate dibasic, 85 mM sodium chloride, 18.7 mM ammonium chloride, 2 mM magnesium sulfate, 0.1 mM calcium chloride, and 0.4% glycerol), resuspended in M9, and streaked onto M9/agar plates containing either carbenicillin (100 μg/mL) or carbenicillin and 1 mM L-tryptophan (Sigma). Plates were incubated at 37°C and allowed to grow for two days, after which images of the plates were taken.
|
10.1371/journal.ppat.1004500 | Silencing by H-NS Potentiated the Evolution of Salmonella | The bacterial H-NS protein silences expression from sequences with higher AT-content than the host genome and is believed to buffer the fitness consequences associated with foreign gene acquisition. Loss of H-NS results in severe growth defects in Salmonella, but the underlying reasons were unclear. An experimental evolution approach was employed to determine which secondary mutations could compensate for the loss of H-NS in Salmonella. Six independently derived S. Typhimurium hns mutant strains were serially passaged for 300 generations prior to whole genome sequencing. Growth rates of all lineages dramatically improved during the course of the experiment. Each of the hns mutant lineages acquired missense mutations in the gene encoding the H-NS paralog StpA encoding a poorly understood H-NS paralog, while 5 of the mutant lineages acquired deletions in the genes encoding the Salmonella Pathogenicity Island-1 (SPI-1) Type 3 secretion system critical to invoke inflammation. We further demonstrate that SPI-1 misregulation is a primary contributor to the decreased fitness in Salmonella hns mutants. Three of the lineages acquired additional loss of function mutations in the PhoPQ virulence regulatory system. Similarly passaged wild type Salmonella lineages did not acquire these mutations. The stpA missense mutations arose in the oligomerization domain and generated proteins that could compensate for the loss of H-NS to varying degrees. StpA variants most able to functionally substitute for H-NS displayed altered DNA binding and oligomerization properties that resembled those of H-NS. These findings indicate that H-NS was central to the evolution of the Salmonellae by buffering the negative fitness consequences caused by the secretion system that is the defining characteristic of the species.
| H-NS is an abundant DNA-binding protein found in enteric bacteria including the important pathogens Escherichia, Salmonella, Vibrio, and Yersinia, that plays a primary role in defending the bacterial genome by silencing AT-rich foreign genes. H-NS has been hypothesized to facilitate the evolution of bacterial species by acting as a buffer against the negative consequences that can occur when new genes are incorporated into pre-existing genetic landscapes. Here experimental evolution and whole-genome sequencing were employed to determine the factors underlying the severe growth defects displayed by Salmonella strains lacking H-NS. Through tracking the evolution of several independently derived mutant lineages, we find that compensatory mutations arise quickly and that they occur in loci related to virulence. A frequent outcome was loss of the Salmonella Pathogenicity Island-1, the defining genetic island of the genus Salmonella. Among other things these findings demonstrate that H-NS has enabled the birth of a new and important bacterial pathogen by buffering the fitness consequences caused by overexpression of SPI-1. These findings are likely generalizable to pathogens such as E. coli, Yersinia, Shigella, and Vibrio cholerae, all of which maintain a pool of “expensive” AT-rich virulence genes that are repressed by H-NS.
| Horizontal gene transfer (HGT) has profoundly shaped the course of bacterial speciation and diversification. The uptake of ‘pre-assembled’ genetic loci involved in antibiotic resistance, virulence, phage resistance or novel modes of metabolism can instantly confer beneficial phenotypes to the recipient cell. HGT events have been critical in the evolution of almost all bacterial pathogens from their non-pathogenic progenitors [1]–[5]. Two of the critical events when the Salmonellae diverged from their last common ancestor with E. coli were the acquisition of the Salmonella Pathogenicity Island-1 (SPI-1) and the tetrathionate reductase ttr gene clusters [6]–[8]. SPI-1 is a 40 kb genomic island encoding a Type 3 Secretion System (TTSS) required for triggering inflammation and for invasion of cells lining the intestinal mucosa [9]–[12]. Together these systems enable Salmonella to outcompete other microbes in the mammalian gut where SPI-1 induces a potent oxidative inflammation that generates tetrathionate, which then serves as a terminal electron acceptor for anaerobic respiration that is available solely to Salmonella but not other gut microbes [7].
Despite its overall importance to bacterial evolution, any individual HGT event is more likely to reduce bacterial fitness than to improve it. Even potentially beneficial genes can disrupt regulatory networks or drain metabolic resources away from the production of energy or biomass if they are not properly regulated [13]. Indeed, studies examining the barriers to new gene acquisition found that genes expressed at high levels are much more likely to be selected against in the new host [14], [15]. Virulence-associated genes, including those that encode secretion systems like the TTSS, can be particularly costly and are often lost in the absence of purifying selection (e.g. virulence attenuation by laboratory passage) [16]–[19]. For example, triggering TTSS activation from the Shigella virulence plasmid in liquid media causes the destabilization and eventual loss of the plasmid from the population [20].
The nucleoid associated protein H-NS was proposed to buffer the fitness costs associated with HGT by silencing genes with a %GC content significantly lower than the host genome average and are therefore likely to have been acquired from a foreign source [21]–[25]. H-NS confers this benefit both by counteracting transcription at standard promoters and by preventing spurious transcription within an adenine and thymine-rich (AT-rich) open reading frame at sequences that can adventitiously resemble a bacterial promoter [26]. H-NS exhibits low sequence specificity and targets DNA by recognizing specific structural features in the minor grove of AT-rich DNA [27], [28]. H-NS polymerizes along target AT-rich sequences by virtue of two independent dimerization domains, leading to the formation of extended nucleoprotein filaments [29]–[32]. As a result of its activity, H-NS regulates the majority of horizontally acquired sequences in species such as E. coli, Yersinia, Shigella and Salmonella [1], [33]–[35].
Members of the H-NS protein family are distributed between the alpha, beta and gamma proteobacteria. Functional analogues that bear minimal sequence or structural resemblance to H-NS have been identified in Pseudomonas sp. (MvaT and MvaU) and Mycobacteria sp. (Lsr2) [36], [37]. While global gene expression data sets from Escherchia coli (E. coli), Yersinia enterolitica (Y. enterolitica), Salmonella enterica Sv. Typhimurium (S. Typhimurium), Pseudomonas aeruginosa (P. aeruginosa) and Mycobacteria smegmatis (M. smegmatis) point to a common role for the H-NS/MvaT/Lsr2 proteins as silencers of foreign AT-rich sequences, the fitness consequences of mutating the xenogeneic silencers among these species differs significantly [21]–[24], [38]–[40]. In P. aeruginosa, MvaT and MvaU together are essential and depletion of both of these proteins results in the activation of the Pf4 prophage, which kills the bacterial cell [41]. In most strains of E. coli, mutations in hns mildly impede growth rates whereas failed attempts at constructing hns mutants in Y. enterolitica and Y. pseudotuberculosis strains strongly suggest hns is an essential gene in Yersinia sp. [42], [43]. S. Typhimurium strain 14028s hns mutants are only viable if additional mutations are present in either the PhoP-PhoQ two component signaling system or the stationary phase sigma factor RpoS [22]. What remains unclear is why global H-NS mediated gene silencing is so critical for the fitness of S. Typhimurium and Y. enterolitica, but is largely dispensable to other closely related species such as E. coli.
Several members of the Enterobacteriaceae including E. coli, S. Typhimurium and Shigella flexneri (S. flexneri) encode a second H-NS-like protein, StpA. StpA shares 53% sequence identity with H-NS as well as several functional properties, such as the ability to self-associate and bind AT-rich DNA [44]–[48]. H-NS and StpA also share a similar domain architecture exemplified by the detection of StpA/H-NS heterodimers in vivo and in vitro [49]–[52]. Global transcript analysis and ChIP-on-chip data sets indicate StpA and H-NS co-localize in E. coli and S. Typhimurium, but the loss of stpA only affects the transcript levels of a subset of these loci [47], [48]. In fact, loss of StpA alone does not generate observable phenotypes but will further impair the fitness of strains lacking H-NS [45], [53], [54]. The mild effects of stpA depletion may be attributed to low intracellular StpA concentrations [46], [55]. StpA is a substrate of the Lon protease and a StpA point mutation, F21C, that imparts resistance to proteolytic cleavage also restored stationary phase viability to an E. coli hns mutant strain [56]. Other reports, however, suggest H-NS and StpA exhibit similar expression levels with the StpA protein reaching 25 000 copies per cell at mid-exponential phase and H-NS reaching 20 000 copies [57]. Despite significant sequence homology between H-NS and StpA, the basis for their functional dissimilarities remains unknown.
In this study, we employed an experimental evolution strategy to select for mutations that compensate for the strong fitness defects of S. Typhimurium hns mutants. Using whole genome sequencing we identified parallel adaptations in many of the hns mutant lineages including genomic deletions in the pathogenicity locus SPI-1 and non-synonomous changes in the gene encoding StpA. The stpA mutations altered residues in the oligomerization domain and several enhanced the ability of StpA to silence hns regulated genes without having an effect on StpA expression levels. Much of the fitness defect in the hns mutants could be attributed to overexpression of SPI-1. This work provides compelling evidence that H-NS potentiates bacterial speciation by improving bacterial tolerance for horizontally acquired sequences. These findings also suggest that fitness-cost buffering by xenogeneic silencing proteins contributes to the observed tendency for genomic islands to be AT-rich.
Disruption of the hns gene in the wild type S. Typhimurium 14028s strain background severely restricts its growth rate to the point where cultivation is difficult [22]. However, we previously demonstrated that hns mutations can be achieved in strains that harbor additional mutations in the gene encoding the alternative sigma factor RpoS (σS or σ38). Alleles that reduce σS activity frequently arise during laboratory passage and are present in another commonly used Salmonella laboratory strain, LT2. The alleviating effect of rpoS mutations in the hns mutants may be due to the fact that loss of H-NS dramatically improves the stability of RpoS [58], which may cause the inappropriate overexpression of stationary-phase genes and interfere with the expression of housekeeping genes controlled by RpoD. To facilitate this study the hns gene from S. Typhimurium 14028s was replaced with a kanamycin resistance cassette in a background harboring a 5 amino acid in frame deletion within the coding region of rpoS that reduces RpoS activity (referred to as rpoS*) [22]. Although this additional mutation improved the tolerance of 14028s for hns mutations, Δhns/rpoS* strains continue to display severe growth defects including dramatically reduced colony size.
In the course of an earlier microarray study of a S. Typhimurium Δhns/rpoS* strain we noted one isolate appeared to lose a large cluster of genes at some point during laboratory passage [22]. To identify the nature of this deletion the isolate was further analyzed by whole genome sequencing where reads were assembled against the S. Typhimurium 14028s reference genome (Genbank ID CP001363.1) using Geneious Pro 5.5.6 software. This analysis revealed that the isolate incurred a 10 kb genomic deletion spanning nucleotides 1,334,560 to 1,344,664 (Figure 1A). The deleted region is highly AT-rich (GC% = 40% as compared to the genome average of %GC = 52) and encodes several putative envelope proteins including the PhoP activated genes pagC, pagD, pliC, envE, envF and msgA [59]. Multiple studies have shown that expression of pagC is strongly repressed by H-NS, and the spontaneous loss of these genes from the Δhns isolate suggested that hns mutants are genetically unstable and may shed horizontally acquired sequences during passage [21], [22], [60].
We sought to experimentally determine if the loss of horizontally acquired sequences is a reproducible outcome of deleting hns from S. Typhimurium, as well as to identify novel compensatory mutations that may alleviate the fitness defects associated with the loss of H-NS. Toward this end an in vitro evolution screen was performed where six independently derived freshly constructed Δhns/rpoS* mutant lineages were serially passaged alongside six lineages of the isogenic rpoS* background (the “wild type” strain) in Luria-Bertani broth for 30 days, or approximately 300 generations (Figure 1B). The lineages were designated WT or Δhns, “A” through to “F”. Each day during the experiment, aliquots from the cultures were stocked and stored at −80°C to enable the retrospective analysis of genomic changes in each lineage over time. At the end of the evolution period, the growth rates of the passaged wild type and the passaged Δhns lineages were monitored alongside their unpassaged (day 0) counterparts (Figure 1C). All six lineages lacking H-NS displayed significant increases in their growth rates compared to their respective day 0 clone, while the wild type lineages displayed modest improvements in growth (Figure 1C). Notably, by day 30 the Δhns lineages all exhibited growth rates similar to that of the wild type strains at day 30.
To identify mutations that arose throughout the evolution period, genomic DNA from the passaged WT and Δhns lineages and their progenitor lines was analyzed by Illumina whole genome sequencing. In total, the six Δhns lineages acquired 15 missense mutations, 2 small deletions, 2 small insertions and 5 chromosomal deletions larger than 10 kb (Table 1). Most striking was the high degree of similarity in these mutations. Five of six Δhns lineages incurred unique 10–50 kb deletions within the Salmonella Pathogenicity Island 1 (SPI-1) and all six Δhns lineages accumulated missense mutations within the stpA gene encoding the H-NS paralogue StpA (Figure 2, Table 1).
In agreement with our earlier observations [22], three Δhns lineages acquired mutations in the genes encoding the PhoP/PhoQ two component system that activates many H-NS repressed genes involved in virulence, acid stress, resistance to antimicrobial peptides and intramacrophage survival [59], [61]–[64]. Specifically, lineages A and E acquired frameshift mutations in PhoP and PhoQ respectively while Δhns lineage B acquired a missense mutation (Y320D) in the cytoplasmic sensor kinase domain of PhoQ.
Throughout the experiment each Δhns lineage acquired a total of three to four mutations with the exception of Δhns lineage D, which acquired eight. It is notable that Δhns lineage D incurred the largest chromosomal deletion that extended beyond SPI-1 into the locus encoding mutS and mutL, essential components of the methyl-directed mismatch repair pathway [65]. The loss of either mutS or mutL from E. coli has been shown to result in a mutator phenotype and may explain the accumulation of other missense mutations specific to the Δhns lineage D, namely idnK(E62G), mutY(D316N), yecS(P169S), yhfC(M255V) and stm1881(V321A) [66]. Analysis of the SPI-1 deletion junction regions revealed that 3 of the 5 deletions occurred without any homology in the sequences flanking the deleted segment. The other 2 SPI-1 deletions occurred between segments homologous in only 4 nucleotides. This suggests that RecA mediated recombination did not play a role in the loss of this island in the Δhns mutants (Figure S1).
Analysis of the wild type lineages revealed that comparatively fewer genetic changes arose during the course of the experiment. 3 of the 6 wild type lineages acquired large chromosomal deletions that extended from 10 kb to 58 kb downstream of the uvrC locus (Table 2). Common to all three deleted fragments were components of the uvrABC nucleotide excision pathway and constituents of the flagellar apparatus. Under the laboratory growth conditions used in this study, expression of the uvrABC and flagellar genes likely resulted in a disadvantageous use of cellular resources. Apart from these deletions no mutations common among the wild type lineages were observed.
To determine the timeline of the genetic changes that took place, genes of interest were PCR amplified from the frozen daily stocks of the hns mutant lineages and the PCR products were submitted for Sanger sequencing. This assay enabled the detection of mutant alleles soon after they arose in a given lineage and the relative proportion of the wild type and mutant alleles in the population at each day could be estimated from the sequencing chromatograms by analyzing the dual fluorescence peaks at a particular nucleotide. The relative signal strength of wild type vs. mutated nucleotides was used to approximate the emergence and dominance of each mutation in each population over time.
To determine when the large chromosomal SPI-1 deletions arose a PCR assay was employed; amplifying a region bridging the deleted segment. This detection method did not enable us to estimate the relative proportion of SPI-1 deletion strains in the population.
We found the mutations in the PhoP/PhoQ regulatory system and the SPI-1 deletions were acquired by the hns mutant lineages in the early stages of the passaging period, prior to the missense mutations in stpA (Figure 3). The PhoP/PhoQ and SPI-1 mutations were detected as early as day 2 of the evolution period in Δhns lineages A, B and D, suggesting these mutations confer the greatest growth advantages and/or are most easily acquired.
Of particular interest is the Δhns lineage C, which did not obtain inactivating mutations in either the PhoP/PhoQ or SPI-1 but displayed a comparable increase in fitness as Δhns lineages A, B, D, E and F in liquid growth assays (Figure 1C). Δhns lineage C acquired a stpA missense mutation (M4T) by day 5 that persisted at low frequency until it also acquired a second mutation in the housekeeping sigma factor RpoD (G471D), at which point the Δhns/stpA/rpoD mutant rapidly outcompeted both the Δhns and Δhns/stpA mutant strains in the population by day 13. To address the concern that lineage C acquired SPI-1 inactivating mutations that were not detected with the Geneious Pro software we performed a reference alignment of the raw Δhns lineage C paired end reads to the S. Typhimurium 14028s reference genome using the Bowtie software package and also preformed a de novo genomic assembly of the evolved Δhns lineage C with Velvet [67], [68]. A list of variants from both the Bowtie and Velvet assemblies was generated with Samtools and no other mutations besides for the StpAM4T and RpoDG471D variants were identified [69].
The fact that five out of six Δhns lineages rapidly and independently incurred deletions within the SPI-1 locus suggested that SPI-1 misregulation is a major contributor to fitness defects in S. Typhimurium Δhns mutants. SPI-1 expression is repressed by hns and activated by a complex positive feedback loop where the production of the HilD regulatory protein induces the expression of HilA, a transcription factor that directly activates expression of the TTSS and effector proteins [70]. To determine the degree to which SPI-1 impairs growth of the S. Typhimurium Δhns mutant, we deleted the 40 kb genomic island from a wild type strain prior to introducing the hns deletion by transduction. The SPI-1 deletion significantly improved the growth of the Δhns strain and also provided a mild improvement in growth of the wild type strain (Figure 4A). The region of SPI-1 lost in all Δhns lineages included the promoter upstream of hilD. Introduction of a hilD mutation into the Δhns background conferred a growth benefit similar to that of the 40 kb SPI-1 deletion (Figure 4B). These results indicate that in the absence of H-NS, SPI-1 is activated through a hilD dependent pathway and that the uncontrolled expression of SPI-1 encoded virulence determinants significantly impairs Salmonella growth.
Salmonella enterica harbors a second pathogenicity island, SPI-2, that encodes a type-3 secretion system distinct from the one encoded on SPI-1. Lucchini et al., previously reported that construction of a Salmonella ΔssrA/Δhns double mutant unable to express the genes encoded in SPI-2 significantly increased the growth rate of the Δhns strain (grown in LB media and using strain LT2) [21]. To determine if inactivation of SPI-2 encoded TTSS would offer the same fitness benefit as deletion of SPI-1 from a Δhns background, we introduced a 25 kb SPI-2 genomic deletion into Δhns and Δhns/ΔSPI-1 strains (Figure S2). Inactivation of SPI-2 did not significantly improve growth of either the Δhns or Δhns/ΔSPI-1 14028s strain to the same extent as loss of SPI-1. A similar experiment was conducted in LPM (low pH, low Mg2+ and low phosphate) media known to activate SPI-2 to determine if fitness of the Δhns mutant would be adversely affected in a manner dependent on SPI-2. The Δhns mutant failed to grow in this media but this growth defect was not alleviated in the ΔssrA/Δhns double mutant indicating that other factors, not SPI-2, impact fitness in our strain under these particular conditions.
The only gene that acquired mutations in all six passaged Δhns lineages encodes the H-NS paralogue StpA. All of the acquired stpA mutations resulted in single amino acid substitutions or in-frame insertions that map to the predicted N-terminal and central dimerization domains of the protein. Because disruption of stpA in a Δhns background is known to exacerbate hns mutant phenotypes, we found it unlikely that these substitutions impaired stpA function. Intriguingly, the StpA mutations arise exclusively at sites where the unchanged amino acid is not conserved with H-NS, and the residue changes appear to render StpA more “H-NS-like” (Figure 5A). We hypothesized that the stpA mutations impart H-NS-like silencing properties to StpA and therefore partially compensated for the loss of hns at loci outside of SPI-1 in the serial passaging experiment.
To test the ability of the StpA variants to complement hns mutant phenotypes, we cloned the stpA locus from each passaged Δhns lineage and wild type stpA into a low copy vector with the native stpA promoter. The resulting plasmids were pStpAWT, pStpAT37I cloned from Δhns lineage A, pStpAT37I/E42ins from Δhns lineages B and D which both acquired the T37I substitution and an E42 insertion, pStpAM4T from Δhns lineage C, pStpAA77D from Δhns lineage E and pStpAK38Q/F76L from Δhns lineage F. The StpA plasmids were transformed into a Δhns/ΔstpA S. Typhiumurim background in order to determine whether or not the isolated StpA variants could ameliorate bacterial fitness in the absence of hns. Introducing either pStpAWT or StpAT37I did not significantly improve growth of the Δhns/ΔstpA mutant (Figure 5B). On the other hand expression of the StpAM4T mutant significantly improved bacterial fitness in the liquid growth assay. Likewise, the StpAT37I/E42ins variant also offered an observable growth advantage. The strains expressing StpAA77D and StpAK38Q/F76L initially displayed a slight growth advantage and then plateaued at a similar final optical density as the StpAWT expressing strain.
Given that expression of the StpA variants identified in the serial passaging experiment enhanced bacterial fitness to varying degrees, we next tested the ability of the modified StpA proteins to complement the impaired motility phenotype of hns mutants. H-NS is required for both the expression and assembly of a functional flagellum [71]–[73]. H-NS indirectly stimulates flagellar gene expression by repressing hdfR, a known repressor of the flhDC regulatory locus and, in addition, H-NS directly binds to the flagellar protein FliG and helps organize rotor subunit assembly [22], [74]. StpA has also been shown to bind FliG, but does not promote motility in the absence of H-NS unless cellular StpA levels are artificially elevated [74]. To determine if the StpA variants stimulate motility to a greater extent than wild type StpA, we employed the same strains used in the liquid growth assays and measured their radial swarming diameters on soft agar motility plates. After a 16 hr incubation period, wild type S. Typhimurium displayed a swarming diameter of 62 mm (Figure 5C). Similar to the hns mutant strain, the S. Typhimurium Δhns/ΔstpA strains harboring pStpAWT and pStpAT37I did not migrate beyond the original inoculation zone. Remarkably, the StpA variants StpAM4T, StpAA77D and StpAK38Q/F76L restored motility to the Δhns/ΔstpA strain by 30%, 44% and 34% that of the wild type strain respectively (Figure 5C). StpAT37I/E42ins provided a small yet significant increase in swarming diameter to 16% the wild type diameter.
One possibility by which the StpA variants could restore motility to the Δhns mutant would be if the single amino acid substitutions increase StpA protein stability. Intracellular StpA pools are reportedly subject to proteolysis by the Lon protease in strains lacking hns [56]. In this study a mutation in the N-terminal dimerization domain of StpA, F21C, was shown to impart resistance to proteolysis and increase intracellular StpA concentrations. To determine if any of the StpA mutations identified in our laboratory passage screen influenced protein levels, the amount of intracellular StpA was quantified by western blot analysis. Δhns strains harboring epitope tagged StpA or its variants was probed with an α-FLAG antibody. DnaK levels were analyzed on the same blot as a loading control. Similar to the StpAF21C variant, StpAT37I and StpAT37I/E42 accumulated to higher intracellular levels than StpAWT (Figure 6). In contrast the variants StpAM4T, StpAA77D and StpAK38Q/F76L were detected at similar levels to that of StpAWT. This suggests that the StpA variants identified in this study fall into one of two categories, mutations that increase intracellular StpA levels similar to the previously identified StpAF21C variant, and a novel class of mutations that do not significantly alter intracellular StpA levels. Notably, it was the latter class of variants that provided partial complementation for the loss of hns in the growth and motility assays suggesting that the amino acid substitutions M4T, A77D and K38Q/F76L alter the functional properties of StpA and not its stability.
Much like H-NS, StpA has also been implicated in silencing AT-rich regions of the genome. Although the set of genes under control of StpA shares significant overlap with the set of genes regulated by H-NS, in the absence of H-NS, the silencing activity of StpA alone does not provide sufficient repression of H-NS regulated loci [46], [48], [75]. To determine if the missense mutations acquired throughout the evolution of the Δhns lineages enhanced StpA's silencing activity, we measured the steady state transcript levels of four model H-NS and StpA regulated loci from a Δhns/ΔstpA strain harboring pStpAWT, pStpAM4T, pStpAA77D and pStpAF21C. The StpAM4T and StpAA77D variants were chosen for transcript analysis because they provided the greatest restoration of the Δhns growth and motility defects without altering protein stability, while the StpAF21C variant was included to determine the regulatory consequences of increased intracellular StpA levels. Also included in the analysis were a Δhns complemented strain (Δhns+pHNS) and a Δhns strain, which served as reference points for repressed and derepressed transcript levels. cDNA from mid-log cultures was analyzed by Q-PCR with primers specific to proV, hilA, ssrA and yciG. proV is a well studied H-NS regulated gene target that resides outside the Salmonella pathogenicity islands, while hilA and ssrA are transcriptional activators encoded within SPI-1 and SPI-2 respectively. yciG is part of the rpoS regulon and was previously shown to be highly induced in a Salmonella SL1344 strain lacking stpA [48].
Relative to the Δhns complemented strain, the transcript levels of proV, hilA, ssrA and yciG increased by 20-fold or greater in the Δhns strain (Figure 7). The expression of yciG is highly repressed in the Δhns+pHNS strain, its transcript levels were lower than the detection limit of the Q-PCR cycler and could not be reported with confidence. The Δhns/ΔstpA strain harboring pStpAWT displayed a greater increase in the transcripts levels of proV, ssrA and yciG compared to the Δhns strain, while hilA transcript levels were reduced by 4.5-fold in the presence of pStpAWT. Substituting StpAWT with StpAM4T significantly reduced the expression levels of proV and ssrA by approximately 2-fold and 10-fold respectively. The StpAA77D variant provided even greater repression of proV and ssrA by reducing their transcript levels by 4-fold and 20-fold relative to StpAWT. Similar to the Δhns+pHNS strain, both StpAM4T and StpAA77D maintained yciG expression levels close to the detection limit of the sensor. In contrast, the StpAF21C variant that accumulates to higher intracellular levels than StpAWT did not maintain significantly lower expression levels of any of the four genes tested relative to the pStpAWT strain. This further establishes that the StpAM4T and StpAA77D variants as a novel set of mutations that enhance StpA silencing activity without affecting protein stability.
While the two single point mutations, M4T and A77D, significantly enhanced StpA's silencing activity at the proV, ssrA and yciG promoters regions these substitutions did not provide increased repression of hilA, encoding the SPI-1 transcriptional activator HilA. hilA expression is induced by three transcriptional activators, HilC, HilD and RtsA [70]. In the absence of H-NS it is possible that silencing complexes generated by StpAM4T and StpAA77D, although more effective than StpAWT, were unable to impede the combined HilC and HilD-mediated activation of hilA.
We repeated our in vitro evolution on an expanded number of freshly constructed hns deletion mutants to determine if loss of hns invariably led to mutations in stpA and, if so, to use this technique as a novel method of mapping functional residues in stpA. Toward this end hns deletion mutations were introduced by transduction into the rpoS-low strain to generate 12 independent lineages. To assess the impact SPI-1 may have on the evolution of stpA another 12 linages were generated by introducing the hns mutation into a strain already lacking SPI-1. Each of the 24 lineages were serially passaged in LB media over the course of 21 days and the stpA genes of each lineage were amplified by PCR and sequenced.
Sequencing of the stpA genes revealed missense mutations in 10/12 of the hns mutants in the rpoS background and 12/12 of the rpoS*/SPI-1 mutant background (Table 3). Remarkably the two hns mutant strains that did not acquire misssense mutations in stpA did acquire silent mutations, suggesting that either that stpA is prone to mutation in the absence of hns or that the presumably silent mutations actually affect StpA levels or function by increasing mRNA stability or by altering codon usage. As before all missense mutations mapped to the oligomerization domain between residues 2 and 80 of the stpA protein. Furthermore some lineages acquired as many as 4 different nucleotide substitutions. The fact that 30 independent lineages (24 in this experiment and 6 in the initial experiment) acquired mutations in stpA and that none of these were nonsense mutations confirms that there is strong selective pressure to acquire mutations in stpA in the absence of H-NS.
Notably there were some differences observed in the specific mutations acquired between the two lineages (those with or without SPI-1). In the presence of SPI-1 the StpA protein was altered at several different residues but a cluster of mutations occurred at or near codon 38 (nucleotides 112–114) encoding lysine including a silent mutation at nucleotide 111. Strains that evolved in the absence of SPI-1 acquired a notably different set of mutations where all but one lineage acquired a mutation at nucleotide 110 resulting in the StpA(T37I) variant. Additional mutations changed the asparagine at positions 2 or 7 to an aspartic acid (N2D or N7D). This suggests that the pressures that select for mutations in StpA may differ in the absence of SPI-1.
The results of the evolution experiment provided an opportunity to map what single or double residue changes in StpA would be sufficient to engender it with H-NS-like functionality. This functionality of each StpA variant was assessed by their ability to restore motility (Figure 8A) when expressed in the hns mutant background. This assay was chosen because our data with the earlier StpA variants indicated motility restoration correlates closely with their to silence H-NS regulated loci. These assays uncovered functional changes in single amino acids that cluster to two discrete regions of the StpA protein (Figure 8). The functional variants StpAN2D, StpAM4T, and StpAN7D map to the short helix 1 that lies within the N-terminal dimerization domain while the variants StpAF76V, StpAF76L, StpAA77D and StpAM78K all map to helix 4 which is contained in the central dimerization domain. Other single residue StpA variants, where changes mapped to helix 3 or the short linker segments that connect helix 3 to the other helices, failed to restore significant motility to the hns mutant. Modeling these changes on the previously published H-NS oligomer structure show that the individual changes that confer H-NS-like function to StpA are buried within the dimerization interfaces or present on the outer, convex, surface of the H-NS filament while the residues that do not lie predominantly on the concave surface of the filament, and are largely predicted to have surface exposed side chains (Figure 8B). It is important to note the StpA residues were mostly assessed individually (only two double-mutants were assessed) and that some residues that appear to have no gain of function in our assays may have a more dramatic impact in combination with other changes.
Electrophoretic mobility shift assays were used to determine if changes in the StpA variants that led to increased “H-NS-like” function manifest as differences in their ability to form nucleoprotein complexes on DNA. Like H-NS, StpAWT displays cooperative binding to a model 289 bp AT-rich sequence (%GC = 34) but forms nucleoprotein complexes are consistently observed to have significantly lower mobility than those formed by H-NS on the same DNA target (Figure 9). Remarkably the nucleoprotein complexes formed by the StpAM4T and StpAA77D variants formed complexes with motility more similar to H-NS than wild type StpA. StpAM4T formed two complexes on DNA, one that migrated with the top band of the DNA ladder like StpA and one that migrated further into the gel at the same position as the H-NS complex. StpAA77D almost exclusively formed a single H-NS like complex. StpAT37I, which had enhanced protein levels in vivo, but failed to complement for H-NS for either motility or silencing, formed a lower mobility nucleoprotein complex identical to that of the wild type StpA protein. Notably, there were no differences in overall affinity for DNA between the different variants.
This data indicates that subtle changes in the dimerization domains of StpA can generate large and quantifiable differences in properties of the nucleoprotein complex and that the functional differences observed between StpA variants manifest as differences in their effects on nucleoprotein structure. At high protein concentrations both StpA and H-NS have the ability to spontaneously oligomerize into higher order structures in the absence of DNA, a phenomenon that can be measured by changes by analytical gel filtration chromatography. We assessed the gel filtration profiles of StpA and its variants (Figure 10) to determine if any changes in their oligomerization states could be observed. StpAWT and the StpAT37I, which do not effectively substitute for H-NS, displayed two prominent peaks with calculated molecular weights of approximately 450 and 150 kDa (StpA monomer is ∼15 kDa). The chromatographic profiles of the StpAM4T and StpAA77D proteins indicate that these proteins have a dramatically reduced propensity to form the oligomeric species that elutes early during chromatography. We note that the asymmetrical rod-like structure of the StpA and H-NS oligomers prevent an accurate determination of molecular weight based on mobility through the column when compared to a set of globular standards. Differences in shape or flexibility would also manifest as different elution profiles by gel filtration. Nevertheless these findings when taken as a whole indicate that the functional differences between the StpA variants (and also the functional differences between H-NS and StpA) are primarily due to differences in manner of their oligomerization and not in the specificity of their DNA binding domains.
The xenogeneic silencing model predicts that the selective silencing of foreign DNA accelerates bacterial evolution by reducing the fitness cost associated with HGT. While multiple studies have established a role for the H-NS, MvaT and Lsr2 protein families in regulating newly acquired sequences, the evolutionary advantage of foreign gene repression by H-NS and its impact on genome content had not been assessed by experimental evolution [21], [22], [39], [40]. The genetic adaptations we identify that improve growth in strains lacking H-NS indicate that xenogeneic silencing played a major role in the evolution of the Salmonellae by buffering the fitness consequences caused by the SPI-1 encoded TTSS, a defining characteristic of the species. Indeed a recent study on the evolution of Salmonella revealed that, while many sequences acquired by HGT will adopt the %GC of their host over time, the major pathogenicity islands have selectively retained their AT-richness, presumably to maintain their silencing by H-NS [6]. The fact that we observed large deletions in SPI-1, rather than inactivating point mutations or small indels, is somewhat surprising and suggests that this region may be naturally unstable and prone to gene loss. The 5 deletions independently occurred at different sequences, each with limited no flanking homology, suggesting that replication errors and not homologous or site-specific recombination likely caused the loss of these regions from the genome.
Multiple lines of evidence suggest that maintaining a TTSS represents a costly investment of cellular resources. Induction of the Yersinia TTSS by low calcium essentially halts bacterial growth and the plasmid-encoded Shigella TTSS is readily lost during laboratory passage [20]. An association between impaired bacterial growth and SPI-1 expression in wild type cells was recently reported in a study conducted by Sturm et al [76]. This study tracked the spontaneous induction of the SPI-1 encoded TTSS at a single cell level using time-lapse microscopy imaging. Sturm et al. correlated the expression of the TTSS with retarded growth rates that were alleviated by mutations in the SPI-1 activator hilA. Under the conditions employed in this study the hns mutant strains only incurred genomic deletions within SPI-1. However, a targeted disruption of SPI-2 was previously shown to partly improve growth in an hns mutant background [21]. We believe the loss of SPI-1 and not SPI-2 from our hns mutant lineages likely reflects that the conditions we employed in this study favored SPI-1 expression. It is important to note that the conditions employed during the experimental evolution experiment were arbitrarily chosen and it is entirely likely that subtle changes in environment will significantly impact which loci will impinge on fitness in the absence of H-NS. While the wild type lineages passaged in parallel to our hns mutants did not acquire mutations in the SPI-1 locus, genomic deletions encompassing components of the flagellar apparatus were noted in 3 out of 6 of the wild type lineages. Flagella and TTSS are evolutionarily related and highly homologous in both primary sequence and structure [77]. The fact that the flagellar loci of the hns mutant lineages did not acquire mutations is consistent with the fact that hns mutants fail to express flagella to begin with.
Pathogens like Salmonella spend a substantial amount of time outside of the host environment and our studies suggest that H-NS is essential for enteric bacteria to retain virulence in the absence of selective pressures. Naturally occurring SPI-I deletions have occasionally been identified among environmental Salmonella isolates that have consequentially lost the ability to invade host cells [78]. Spontaneous SPI-1 mutations are thought to arise throughout host infection generating a subpopulation of “avirulent defectors” that propagate much faster than their TTSS-positive predecessors [79]. Diard et al demonstrated that Salmonella infection with a constitutively active SPI-1 TTSS strain resulted in a sharp rise of the genetically avirulent subpopulation and consequently premature clearing of the infection [79].
The only hns mutant lineage that did not incur a SPI-1 deletion, Δhns lineage C, acquired a missense mutation in rpoD (RpoDG471D). The recent crystal structure of the E. coli RNAP/σD holoenzyme shows RpoD residue G471 is located in an exposed loop region, enriched in highly conserved aromatic and positively charged residues [80]. An alignment of the E. coli RNAP/σD structure with the RNAP/σD initiation complex from Thermus thermophilus reveals the conserved loop region harboring residue G471 is in close proximity the template strand during transcriptional initiation [81]. It is possible that introduction of the negatively charged aspartic acid residue at position 471 could hinder transcriptional initiation and result in reduced expression of the SPI-1 locus, however it is currently unclear how this mutation would affect SPI-1 but not impede the expression of many other important σD targets.
Another central and important outcome of this study was the identification StpA oligomerization variants that partially compensate for several H-NS dependent phenotypes. Many of these mutations do not increase cellular StpA protein concentrations, as has been observed previously [46], [54]. Notably a recent study on a spontaneous mutant that improved fitness of an E. coli strain lacking Hha and YdgT, two molecules that collaborate with H-NS to facilitate gene silencing, identified a promoter mutation that dramatically enhanced H-NS levels [82]. StpA in S. Typhimurium was recently proposed to repress the rpoS regulon during exponential growth and the major caveat of our study is that we started with strain encoding a defective RpoS, which would alleviate the selective pressure to maintain a wild type copy of StpA [48]. H-NS also represses numerous genes activated by rpoS in response to cellular stress [83]–[85]. A study of the hdeAB promoter region suggested H-NS repression was overcome by the RNAP•σS complex, while RNAP associated with the house keeping sigma factor σD was more effectively inhibited by the presence of H-NS [85]. A similar finding was also reported for the dps promoter [86]. One model that could be extrapolated from these observations is that StpA restricts transcription of the RNAP complexed with σS while H-NS more efficiently represses RNAP bound to σD. Complicating this model is the fact that H-NS and StpA can heterodimerize and that each may individually regulate cellular σS concentrations [48]–[50], [87], [88].
The story that is emerging from this and other recent studies is that subtle changes in local nucleoid architecture, directed by the structure of the oligomerized protein, underlies the diverse functions ascribed to the H-NS like molecules. Several findings indicate that changes in DNA shape and tension are the relevant outputs of this class of transcriptional modulators; a mode of gene regulation that is particularly challenging to study using conventional assays like EMSA and footprinting. Our results indicate that StpA and H-NS differ primarily not in their ability to bind AT-rich DNA per se, in fact StpA binds DNA with an apparent affinity higher than that of H-NS, but by some physical property that manifests as a change in promoter architecture once bound by the protein. Due to its apparent higher affinity for DNA and elevated propensity to form higher-order oligomers in conventional assays one would predict that StpA would be a more effective silencer than H-NS in most situations. We note that there are significant qualitative differences in the shifted DNA complexes between the StpA variants that can complement for the loss of H-NS and those variants that cannot. This supposition is further supported by recent studies on the H-NS-like transcriptional activator Ler and H-NS paralogs encoded on plasmids demonstrating that the central linking domain, not the DNA binding domain, is the primary determinant in how these molecules functionally differ from H-NS [89], [90].
Evidence that H-NS, StpA, and the “H-NS-like” Ler proteins each form characteristically distinct higher order protein/DNA complexes has been more directly provided by recent atomic force microscopy imaging studies and single molecule “DNA stretching” experiments [91]–[94]. Lim et al reported that StpA-induced DNA/protein filaments were significantly more rigid than those produced by H-NS, and that the StpA filaments were insensitive to changes in pH, temperature, and osmolarity; conditions known to disrupt H-NS-DNA binding [92]. Another observation that might support divergent oligomerization properties of StpA and H-NS is that StpA can silence the E.coli bglG operon, but only in the presence of H-NS molecules deficient in DNA binding [50], [95]. This observation was used to suggest that the H-NS proteins can heterodimerize with StpA to facilitate silencing of bglG. However, based on our new findings, we cannot exclude the possibility that the hns mutant strain used in that study acquired mutation(s) in stpA during routine lab passaging that enabled it to act like H-NS.
The fact that compensatory stpA mutations arise rapidly and reliably in the absence of H-NS is a worrying outcome of this study. Complicating matters further is the apparent functional heterogeneity in the various stpA mutations we uncovered, i.e. the different compensatory mutations do not share the exact same properties. It is unclear how much care has been taken in the maintenance of the various hns mutant strains employed in many prior studies and in all but one case it is clear that the stpA locus was not sequenced to check for mutations. Regrettably this leaves some doubt regarding the validity of earlier studies on the phenotypes of strains lacking H-NS. Given their genetic instability, all future work on hns mutants in either E. coli or Salmonella should be performed on multiple freshly constructed (transduced) isolates and laboratory passaging of such strains should be kept to a minimum. Whenever possible the genomes of hns mutants should be re-sequenced to verify that phenotypes ascribed to H-NS are not, in fact, due to a mutation in a different gene.
The plasmids and strains employed in this study are listed in Table 4 and a complete list of oligonucleotides sequences is provided in Table 5. In a previous study, a FLAG-epitope tag was incorporated into the XhoI and BamHI sites of the low copy vector pHSG576 to generate pWN425 [22]. The stpA coding sequence and 206 bp upstream region (comprising nucleotides 2976460 to 2968067 in the S. Typhimurium 14028s genome Genbank ID CP001363.1) was PCR-amplified from S. Typhimurium 14028s genomic DNA with primers ALO115 and ALO116. The amplified fragment was ligated into the PstI and BamHI sites of pHSG576 backbone for expression of StpA harboring a C-terminal FLAG epitope tag. The StpA coding sequence and promoter region were incorporated into pHSG576 in the opposite orientation of the lac promoter, such that stpA expression levels were controlled by the native stpA promoter. The resulting plasmid pStpAWT was used for complementation studies. Similarly, plasmids harboring the StpA variants identified in the experimental evolution screen were constructed using the pHSG576 backbone with a C-terminal FLAG epitope tag. The mutated StpA alleles were PCR amplified from the genomic DNA of their respective hns mutant lineages that had been passaged for 30 days. The same primer pair used to amplify the wild type stpA coding sequence and 5′ promoter region was employed. The mutant stpA allele PCR fragments were inserted into the PstI and BamHI sites of vector pHSG576 harboring the FLAG epitope sequence 3′ of the BamHI site. The plasmids generated and the corresponding hns mutant lineage that the stpA alleles were cloned from were as follows: pStpAT37I from Δhns lineage A, pStpAT37I/E42insert from Δhns lineages B and D, pStpAM4T from Δhns lineage C, pStpAA77D from Δhns lineage E and pStpAK38Q/F76L from Δhns lineage F. The sequences of all the plasmids constructed in this study were confirmed by Sanger Sequencing at the TCAG Sequencing Facility (Centre for Applied Genomics, Hospital for Sick Children).
The Salmonella enterica serovar Typhimurium 14028s strains used in this study possess a mutant rpoS allele (called rpoS*) that encodes a five residue in-frame deletion that significantly reduces RpoS (σ32) activity [22]. The single stpA and hns chromosomal deletion strains were previously constructed using the lambda red recombinase method described by Datsenko and Wanner [96]. The stpA gene from S. Typhimurium 14028s was replaced with a kanamycin resistance cassette amplified from plasmid pKD4, flanked by FRT recombinase sites. The kanamycin resistance cassette was subsequently flipped out of the chromosome by introducing the pCP20 plasmid expressing the FLP recombinase. This generated a S. Typhimurium ΔstpA strain without antibiotic resistance markers. To test the ability of the StpA variants to compensate for the loss of H-NS, each of the StpA complementation plasmids were transformed into the S. Typhimurium 14028s ΔstpA mutant. Next, the hns null allele was moved into the ΔstpA mutant strains harboring the StpA complementation plasmids by P22 transduction. The resulting clones were selected for on Miller's Luria Bertani (LB) 1% agar plates supplemented with 50 µg/ml kanamycin (to select for the hns null mutation) and 20 µg/ml chloramphenicol (to select to the StpA plasmids).
SPI-1 deletion mutants were constructed by deleting a 44.4 kb region spanning the SPI-1 region using the lambda red recombinase method described by Datsenko and Wanner [96]. In brief, the region between S. Typhimurium 14028s genome coordinates 3005740–3050161 (Genbank ID CP001363.1) in each parent strain was replaced by a chloramphenicol resistance cassette flanked by FRT recombinase sites from plasmid pKD3 using primers ALO76 and ALO77. Each knockout mutation was then transduced into a fresh strain background by P22 HT105/1 int-201 transduction. Similarly, the SPI-2 deletion mutants were generated using the lambda red recombinase method. A chloramphenicol resistance cassette was amplified from plasmid pKD3 with primers ALO83 and ALO84, which were designed with flanking sequences complementary to the SPI-2 region. Following lambda red recombinase with the amplified PCR product, a 25 kb SPI-2 deletion spanning nucleotides 1,486,143–1,511,465 (Genbank ID CP001363.1) was introduced into the S. Typhimurium 14028s genome. The SPI-2 mutation was then transduced into a fresh strain background by P22 HT105/1 int-201 transduction. The double ΔSPI-1/ΔSPI-2 mutants strain were generated by first flipping the ΔSPI-1 chloramphenicol resistance cassette out of the chromosome by introducing the plasmid PCP20 expressing the FLP recombinase, and then introducing the SPI-2 deletion via P22 transduction.
Strains containing the hilD mutants were constructed by P22 transduction of a previously constructed mutation provided generously by the lab of Dr. Ferric Fang at the University of Washington [97].
An hns gene knockout from S. Typhimurium 14028s harboring a kanamycin resistance cassette in place of hns was moved into a fresh S. Typhimurium 14028s background containing a mutated rpoS allele (rpoS*) via P22 phage transduction. The transductants were selected on LB-agar plates supplemented with 50 µg/ml kanamycin. Six independently derived colonies from the original transduction were streaked twice on LB-kanamycin plates to eliminate trace P22 phage lysate, with each passage on solid media corresponding to a 16 hour incubation period at 37°C. All six transductants harbored the kanamycin resistance cassette in place of hns and were free of contaminating P22 phage as determined by PCR. These hns mutant isolates were selected to inoculate 5 ml LB in conical 25 ml polypropylene culture tubes. The cultures were grown at 37°C with 200 rpm shaking and every 24 hr, 5 µl from each culture was transferred to 5 ml of fresh media. The 1∶1000 dilution corresponds to approximately 9.96 doublings a day for a total of ∼300 doublings over the course of the 30 day evolution period. Daily samples from each lineage were taken and stored at −80°C in culture media supplemented with 10% DMSO for later analysis.
Samples from the frozen DMSO stocks representing day 1 and day 30 of the evolution period were scraped into LB media and grown at 37°C with shaking until mid-stationary phase (approximately 8 hours). The genomic DNA from approximately 4×109 cells from each culture was purified using the Qiagen DNeasy blood and tissue kit. 5 µg of the purified DNA in 130 µl water was sheared to a mean fragment size of 400 nt using a Covaris S2 focused ultrasonicator (Woburn, Massachusetts). The fragmented DNA was concentrated in a centrifugal evaporator to less than 34 µl and treated with the End-IT DNA repair kit from Epicenter to blunt-end the DNA. Following a 1 hr incubation period at room temperature, 50 µl H20 and 400 µl buffer QG from the Qiaquick Gel extraction kit were added to the blunted DNA fragments. The samples were purified with the QIAquick spin columns (Qiagen) and eluted twice in 15 µl elution buffer (total elution volume 30 µl). A-tails were added to the blunted fragments using the Klenow Exo-minus enzyme from Lucigen for 1 hr at room temperature and the reaction was terminated with the addition of 400 µl Quiagen QG buffer. After a second purification with the QIAquick spin columns, the eluted DNA (30 µl) was reduced to a volume of 9.25 µl in the centrifugal evaporator. Preannealed dsDNA adapter oligonucleotides were ligated to each sample overnight at 16°C using the Fast-Link DNA Ligation kit (Epicentre). These adapters were generated by mixing equimolar parts of a desalted common oligonucleotide (5′-AAT GAT ACG GCG ACC ACC GAG ATCTAC ACT CTTTCC CTA CAC GAC GCT CTT CCG ATC*T-3′), where C* indicates the addition of a phosphothioate group, and a unique indexing oligonucleotide with partial complementarity (5′Phos-GATCGGAAGAGCGGTTCAGCAGGAATGCCGAGACCGNNNNNNNNATCTCGTATGCCGTCTTCTGCTTG-3′), where N indicates a unique 8 nt barcode. The samples were separated on a 2% agarose gel and a slice containing fragments of approximately 400–450 nucleotides was extracted purified with the Qiagen gel extraction kit. The samples were then amplified in a PCR cycler for 16 cycles, purified once again with the Qiagen Gel Extraction Kit, and were quantified spectrophotometrically. Equal quantities of each library were combined and sequenced by the Donnelly Sequencing Centre (Toronto) in a partial lane of a 130 nt×8 nt index×100 nt paired-end run on an Illumina HiSeq2000 instrument using v3 chemistry. To achieve greater depth of coverage for the wild type E lineage at 30 days, this library was resequenced on a partial lane of a 101 nt×8 nt×101 nt HiSeq2500 run. The unique 8 nt barcode sequence present on the ligated adapters enabled the identification of each sample during downstream analysis.
Paired-end Illumina reads from each strain were reference assembled to the published S. Typhimurium 14028s genome (Genbank ID CP001363.1) using the Geneious Pro 5.5.6 software package on “medium-low sensitivity/fast”, which corresponds to the following settings: maximum gaps per read 10%, maximum gap size 15, minimum overlap identity 80%, minimum overlap 25 nt, and maximum mismatches per read 20%. The mean genomic depth of coverage ranged from 32.2%–134.2%. Single nucleotide polymorphisms and small deletions and insertions (SNPs/INDELS) that arose in each lineage were identified by comparing the genomes of each lineage at Day 30 to their corresponding genome sequence at Day 1 using the Geneious Pro 5.5.6 “Find Variants/SNPs” tool with the minimum depth of sequence coverage 25-fold and the variant frequency set to 0.8. In addition, the raw Illumina reads of the hns mutant lineage C genomic DNA from Day 30 were aligned to the published 14028s genome using Bowtie version 1.0.0 and de novo assembled using Velvet version 1.2.1.0 [67], [68]. A list of the SNPs/INDELS from the Bowtie and Velvet assemblies of hns lineage C was generated using Samtools [69].
SNPs were then confirmed by sequencing PCR products from each strain. For each SNP, the corresponding gene was amplified from the DMSO stock of each passage by PCR using gene-specific primers: stpA (ALO117/118), rpoD (ALO122/123), idnK (ALO139/140), mutY (ALO141/142), phoP (ALO145/146), phoQ (ALO143/144), yecS (ALO151/152), yhfC (153/154) (Table 4). The resultant PCR product was then purified using EZ-10 Spin Column PCR Purification Kit (Biobasic) and sent for Sanger Sequencing at TCAG Sequencing Facility (Centre for Applied Genomics, Hospital for Sick Children). The passage when each mutation occurred was similarly determined by sequencing individual loci from the samples stored daily during the course of the experiment. Sequencing chromatograms were visually compared for the emergence of the mutant nucleotide change. Emergence of the SPI-1 deletions were detected by PCR amplifying the region spanning the deletion sites from the DMSO stock of each passage using the following primer pairs: ALO127/ALO128 for Δhns lineage A, ALO129/ALO130 for Δhns lineage B, ALO131/ALO132 for Δhns lineage D, ALO133/ALO134 for Δhns lineage E and ALO135/ALO136 for Δhns lineage F. The PCR products were purified using EZ-10 Spin Column PCR Purification Kit (Biobasic) and sent for Sanger Sequencing at TCAG Sequencing Facility (Centre for Applied Genomics, Hospital for Sick Children).
Overnight cultures (5 ml LB) were inoculated from single colonies grown for approximately 16 hours at 37°C with 200 rpm shaking. Cultures for each strain were then adjusted to an O.D. at 600 nm of 0.5 and then diluted an additional 1∶100. 200 µl of each culture was then dispensed in triplicate into a clear, flat-bottom 96-well plate (Sarstedt), the plate was covered with the plate lid and grown overnight with shaking at 37°C in a TECAN Infinite M200 Pro microplate reader. Optical density readings at 600 nm were recorded every 15 minutes for 18 hours.
Overnight cultures (5 ml LB) were initiated from single colonies and grown for 16 hour at 37°C with shaking at 200 rpm. The next day cultures were each adjusted by dilution to an O.D. at 600 nm of 0.1. Equivalent colony forming units in each of the diluted cultures were verified by plating serial dilutions. 5 µl of the O.D.600nm 0.1 cultures was spotted into the center of 25 ml soft agar plates (LB 0.35% agar). The plates were incubated for 12 hours at 37°C and the radial swarming diameters were measured. The motility assays were replicated three times and in each assay the strains were plated in triplicate.
Overnight cultures were diluted 1∶200 in 200 ml LB media containing 20 µg/ml chloramphenicol. Sample volumes of 50 ml, 12 ml and 1.5 ml were removed from the cultures at O.D. 600 nm of 0.1, 0.6 and 1.5 respectively. Cells were harvested by centrifugation at 5000× g for 15 min at 4°C. The cell pellets were resuspended in cell lysis buffer containing 9.32 M urea, 2.67 M thiourea, 40 mM Tris, and 86.78 mM 3-(3- cholamidopropyl)-dimethylammonio-1-propanesulfonate (CHAPS; pH 8.5). Cells were lysed by sonication and the total protein concentrations were quantified using Bradford assay (Bio-Rad). 30 µg of total protein was combined with 2× SDS PAGE loading dye and separated on a 16% polyacrylamide SDS Tris-Tricine gel. Transfer to a nitrocellulose membrane was performed with the Bio-Rad semidry electrophoretic transfer cell at 15 V for 1 h. The membrane was blocked at 4°C over night in TBST 1× Tris-buffered saline, 0.05% Tween 20) with 5% skim milk powder. The membrane was probed with Rabbit anti-FLAG M2 antibody (Sigma) diluted 1∶1000 in TBST with 5% (w/v) skim milk for 1 h at room temperature, followed by goat anti-rabbit secondary antibody conjugated with horseradish peroxidase (Sigma, diluted 1∶10,000 in TBST with 5% milk) for 1 h at room temperature. DnaK was probed as a loading control using a mouse primary antibody (Enzo Life Sciences, 1∶1,000 in TBST with 5% milk) followed by a goat anti-mouse secondary antibody conjugated with horseradish peroxidase (Enzo Life Sciences, 1∶10,000 in TBST with 5% milk).
Constructs overexpressing StpAwt, StpAT37I, StpAM4T and StpAA77D were transformed in BL21Δhns (DE3) strain. The resulting strains were cultured in Luria Bertani (LB) until OD600 nm = 0.6. IPTG was added to a final concentration of 1 mM prior to growing the cultures for 16 h at 18°C. Cells were spun at 2500×g for 30 min, resuspended in 20 mL cell lysis buffer (20 mM Tris pH 8, 500 mM NaCl, 5 mM imidazole, 5 mM β-mercaptoethanol) and sonicated. The cellular debris was removed by centrifugation at 20 000 g for 15 min. Ni2+ resins were incubated with supernatant for 1 h on a rocking platform, washed twice with washing buffer (20 mM Tris pH 8, 500 mM NaCl, 30 mM imidazole, 5 mM β-mercaptoethanol), and eluted with elution buffer (20 mM Tris pH 8, 500 mM NaCl, 500 mM imidazole). Proteins were then purified further by size exclusion chromatography using Superdex 200 16/60 column from GE healthcare using storage buffer (20 mM Tris pH 8, 1 M NaCl, 1 mM EDTA and 5% glycerol). Fractions containing protein were concentrated using Millipore Amicon Ultra centrifugal filter 3K and stored at −80°C. H-NS6HIS protein was purified by nickel affinity chromatography as previously described [98]. Ni2+ purified H-NS6HIS was dialyzed in buffer A (20 mM Tris pH 8, 1 mM EDTA, 200 mM NaCl and 5% glycerol) overnight prior to being loaded onto a 5 mL Hitrap Heparin HP column, and eluted using a linear gradient of low salt buffer (20 mM Tris pH 8, 1 mM EDTA, 150 mM NaCl and 5% glycerol) with high salt buffer (20 mM Tris pH 8, 1 mM EDTA, 1 M NaCl and 5% glycerol) over 120 mL. Peak fractions were analyzed by SDS-PAGE and then dialyzed in loading buffer prior to storage at −80°C.
Total RNA was purified and reverse transcribed as previously described [98]. The resulting cDNA was analyzed by real-time quantitative PCR (Q-PCR) with primers specific to ssrA (SSA198/199), hilA (SSA200/201), proV (SSA202/203) and yciG (SSA232/233). gyrB served as an internal control for normalization and was analyzed with primer set WNp233/234. Q-PCR was performed with the SsoFast Evagreen Supermix (Bio-Rad) according to the manufacturer's instructions.
Two DNA fragments were employed in this assay; a 289 bp fragment of hilA (%GC = 34) from S. Typhimurium 14028s genomic DNA, and a 204 bp GC-rich fragment of PA3900 (%GC = 74) from Pseudomonas aeruginosa strain PAO1. The hilA fragment was amplified by PCR using primers GT068 and GT077 and the PA3900 fragment was amplified using primers GT049 and GT050 (Table 5). Various concentrations of purified H-NS, StpAWT and relevant variants of StpA were incubated with 10 nM DNA in binding buffer (15 mM HEPES pH 7.9, 40 mM KCl, 1 mM EDTA, 0.5% DTT, 5% glycerol) for 30 minutes. 4 µl of 6× Fermentas loading dye was added to each 20 µl reaction immediately prior to separation by gel electrophoresis for 2.5 h at 70 V on a 6% native polyacrylamide gel at 4°C (buffered with Tris acetate EDTA). Gels were stained with SYBR Green for 20 minutes at room temperature, washed twice with ddH2O, and DNA complexes were visualized with ultraviolet light.
|
10.1371/journal.pntd.0005710 | 2b-RAD genotyping for population genomic studies of Chagas disease vectors: Rhodnius ecuadoriensis in Ecuador | Rhodnius ecuadoriensis is the main triatomine vector of Chagas disease, American trypanosomiasis, in Southern Ecuador and Northern Peru. Genomic approaches and next generation sequencing technologies have become powerful tools for investigating population diversity and structure which is a key consideration for vector control. Here we assess the effectiveness of three different 2b restriction site-associated DNA (2b-RAD) genotyping strategies in R. ecuadoriensis to provide sufficient genomic resolution to tease apart microevolutionary processes and undertake some pilot population genomic analyses.
The 2b-RAD protocol was carried out in-house at a non-specialized laboratory using 20 R. ecuadoriensis adults collected from the central coast and southern Andean region of Ecuador, from June 2006 to July 2013. 2b-RAD sequencing data was performed on an Illumina MiSeq instrument and analyzed with the STACKS de novo pipeline for loci assembly and Single Nucleotide Polymorphism (SNP) discovery. Preliminary population genomic analyses (global AMOVA and Bayesian clustering) were implemented. Our results showed that the 2b-RAD genotyping protocol is effective for R. ecuadoriensis and likely for other triatomine species. However, only BcgI and CspCI restriction enzymes provided a number of markers suitable for population genomic analysis at the read depth we generated. Our preliminary genomic analyses detected a signal of genetic structuring across the study area.
Our findings suggest that 2b-RAD genotyping is both a cost effective and methodologically simple approach for generating high resolution genomic data for Chagas disease vectors with the power to distinguish between different vector populations at epidemiologically relevant scales. As such, 2b-RAD represents a powerful tool in the hands of medical entomologists with limited access to specialized molecular biological equipment.
| Understanding Chagas disease vector (triatomine) population dispersal is key for the design of control measures tailored for the epidemiological situation of a particular region. In Ecuador, Rhodnius ecuadoriensis is a cause of concern for Chagas disease transmission, since it is widely distributed from the central coast to southern Ecuador. Here, a genome-wide sequencing (2b-RAD) approach was performed in 20 specimens from four communities from Manabí (central coast) and Loja (southern) provinces of Ecuador, and the effectiveness of three type IIB restriction enzymes was assessed. The findings of this study show that this genotyping methodology is cost effective in R. ecuadoriensis and likely in other triatomine species. In addition, preliminary population genomic analysis results detected a signal of population structure among geographically distinct communities and genetic variability within communities. As such, 2b-RAD shows significant promise as a relatively low-tech solution for determination of vector population genomics, dynamics, and spread.
| Vector control has been the mainstay of Chagas disease control strategies in Latin America. Several Latin American countries implemented nation-wide insecticide-spraying programs to eradicate Chagas disease vector populations in human dwellings over the last 30 years. These campaigns resulted in a dramatic reduction in vectorial transmission [1–3]. Despite this success, domicile recolonization is a constant threat due to the ability of several triatomines species to disperse from sylvatic to domestic/peridomestic environments and establish local domestic populations [4–8].
Triatomines, members of the arthropod family Reduviidae, subfamily Triatominae, commonly known as kissing bugs, are distributed from the southern United States to central Argentina [9]. Over 130 species are identified, but only a few dozen are known to transmit Chagas disease [10]. In Ecuador, Triatoma dimidiata and Rhodnius ecuadoriensis are main vectors of Chagas disease, with the latter widely distributed from coastal and southern Ecuador to northern Peru [11,12].
Multiple molecular genetic studies exist which attempt to explain genetic structure and gene flow in triatomine populations [8,13–20]. An example of those tailored to address defined epidemiological hypotheses include that of Fitzpatrick et al. [13]. Fitzpatrick et al. confirmed that gene flow (and therefore vector dispersal) occurs between sylvatic, domicile and peridomicile ecotopes in Venezuelan Rhodnius prolixus based on pairwise FST values derived from both cytochrome b (cytb) and nine microsatellite loci. R. prolixus is the major vector species in Venezuela and Colombia, as well as Andean and Central American countries. Fitzpatrick et al.’s data suggested that colonization of domestic locales by wild triatomines is indeed possible in the region, and these findings had major implications for control. Other species have also been the subject of study. Population genetic data from Triatoma infestans based on ten microsatellite loci showed fine-scale genetic structure in domestic populations several years after the spraying of insecticides [18]. In this case, genetic data were tested under two different models of dispersal: isolation by distance and hierarchical island with stratified migration. The latter best reflected vector genetic structure among the sample sites. Finally, Almeida and colleagues [20] compared cytb and 8 microsatellite loci in Triatoma brasiliensis to investigate its genetic structure and to assess gene flow among sylvatic and domestic/peridomestic populations. As with Fitzpatrick et al. [13], pairwise comparison of FST values obtained from microsatellite loci analysis also demonstrated connectivity between locales.
Given that vector control remains the mainstay of Chagas disease intervention strategies, greater understanding of vector genetics and dispersal is urgently required. Of particular importance are genotyping approaches that provide very high resolution at local, epidemiologically relevant scales, as well as the ability to share and combine datasets across different studies and research groups. Microsatellite loci offer little flexibility in terms of shareability as data standardization guidelines for amplicon size estimation and allele nomenclature between laboratories, although possible [21], are rarely established, time-consuming and expensive to resolve, an issue already seen in Trypanosoma cruzi typing [22]. Likely as a function of funding constraints, molecular genetic research on triatomine vectors, and on Chagas disease in general, has been relatively late to arrive on the ‘omics’ scene. The belated publication of R. prolixus genome in 2015, as compared to other vector species, represents a step in the right direction and has revealed much about the core adaptations that underpin the biological success of triatomines [23]. A number of expressed sequence tags have been developed for T. infestans [24,25]. However, in general, genome sequencing efforts in triatomines so far have yielded little benefit to scientists and public health professionals attempting to map vector dispersal.
In tandem with the emergence of high throughput next generation sequencing (NGS) approaches, several groups have pioneered the use of restriction enzymes (REases) on restriction site-associated DNA sequencing (RADseq) protocols to allow a small fraction of the genome to be sequenced across multiple samples [26–34]. Several variants of the RADseq technique currently exist [35–39]; however, protocol choice to address a specific research question must balance technical issues, budget and laboratory capacity [40].
The 2b-RAD genotyping strategy specifically uses Type IIB restriction enzymes (IIB-REases) for genomic DNA (gDNA) digestion [38]. Advantages of this protocol include simplicity and cost-efficiency, since it is carried out in 3 steps in the same 96-well plate, as compared to 4–6 steps required in other RADseq protocols [35–37, 39]. Furthermore, library preparation can be achieved with no more than a PCR machine and a standard agarose gel. Moreover, IIB-REases capacity to generate identically sized 2b-RAD tags (IIB-REase-dependent) across all samples [38,40] and cleave at both strands of DNA removes the need for a post-digestion fragment size selection step. These characteristics also prevent fragment size [41] and strand [42] sequencing bias, which can compromise genotyping calls, as seen in other RADseq protocols. One disadvantage compared to other RADseq methods is that 2b-RAD may be inappropriate where accurate mapping against a highly duplicated/polyploid reference genome is required due to short fragment size production (33–36 bp) [43]. Finally, bias from PCR duplicates, sequencing errors and allele dropout can be introduced in all RADseq protocols.
In our study, we were able to rapidly and cost-effectively generate several hundred Single Nucleotide Polymorphism (SNP) markers for R. ecuadoriensis allowing for resolution of regional population genetic structure. Furthermore, by comparing the performance among the three IIB-REases, we were able to recommend the appropriate IIB-REase and read depth to employ in order to yield a given number of SNP markers for R. ecuadoriensis, and presumably for other members of the Rhodnius genus.
A total of 20 samples of R. ecuadoriensis were selected from the communities of La Extensa, Chaquizhca, and Coamine in Loja Province (southern Andean region), and from the community of Bejuco in Manabí (central coast) in Ecuador (Fig 1). Triatomines were captured in previous field surveys [44–46] from June 2006 to July 2013 (see S1 Table for further sample information). For each sample, head, legs and thoraxes were dissected and preserved in 100% ethanol.
A salt extraction protocol modified from Aljanabi and Martinez [47] was used to extract total gDNA from R. ecuadoriensis heads, legs and thoraxes (hindgut excluded). The modified protocol involved an additional overnight chitinase digestion step, as well as one overnight 75% ethanol wash to ensure purity (Table 1 and Fig 2). gDNA concentrations and purity ratios assessments were obtained by using NanoDrop ND-1000 Spectrophotometer (NanoDrop Technologies, Inc.). Integrity of the extracted DNA was evaluated by agarose electrophoresis and highly fragmented samples were excluded from subsequent analysis.
Initial selection of potential IIB-REases for our 2b-RAD protocol involved an in silico digestion of the R. prolixus genome, which is available from Genbank (accession code: KQ034056.1). For this purpose, 7 REases (AlfI, CspCI, BsaXI, SbfI, EcoRI, BcgI and KpnI) were screened. Three IIB-REases, namely AlfI, BcgI and CspCI were chosen based on the total number of restriction fragments produced in silico for the draft R. prolixus genome (www.vectorbase.org), financial resources, known efficiency in previous studies [31–33,38] and authors’ previous experience working with those enzymes [48]. We expected REases with abundant in silico restriction sites to show larger coverage variability among samples, at lower read depths. On the contrary, REases with less abundant restriction sites in silico could provide more exploitable markers at lower read depths.
Libraries were prepared using the 2b-RAD protocol proposed by Wang et al. [38] (Table 1 and Fig 2). Reaction mix and PCR conditions varied (S2 Table) depending on which IIB-REase was used. First, approximately 100–400 ng of high-quality gDNA from each sample was digested separately by each IIB-REase, producing IIB-REase-specific, uniform length fragments (32 bp, 35 bp and 33 bp for AlfI, BcgI and CspCI, respectively) with random overhangs. To confirm that the restriction reaction took place appropriately, equal amounts of digested DNA (dDNA) and gDNA from the same sample were visualized on a 1% agarose gel. Subsequently, the dDNA of each sample was ligated to a pair of partially double-stranded adaptors with compatible and fully degenerated overhangs (5’NNN3’). Finally, the obtained 2b-RAD tags were amplified to introduce a sample-specific 7bp barcode and the Illumina NGS annealing sites using two different pairs of sequencing primers. A 1.8% agarose gel electrophoresis of the PCR products was performed to verify the presence of the expected 150 bp target band (fragment, barcodes and adaptors included). In order to ensure an approximately equimolar contribution of each sample to the library, the exact amount of each PCR product was measured from the intensity of the target band in a digital image of the 1.8% agarose gel. We prepared three libraries in total, one for each IIB-REase, according to the relative concentration of each sample. The purification of the libraries from high-molecular weight fragments and primer-dimers was achieved first by removing the target band on agarose gel from each sample among the three libraries and eluting them in water overnight, followed by DNA capture with magnetic beads (SPRIselect Beckman Coulter) based on the Solid-Phase Reversible Immobilization method [49]. The DNA concentration in the purified libraries was quantified with a Qubit Fluorometer (Invitrogen) and the libraries were assembled in one single pool according to their relative concentrations. The library pool was sequenced on MiSeq (Illumina, San Diego, CA, USA) with a single 1x50 bp setup using ‘Version2’ chemistry at the Science for Life Laboratory (SciLifeLab, Stockholm, Sweden), which also implemented the reads demultiplexing and quality-filtering (Table 1 and Fig 2). Raw sequencing data has been uploaded to the Dryad Digital Repository (10.5061/dryad.02bf1).
The quality of demultiplexed and quality-filtered raw reads was verified by using FastQC software [50]. Subsequently, custom-made Python scripts were used for trimming the adaptors and then filtering the reads on the IIB-REase-specific recognition site (Table 1 and Fig 2). For each of the three libraries (AlfI, BcgI and CspCI) we sought to determine the relationship between sequencing effort (number of reads) and the total yield of polymorphic loci (set at up to two SNPs per locus). Therefore, we subsampled the total number of reads for each library in each individual using the fasta-subsample package from MEME SUITE [51] portal. This script randomly subsampled 25%, 50%, and 75% of total reads in triplicate to assess variability. This process resulted in 10 datasets per IIB-REase library: nine representing the three subsampling repetitions of the fixed percentages and only one from the total (100%) reads.
To estimate the polymorphic loci growth rate among the three IIB-REases, a nonlinear least square fitting (NLS) approach [52,53] was used with the R software [54] package NLS [55]. Specifically, NLS algorithm fits to the data by approximating a nonlinear function to a linear one, applying an iterative process to calculate the optimal parameter values for the growth rate [52,53,56]. Different built-in NLS models were tested in order to find the best fit to our data. These models were represented each with a different version of the Power-law equation [57]:
Y=aXb
(1)
Here, Y is the expected number of polymorphic loci at reads yield X; a is the estimate starting amount of Y when X is close to 0; b is the estimate of the relative change of Y in relation to a unit change in X (slope). A detailed description of the equations used for each dataset is provided in S3 Table.
All datasets created were analyzed separately using STACKS software version 1.42 [58], in which in silico assembly of loci and individual genotyping was performed by running the DENOVO_MAP.PL pipeline (Table 1 and Fig 2). STACKS algorithm, first, reconstructs stacks (alleles) from exactly matching reads of each sample (-m). These stacks are then either merged with others to form a single polymorphic locus or kept as separate monomorphic loci depending on the number of nucleotide mismatches (-M). Stacks with repetitive sequences are removed from the pipeline. Finally, each sample information is stored in a catalog (stored in the MySQL repository) containing the consensus (-n) of all loci and alleles in the entire population (See [58] tutorials).
Due to the failure of the protocol in one of the samples from the AlfI library (likely as a result of low gDNA quality), we decided to discard this sample from the other two datasets to avoid biased results in the de novo assembly. After several parameter adjustments, we set the minimum number of identical raw reads necessary to create a stack (-m) to 5. We kept the number of mismatches allowed between loci when building a locus in a single individual (-M) and when comparing across all individuals to build the population catalogue (-n) at default values. The bounded SNP calling model for identifying a SNP and estimating the sequencing error rate for calling at that SNP (—bound) ranged from 0 to 0.05. Finally, the significance level required to call a heterozygote or homozygote (—alpha) was set to 0.01. The EXPORT_SQL.PL utility was used to export loci shared by at least the 80% and the 90% of samples with the same polymorphism level (loci with up to 2 SNPs) from the MySQL database for all datasets analyzed in STACKS for each IIB-REase (Table 1 and Fig 2).
Although both total number of samples (N = 19) and sample size per community (N = 4–5) were low, we conducted pilot explorations of the population structure of R. ecuadoriensis in the study area. We retained polymorphic loci shared by at least 90% of the samples, characterized by the presence of 1 and 2 SNPs and with a minor allele frequency of 0.01. We performed preliminary genomic analysis using two different datasets: i) one dataset contained 361 polymorphic loci obtained from 18 samples processed with the BcgI IIB-REase (one sample was excluded from the analysis due to the high level of missing data) and ii) the second dataset contained 1225 polymorphic loci obtained from 19 samples processed with the CspCI IIB-REase. The number of markers obtained for the AlfI dataset derived from digestion with AlfI was too low to be used for the preliminary assessment of genomic structure of this particular sample. During the genotype calling, it is possible for more than one SNP to appear within the same region. When two SNPs were recovered at a single locus, a conservative approach was used to retain the first SNP for analysis, thereby excluding tightly linked SNP variation.
ARLEQUIN version 3.5 [65] was used to calculate non-hierarchical analysis of molecular variance [AMOVA; 66]. To deal with missing data, the locus-by-locus option was set. Bayesian clustering implemented in STRUCTURE 2.3.4 [67] was conducted to investigate the most likely number of clusters of genetically related individuals excluding the locality origin (model LOPRIORI). After several trials, a burn-in of 300 000 followed by 3 million runs for K = 1 to K = 4 and 5 iterations per each K value was set; admixture model and correlated allelic frequencies were assumed. The most probable number of clusters was identified from delta K, implemented online with STRUCTURE HARVESTER [68]. Then, in order to confirm our polymorphic loci was Rhodnius sp.-related, we also aligned the total polymorphic loci shared by at least the 90% of samples obtained from BcgI and CspCI datasets to the reference R. prolixus genome using BOWTIE 1 [69]. The highest alignment score (—best) was chosen and no more than 3 mismatches (-v) were allowed.
The extraction method allowed us to obtain RNA-free genomic DNA from all twenty samples with an average DNA concentration (ng/μL) of 62.77 ± 33.75 (s.d.) with average DNA purity ratios of 1.81 ± 0.05 (s.d.) and 1.81 ± 0.62 (s.d.) for absorbance at 280/260 and at 260/230, respectively (see S1 Table for detailed information). The in silico digestion on R. prolixus genome sequence by AlfI, BcgI and CspCI IIB-REases produced 204895, 103268 and 69984 putative cut sites, respectively.
The 2b-RAD experimental approach used in this study was effective for R. ecuadoriensis gDNA samples using any of the three IIB-REases (Fig 3), except for one sample (ID: CQ12, see S1 Table) digested by AlfI (CQ12 was thus not included in the pool for sequencing). A 2b-RAD pool of fifty-nine samples was established from nineteen samples digested by AlfI, twenty by BcgI, and twenty by CspCI IIB-REases.
The Illumina NGS yielded a total of 14.8 million de-multiplexed and quality-filtered reads, approximately 3, 6.2 and 5.6 million reads for AlfI, BcgI, and CspCI, respectively. FastQC analysis showed high per-base quality scores (> 32) for the reads of all samples processed with each of the three IIB-REases. After trimming the adaptors and filtering the IIB-REase-specific recognition site, 2.9, 5.8 and 4.8 million reads for AlfI, BcgI, and CspCI (respectively) were retained (Fig 4). The average trimmed Mreads per sample for each IIB-REase was 0.15 ± 0.06, 0.30 ± 0.04 and 0.25 ± 0.07. The number of reads subsampled and the total polymorphic loci for each IIB-REase are reported in Table 2. STACKS reference genome free runs assembled and identified a catalogue of loci from each of the datasets. The EXPORT_SQL.PL script was used to extract two datasets which included all the polymorphic loci with up to 2 SNPs shared by at least 80% and 90% of samples from each of the set percentages (25%, 50%, 75%, 100%) among the three replicates. We found only minor variation in the number of polymorphic loci called for each of the three subsampling replicates in all IIB-REase libraries. The average number of exported polymorphic loci obtained among replicates and from the total number of reads for each IIB-REase is reported in Table 2.
We observed growth in the number of loci recovered as we increased the read depth for all enzymes (Fig 5). However, while increasing read depth led to corresponding moderate and minor gains in locus number for BcgI and AlfI, respectively, for CspCI this number of loci is highlighted by a greater exponential growth in comparison to the other REases. Our results of best fit model analysis and estimated parameters (S3 Table) for each REase dataset were obtained by assessing different NLS models residual standard error, parameter significant p-values, number of iterations to convergence, the correlation between y and predicted values, and Akaike Information Criterion (AIC). In the first dataset (Fig 5A), we found that logarithmic (y∼a+bln(x)), geometric (y∼axbx) and exponential (y∼ae(bx)) NLS equations best fit to the AlfI, BcgI and CspCI datasets, respectively, allowing the estimation of growth rate parameters α and b (S3 Table). As for the second dataset (Fig 5B), geometric (y∼axbx) and Power-law (y∼axb) equations converged the best fit and parameters estimation for AlfI and BcgI, and CspCI, respectively (S3 Table). Detailed statistical analysis is provided in S1 Code.
The non-hierarchical AMOVA carried out on all four community samples for both datasets (BcgI and CspCI) detected a strong signal of genetic structuring across the study area, with highly significant (P< 0.0001) global FST values of 0.20452 (BcgI) and 0.39327 (CspCI). The most likely number of genetic clusters (K) identified by STRUCTURE was 2 for both datasets: on one side, the 3 samples from Loja region (CE, EX, CQ) were grouped together, and on the other, the sample from Manabí (BJ) was considered as a distinct cluster (Fig 6). The alignment to the R. prolixus reference genome resulted in a 42% and 31% of polymorphic loci aligned for BcgI and CspCI, respectively, likely due to genomic variability between the R. ecuadoriensis and the available R. prolixus reference genome as well as the difficulty in mapping short reads.
Our data demonstrate the power of 2b-RAD as a valid genotyping approach that can be applied to Chagas disease vectors for which either no reference genome exists or, as in our case, a reference genome exists for a species within the same genus. Our data broadly support the assertion of Wang et al. [38] that the 2b-RAD approach provides a simple, cost-effective and robust means of generating genome wide SNP data for non-model organisms. In our experiment, library preparation and sequencing was completed within a month and the cost per sample was approximately $18 USD (library preparation and sequencing cost), as compared to $30 USD per sample in other RADseq methods [39]. In fact, costs and technical complexity are two of the key factors when considering different RADseq protocols for a particular genomic study [59]. Moreover, laboratories/research groups deciding between “going RAD” or “keeping it classic” in terms of genotyping should assess whether a certain marker type addresses the research question at hand and fits their current and future research ambitions along with project budget. A total project/per sample cost analysis study showed that the cost of genotyping using microsatellite loci ($17.58 for 24 loci in four multiplexes) was less expensive compared to SNPs ($39.35 for 288 pooled samples and using a ddRAD-seq protocol [37]). However, it was assumed that a set of 16–24 microsatellite loci and species-specific primers already existed [70], somewhat unrealistic for some non-model organisms in which microsatellite primer development and validation should still be carried out and be considered within the project costs. After a literature search, the authors also pointed out that when studies genotyped microsatellite and SNPs in the same samples, the latter provided higher accuracy and/or precision for parameter estimation.
In our study, we have gone somewhat further than a proof-of-principle by evaluating the performance of three distinct Type IIB restriction enzymes, pre-screened in silico for their performance in terms of marker density against the Rhodnius prolixus genome [23]. Our methodological development aim was to test the predictability of the in silico cutter and to provide recommendations for suitable read depths, marker numbers and sample sizes for studies involving Rhodnius sp. vectors. We expected that an abundant in silico enzyme cutter would provide less usable molecular markers at lower read depths (Fig 4). It is important to highlight that, enzyme performance in silico in terms of number of restriction sites is not necessarily the same in an actual experiment due to genome size, nucleotide distribution, depth of coverage and GC composition [27,40,71]. Thus, a pilot experiment always offers valuable information on actual restriction enzyme performance.
Random re-sampling (rarefaction) of our datasets revealed distinct relationships between read depth and marker (polymorphic locus) number between the different enzymes, CspCI, BcgI and AlfI (Fig 5) broadly in line with predictions of the number of usable markers (Fig 4). As such, CspCI produced the largest amount of polymorphic markers regardless of read depth, evidencing its experimental performance for R. ecuadoriensis and likely for other Rhodnius sp. vectors. AlfI and BcgI, on the other hand, showed a marked tendency of deceleration for marker recovery as read depth increases. However, AlfI does show the initially steeper growth, in line with predictions that AlfI cut sites in the R. prolixus genome are more abundant (AlfI = 204895 sites, BcgI = 103268 sites and CspCI = 69984 sites). Additionally, we were able to fit nonlinear regression models to the data and estimate growth rate parameters for each enzyme (Fig 5). Although the model function varies per enzyme and dataset, all of them follow an exponential growth pattern which is more evident in CspCI datasets. The model function applied to the second CspCI dataset (Fig 5B) did not entirely fit the data; however, it constitutes the best fit compared to generalized linear models or more complex NLS fitting functions. Fitting NLS models to fewer data points for parameter estimation is challenging; however, based on our best-fit selection process we were confident that by substituting x for a determinate read depth we can obtain an estimate of polymorphic loci growth per restriction enzyme. Moreover, both parameters, a and b, are crucial for estimating the starting number of polymorphic loci and shape of the growth curve and understanding how the number of polymorphic loci changes as the number of reads increases. We hope this will be helpful to others planning similar studies.
At the read depth we achieved on one Illumina MiSeq single-ended run across 20 R. ecuadoriensis DNA samples, we generated 1244 markers for CspCI, 367 for BcgI and 68 for AlfI. Even the lowest of these values eclipses the size of marker panels currently in use to explore Triatomine population genetics [8,13–20]. However, to generate read depths to exploit the higher density IIB-REase cutters (e.g. AlfI, BcgI), a HiSeq approach might be more sensible. On the other hand, based on our data, CspCI can be expected to generate the best coverage and over a thousand polymorphic markers for approximately sixty vector samples on one MiSeq run. Interestingly, Graham et al. [72] assessed the impact of degraded gDNA in a modified double-digest-RAD protocol [37] on the MiSeq platform and found a significant correlation between DNA degradation, read quality reduction and loss. They also suggested that a higher throughput platform, HiSeq, and short fragment producer protocols, such as 2b-RAD, could help dealing with degraded gDNA and subsequent sequencing problems. As such, 2b-RAD might be an option for research teams with large and long-term stored triatomine bug collections, in which gDNA might already have started degradation processes. Based on our study, CspCI is the best candidate for generating enough usable markers, seconded by BcgI, and it is likely that a sequencing platform such as HiSeq can exploit a higher number of markers for both enzymes.
As well as ‘range finding’ for the application of 2b-RAD sequencing to triatomine populations, our second aim was to undertake preliminary population genomic analysis to explore genetic structuring in our study region. To this end, we focused on datasets generated with BcgI and CspCI since they presented higher numbers of polymorphic loci. An AMOVA indicated a significant proportion of variation was explained by between-population differences for both datasets. Moreover, we demonstrated the feasibility of our markers to distinguish structuring among populations in both BcgI and CspCI datasets. By using a Bayesian clustering framework our markers from both data sets detected two distinct clusters without previous location information, one of them was Bejuco, the clear geographic outlier with respect to Loja populations. Morphometric and genetic studies of R. ecuadoriensis in Ecuador would also predict a similar pattern of diversification [73,74]. However, inter-population diversification in Loja might be happening [74] at a rate undetectable by coarse test for isolation-by-distance and other conventional population analysis techniques. Our genomic information coupled with a landscape genetics/genomics framework could test whether landscape heterogeneity and environmental variables are driving such processes [64].
Earlier in the manuscript we presented the notion that, fewer steps, simplicity, cost-effectiveness, fragment size and strand bias absence are advantages of using a 2b-RAD protocol compared to other RADseq methods. Nevertheless, researchers must be aware of potential pitfalls and sources of bias accompanying all RADseq protocols, as well as most NGS-based methods. However, development of sophisticated analysis and more powerful software tools to deal with the types of issues produced by most NGS platforms is an active and evolving field of research [75]. During the initial steps of library preparation, degraded gDNA seems to have a greater impact on read quantity and quality in all other RADseq protocols than in 2b-RAD [72]. However, guidelines [27] for assessing gDNA quality should be implemented in all protocols. Another drawback in all RADseq methods is that polymorphism can occur at the restriction site. This so-called allele dropout (ADO) prevents enzymes from cutting at that location and thus precludes recovery of that SNP allele (null allele) [40,76]. ADO will have a direct impact in the estimation of allele frequencies and consequently in overestimation/underestimation of F-statistics as individual heterozygote at the null allele will be recognized as homozygote. However, filtering loci successfully genotyped among a high percentage of the samples can help to remediate the problem [40]. PCR duplicates arise in all RADseq protocols with a PCR step, and only identifiable in protocols with a random shearing digestion (original RADseq protocol [35,36]) as duplicate fragments are identified by having the same length. Another promising approach described by Andrews et al. [40] to identify PCR duplicates is to use degenerated base regions within sequencing adaptors to mark parent fragments. However, Puritz et al. [43] highlighted that, though untested, skewed allele frequencies by PCR artefacts have little effect in statistical bias within loci and thereby genotype calling errors. No less important are sequencing errors introduced in all Illumina instruments. Although several genotype-calling algorithms account for sequencing errors, a high depth sequencing coverage (≥ 20x) is always recommended. Finally, sequencing depth variability among loci could reduce genotyping accuracy for some less covered loci, thus allowing for fewer individuals to be multiplexed per sequencing lane, i.e., increasing cost per sample [40,59].
In our study, most of the above issues encountered in RADseq have been circumvented either during the library preparation or the raw data filtering steps. Nevertheless, our main challenge is the absence of a reference genome to map short reads in order to ensure that all markers do indeed belong to R. ecuadoriensis and not to microorganisms such as bacteria and fungi. Furthermore, it may be important to differentiate between mitochondrial and autosomal loci or sex-specific chromosomes that might have an effect in population divergence analysis. To overcome this difficulty, we adopted a stringent approach during raw data trimming and genotype calling. We focused analyses to loci shared by a high proportion of individuals and removed loci and samples with high amounts of missing data.
Landscape genetics/genomics is a powerful and relatively new approach to explore the underlying spatial processes that affect genetic diversity in biological organisms [63]. Next to isolation-by-distance, isolation-by-resistance is a common null hypothesis tested in landscape genetics when more complex ecological and environmental processes are thought to be at play. The landscape genetics framework and tools such as causal modelling and environmental association analysis have the potential [63,64, 77–79] to uncover whether the same is true for R. ecuadoriensis genetic structuring and dispersal in Ecuador. In our study, the main limitation to carry out a wide range of conventional between and within-population analysis was the sample size per population. Low sample size required our analyses to consider an extended area to resist exploration of processes at finer geographic scales.
The high-resolution genotyping approach we have developed in this study now paves the way for landscape genetic/genomics analysis in vector-parasite systems [64], with genuine potential insights for rational disease and entomological control. For example, landscape genetics approaches expanded our understanding of the natural and human-aided dispersal dynamics of the invasive Asian tiger mosquito, Aedes albopictus [80]. Similarly, insecticide resistance gene spread in Anopheles sinensis has been tracked in China using landscape genetics approaches, demonstrating multiple origins and the importance of long term agricultural insecticide use [81]. More widely, high resolution SNP datasets are increasingly used to explore the local and international spread of important disease vectors (e.g. Aedes aegypti [29,82]).
2b-RAD typing not only promises a potential applicability for population genetic studies but also for linkage and quantitative loci mapping given that marker density can be controlled using selective adaptors [38]. In fact, via its GENOTYPE pipeline, the STACKS package potentiates the construction of genetic maps from F2 or backcrosses of R. ecuadoriensis or other triatomine species.
In conclusion, the decreasing cost and increasingly simplicity of approaches to generate high resolution SNP data puts such tools increasingly in the hands of researchers in endemic countries working on non-model organisms that act as vectors of Neglected Tropical Diseases. An analytical framework to incorporate detailed spatial and environmental variation into genetic analyses is now in place to facilitate a better understanding of the biology and dispersal of disease vectors.
|
10.1371/journal.pntd.0001626 | In Vitro and In Vivo Activity of a Palladacycle Complex on Leishmania (Leishmania) amazonensis | Antitumor cyclopalladated complexes with low toxicity to laboratory animals have shown leishmanicidal effect. These findings stimulated us to test the leishmanicidal property of one palladacycle compound called DPPE 1.2 on Leishmania (Leishmania) amazonensis, an agent of simple and diffuse forms of cutaneous leishmaniasis in the Amazon region, Brazil.
Promastigotes of L. (L.) amazonensis and infected bone marrow-derived macrophages were treated with different concentrations of DPPE 1.2. In in vivo assays foot lesions of L. (L.) amazonensis-infected BALB/c mice were injected subcutaneously with DPPE 1.2 and control animals received either Glucantime or PBS. The effect of DPPE 1.2 on cathepsin B activity of L. (L.) amazonensis amastigotes was assayed spectrofluorometrically by use of fluorogenic substrates. The main findings were: 1) axenic L. (L.) amazonensis promastigotes were destroyed by nanomolar concentrations of DPPE 1.2 (IC50 = 2.13 nM); 2) intracellular parasites were killed by DPPE 1.2 (IC50 = 128.35 nM), and the drug displayed 10-fold less toxicity to macrophages (CC50 = 1,267 nM); 3) one month after intralesional injection of DPPE 1.2 infected BALB/c mice showed a significant decrease of foot lesion size and a reduction of 97% of parasite burdens when compared to controls that received PBS; 4) DPPE 1.2 inhibited the cysteine protease activity of L. (L.) amazonensis amastigotes and more significantly the cathepsin B activity.
The present results demonstrated that DPPE 1.2 can destroy L. (L.) amazonensis in vitro and in vivo at concentrations that are non toxic to the host. We believe these findings support the potential use of DPPE 1.2 as an alternative choice for the chemotherapy of leishmaniasis.
| Leishmaniasis is an important public health problem with an estimated annual incidence of 1.5 million of new human cases of cutaneous leishmaniasis and 500,000 of visceral leishmaniasis. Treatment of the diseases is limited by toxicity and parasite resistance to the drugs currently in use, validating the need to develop new leishmanicidal compounds. We evaluated the killing by the palladacycle complex DPPE 1.2 of Leishmania (Leishmania) amazonensis, an agent of human cutaneous leishmaniasis in the Amazon region, Brazil. DPPE 1.2 destroyed promastigotes of L. (L.) amazonensis in vitro at nanomolar concentrations, whereas intracellular amastigotes were killed at drug concentrations 10-fold less toxic than those displayed to macrophages. L. (L.) amazonensis-infected BALB/c mice treated by intralesional injection of DPPE 1.2 exhibited a significant decrease of foot lesion sizes and a 97% reduction of parasite burdens when compared to untreated controls. Additional experiments indicated the inhibition of the cathepsin B activity of L. (L.) amazonensis amastigotes by DPPE 1.2. Further studies are needed to explore the potential of DPPE 1.2 as an additional option for the chemotherapy of leishmaniasis.
| Protozoan parasites of the Leishmania genus induce cutaneous, mucocutaneous and visceral diseases in man and animals. According to the World Health Organization, about 1.5 million of new human cases of cutaneous leishmaniasis and 500,000 of visceral leishmaniasis are registered annually [1]. Leishmania (Leishmania) amazonensis, one of the causative agents of human cutaneous leishmaniasis in the Amazon region, Brazil, is associated with both the simple and diffuse forms of the disease [2]. The first-line drugs used for treatment of leishmaniasis are pentavalent antimonial compounds, while amphotericin B and pentamidine are used as the second-line chemotherapy. However, the use of these compounds is limited by toxicity to the host and the development of resistance by the parasites [3], [4]. Thus, the development of new leishmanicidal drugs is an important goal and several compounds including synthetic, natural products extracted from plants and marine sources have shown different degrees of efficacy in the treatment of experimental leishmaniasis [5]–[7]. The in vitro and in vivo demonstration that the viability of the Leishmania parasites is reduced by inhibitors of cysteine proteases [8]–[10] encouraged the use of virtual screening to identify additional inhibitors [11], [12]. The demonstration that antitumor drugs may also display antileishmanial activity has also stimulated the screening of these compounds in vitro and in clinical trials [13]. Cyclopalladated complexes have shown in vitro and in vivo antitumor activity and low toxicity in animals [14]–[16] and more recently one of them exhibited lethal effects on human leukaemia cells while was ineffective against normal human lymphocytes [17]. The leishmanicidal and tripanocidal activity of cyclopalladated complexes has also been demonstrated [18]–[20]. Furthermore, there is evidence that palladacycle complexes may destroy tumoral cells by inhibition of cathepsin B activity and their inhibitory effect on Leishmania cysteine proteases in vitro was also demonstrated [18], [21]. The present study describes the effect of one palladacycle compound called DPPE 1.2 on promastigotes, intracellular amastigotes and cutaneous lesions in mice infected with L. (L.) amazonensis.
Eight-week-old female Golden hamsters were obtained from breeding stocks maintained at the Universidade de Campinas (São Paulo, Brazil) and female BALB/c mice 6 to 8 weeks old were acquired from Universidade Federal de São Paulo (São Paulo, Brazil). This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the Brazilian National Council of Animal Experimentation (http://www.cobea.org.br). The protocol was approved by the Committee on the Ethics of Animal Experiments of the Institutional Animal Care and Use Committee at the Federal University of São Paulo (Id # CEP 1844/08).
The L. (L.) amazonensis strain used (MHOM/BR/1973/M2269) was kindly provided by Dr. Jeffrey J. Shaw, Instituto Evandro Chagas, Belém, Pará, Brazil and maintained as amastigotes by inoculation into footpads of Golden hamsters every 4 to 6 weeks. Amastigote suspensions were prepared by homogenization of excised lesions, disruption by four passages through 22-gauge needles, and centrifugation at 250×g for 10 min; the resulting supernatant was centrifuged at 1,400×g for 10 min, and the pellet was resuspended in RPMI 1640. The suspension was kept under agitation for 4 h at room temperature and centrifuged at 250×g for 10 min. The final pellet contained purified amastigotes which were essentially free of contamination by other cells [22].
L. (L.) amazonensis promastigotes were grown at 26°C in 199 medium (Gibco) supplemented with 4.2 mM sodium bicarbonate, 4.2 mM HEPES, 1 mM adenine, 5 µg/ml hemin (bovine type I) (Sigma, St Louis, MO, USA) and 10% fetal calf serum (FCS) (Cultilab, SP, Brazil).
The palladacycle compound DPPE 1.2 (Figure 1) was obtained from N,N-dimethyl-1-phenethylamine (DMPA), complexed to 1,2-ethane-bis (diphenylphosphine) (DPPE) ligand and synthesized as previously described [16]. Stock solutions at 1.45 mM were prepared in dimethylsulfoxide (DMSO); for in vitro use, the drug was diluted to the appropriate concentration in cell culture medium, and for in vivo injections the stock was diluted in PBS.
The promastigote cultures at 1×106 parasites/ml were kept in 199 culture medium as described above containing between 1.25 nM and 150 nM of DPPE 1.2. Parasites were counted daily in a Neubauer chamber for three days. The leishmanicidal effect of DPPE 1.2 on intracellular amastigotes was evaluated in mouse bone marrow derived macrophages infected with L. (L.) amazonensis. Bone marrow-derived macrophages were generated from bone marrow stem cells isolated from BALB/c mice [23]. Cells were counted, added (8×105) and cultured on glass coverslips inserted in 24-well tissue culture plates containing RPMI 1640 medium buffered with 15 mM of HEPES, 20 mM of sodium bicarbonate and supplemented with 1 mM L-glutamine, 20% of fetal calf serum (FCS) and 30% L929 cell conditioned medium (LCCM). Cultures were kept at 37°C in an atmosphere of air/CO2 (95/5%). After 5 days, the medium was changed for RPMI containing 10% of FCS and macrophages were infected at a multiplicity of 2 amastigotes per macrophage. After 24 h, infected cultures were treated with different drug concentrations (150 to 500 nM) for 3 days. The coverslips were fixed with methanol, stained with hematoxylin-eosin (HE) and intracellular amastigotes were counted. Results are expressed by the infection index, obtained by multiplying the percentage of infected macrophages by the average number of amastigotes per macrophage. At least 200 macrophages were scored in each 3 coverslips. Amphotericin B (Sigma-Aldrich, St Louis, MO, USA) and Glucantime (Sanofi-Aventis, Brazil, 300 mg/ml, 81 mg/ml SbV) were used as standard drugs for treatment of L. (L.) amazonensis promastigotes and intracellular amastigotes, respectively.
DPPE 1.2 cytotoxicity to macrophages was tested by a MTT micromethod described previously [24] after incubation of bone marrow derived macrophages with 150 to 2,000 nM of DPPE 1.2 for 3 days. Macrophages were also incubated with the highest concentration of DMSO used for DPPE 1.2 solubilization (0.04%). The formation of formazan was measured by adding 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT; Molecular Probes, Eugene, OR, USA) 0.5 mg/ml and incubation of the cultures at 37°C in the dark. After 4 h the medium was removed, 200 µl of DMSO was added per well and the absorbance was measured using an ELISA reader at 540 nm (Labsystems Multiskan).
For evaluation of in vivo leishmanicidal activity of DPPE 1.2 female BALB/c mice 6 to 8 weeks-old were infected subcutaneously at the right hind-foot with 1×105 L. (L.) amazonensis amastigotes. Fifteen days after infection, the animals were randomly separated in 3 groups of 12 mice each. Treated animals received in the foot lesions every other day doses of 60 mg/kg/day (16.8 mg [Sbv]/kg/day) of Glucantime for 1 month (total of 900 mg/kg–252 mg [Sbv]/kg/day) or doses of 320 µg/kg/day of DPPE 1.2 (total of 4.8 mg/kg). Stock solutions of DPPE 1.2 were prepared daily in PBS after solubilization in DMSO (final concentration of 0.1%). Control group received the same number of injections of PBS. Infection was monitored once a week by measuring the diameter of foot lesions with a dial caliper (Mitutoyo Corp., Japan). Parasite burden from infected feet was determined by a limiting dilution method, as previously described [25].
Serum concentrations of urea, creatinine, bilirubin and transaminases were determined in BALB/c mice at the end of treatment, using sets of commercial reagents (Doles Reagentes e Equipamentos para Laboratórios, Ltda, Brazil).
Proteolytic activity of L. (L.) amazonensis promastigotes and amastigotes was determined by zymography employing electrophoretic separation of parasite lysates under unheated and nonreduced conditions resolved on 10% acrylamide gels containing 0.1% copolymerized gelatin (Gibco-BRL) by low-voltage (50 V) electrophoresis [26]. Proteolytic activity in the gels was detected after 1 h of incubation, under agitation, in 0.1 M sodium acetate buffer, pH 5.0, containing 2.5% Triton X-100, followed by 2 h of incubation in the acetate buffer in the absence of Triton X-100 and Coomassie blue staining. Some gel strips after electrophoresis were incubated in buffer solutions in the presence of either protease inhibitor E-64 (trans-epox-isuccinil-L-leucinamide-(4-guanide-butane) or orthophenanthroline or DPPE 1.2. Molecular weight markers (Pharmacia LKB) were visible on the background of stained gelatin when used in a 5-fold excess.
Cathepsin activities were monitored with the fluorogenic substrates Z-Phe-Arg-AMC (for all cathepsins), Z-Arg-Arg-AMC (for cathepsin B), and Z-Leu-Arg-AMC (for cathepsins K, V, and S) (commercially obtained from Sigma, St. Louis, MO, USA) using 1 µl of L. (L.) amazonensis amastigote cell lysate (1×109 amastigotes disrupted in 200 µl PBS), 2 mM DTT (dithiothreitol), 1 ml of four-component buffer comprised of 25 mM acetic acid, 25 mM Mes (4-Morpholineethanesulfonic acid), 75 mM Tris, and 25 mM glycine, pH 5.0, 10 µM of each fluorogenic substrate and 50 µM of DPPE 1.2. The effect of DPPE 1.2 on the parasite enzyme activity was tested by incubation of the L. (L.) amazonensis lysate with DPPE 1.2 for 2 minutes in the buffer solution a 37°C; the fluorogenic substrate was then added and the fluorescence of the released fluorophore, 7-amino-4-methylcoumarine (AMC), was measured over time. The remaining enzyme activities were determined and expressed as a percentage of the activity of the control experiment. Parasite lysate was also incubated with 10 µM of the fluorogenic substrate Abz-Gly-Ile-Val-Arg-Ala-Lys(Dnp)-OH (Sigma, St. Louis, MO, USA), specific for cathepsin B [27], in the presence of either increasing concentrations of DPPE 1.2 or CA074, a specific inhibitor of cathepsin B. The cathepsin activity was monitored spectrofluorometrically using the fluorogenic substrates on a Hitachi F-2000 spectrofluorometer equipped with a thermostated cell holder. The fluorescence excitation (λEx) and emission (λEm) wavelengths, for the fluorescence of AMC, were set at 380 nm and 460 nm, respectively, while the parameters for the fluorescence of Abz-peptide fragments resulting from the Abz-Gly-Ile-Val-Arg-Ala-Lys(Dnp)-OH hydrolysis were set at λEx = 320 and λEm = 420 nm.
To determine the statistical differences between groups ANOVA and Student's t test were used and P values<0.05 or lower were considered statistically significant. IC50 and CC50 values were determined by GraphPad Prism, version 5.0.
Axenic cultures of L. (L.) amazonensis promastigotes were grown in the presence of 1.25 to 150 nM of DPPE 1.2. Significant inhibition of parasite growth was detected after 2 and 3 days of treatment with 1.25 to 25 nM of DPPE 1.2. At 75 and 150 nM the drug inhibited 84% and 96%, respectively, of parasite growth 1 day after treatment and nearly 100% of promastigotes were killed after 2 and 3 days in the presence of these concentrations of DPPE 1.2. A growth curve similar to control was observed when L. (L.) amazonensis promastigotes were cultured in the presence of the highest concentration of DMSO used for DPPE 1.2 solubilization (0.04%). As a control drug parasites were grown in the presence of amphotericin B. After 72 of incubation, the IC50 values for both drugs were determined (Table 1).
Treatment with DPPE 1.2 resulted in a significant, dose-dependent decrease in infection index of L. (L.) amazonensis-infected macrophages with an inhibition of 92% for 500 nM of DPPE 1.2 (IC50 of 128.35 nM; 95% confidence limits, 111.2–164.2 nM) (Figure 2A). Infected cultures were also treated with the highest concentration of DMSO used for DPPE 1.2 solubilization (0.04%); these concentrations did not reduce the viability or the infection of macrophages (data not shown). The cytotoxicity of DPPE 1.2 on macrophages was evaluated by the MTT method and the CC50 was determined (1,267 nM; 95% confidence limits, 1,15–1,52 nM).
Figure 2B). The IC50 value expressed as µg/ml of pentavalent antimony [Sbv] was 178.5 µg/ml (95% confidence limits, 108.1–294.6 µg/ml) and treatment with the drug at this concentration resulted in 40% of macrophage toxicity (data not shown). The calculated CC50 for Glucantime expressed as µg/ml of pentavalent antimony [Sbv] was 266.3 µg/ml (95% confidence limits, 252.2–290.3 µg/ml).
BALB/c mice infected with L. (L.) amazonensis were treated every other day with 320 µg/kg/day of DPPE 1.2 for 1 month injected in foot lesions. As can be observed in Figure 3, starting from 24 days of treatment the animals which received DPPE 1.2 showed a significant decrease of foot lesion size compared to controls. Starting from 16 days of treatment, the animals that received Glucantime also exhibited significantly smaller foot lesions compared to untreated control, as well as to animals treated with DPPE 1.2.
Parasite load was also evaluated by limiting dilution in foot lesions of BALB/c mice one month after end of the treatment with either DPPE 1.2 or Glucantime. Figure 4 shows that BALB/c mice treated with either DPPE 1.2 or Glucantime displayed a reduction of parasite load of 97% and 99%, respectively, compared to untreated animals.
To evaluate hepato and nephrotoxicity of DPPE 1.2 serum levels of transaminases, urea and creatinine were determined. No statistically significant alterations were detected between groups (data not shown).
Parasite proteolytic activity was determined by zymography after electrophoresis of L. (L.) amazonensis extracts in SDS-PAGE with gelatin coupled gels. Figure 5A shows that most of the proteolytic activity of L. (L.) amazonensis promastigotes that migrates as a 60 kDa band was abolished in the presence of orthophenanthroline, while DPPE 1.2 did not show any effect on this activity. On the other hand, amastigotes displayed a strong activity at a molecular mass of 30–35 kDa that was totally inhibited by either DPPE 1.2 or E-64, indicating that DPPE 1.2 inhibits cysteine protease activity of L. (L.) amazonensis amastigotes.
A spectrofluorometric assay using specific substrates for cathepsins in the presence of DPPE 1.2 was also carried out. Figure 5B shows that the L. (L.) amazonensis amastigote extract exhibited high hydrolytic activity on all substrates tested. Although DPPE 1.2 inhibited the enzymatic activity on all substrates, a significantly higher reduction on cathepsin B activity could be observed in the presence of this palladacycle complex (75%). Figure 5C shows that the activity of L. (L.) amazonensis extract on a most specific substrate for cathepsin B was significantly inhibited either by DPPE 1.2 or CA074 (data not shown). The calculated IC50 values for DPPE 1.2 and CA074 were not significantly different (2.25±0.11 µM and 0.7±0.08 µM, respectively), strongly suggesting that DPPE 1.2 inhibits L. (L.) amazonensis cathepsin B.
The present results document the leishmanicidal effect of the palladacycle complex DPPE 1.2 on L. (L.) amazonensis. This compound destroyed L. (L.) amazonensis promastigotes at very low concentrations. Extension of this study to L. (L.) amazonensis-infected macrophages also showed an effective leishmanicidal activity of DPPE 1.2 against amastigotes, whereas the drug displayed 10-fold less toxicity to macrophages. Although similar leishmanicidal effect was observed with Glucantime, significantly higher concentrations of this antimonial were necessary to destroy L. (L.) amazonensis amastigotes. The leishmanicidal activity of DPPE 1.2 is comparable to that obtained with several compounds tested against L. (L.) amazonensis like mesoionic salt derivatives, primary S-nitrosothiols, aureobasidin A, julocrotine, tamoxifen, elatol [28]–[33]. However, higher concentrations of these compounds were used to destroy L. (L.) amazonensis, whereas an effective leishmanicidal effect was observed with DPPE 1.2 at nanomolar range. The leishmanicidal activity of metal complexes of gold, platinum, iridium, rhodium and osmium has also been investigated. However, most of them destroyed only L. (L.) donovani promastigotes, while few reduced the parasitism in infected animals [34]–[36], impairing the comparison with DPPE 1.2 data. Among other palladium complexes previously tested against Leishmania only one was an effective inhibitor of promastigote growth, while none of them reduced the intracellular amastigote burden [18].
The leishmanicidal effect of DPPE 1.2 was also demonstrated in vivo. Although the reduction of parasite load in foot lesions of L. (L.) amazonensis-infected mice treated with DPPE 1.2 was similar to that obtained with Glucantime, this antimonial compound was used in 200 times higher concentration. Treatment with DPPE 1.2 led to a significant reduction of parasite load in foot lesions (97%), but did not result in sterile cure in infected mice. However, it is important to emphasize that the BALB/c strain is highly susceptible to L. (L.) amazonensis infection. These mice develop a gradual increase of foot lesions characterized by a large infiltrate of macrophages harboring a high number of amastigotes, thus mimicking the anergic form of diffuse cutaneous leishmaniasis caused by L. (L.) amazonensis [2]. The apparent lack of toxicity of DPPE 1.2 to BALB/c mice was demonstrated by hepatic and renal function assays after treatment with the drug, corroborating data that showed the low toxicity of palladacycle complexes in the treatment of mice against tumor cells [16]. More recently, the high selectivity index of one cyclopalladacycle complex with trypanocidal activity was also demonstrated, suggesting the use of this compound for treatment of Chagas' disease [19].
As reported in the literature, the antitumor property of palladacycle complexes can be attributed, at least in part, to their inhibitory activity on the cysteine protease cathepsin B [21]. This information led us to test for the possible effect of DPPE 1.2 on L. (L.) amazonensis protease activity. We showed that DPPE 1.2 did not inhibit the activity of the metalloprotease gp63, the major surface protein of Leishmania promastigotes [37]. On the other hand, the high cysteine protease activity expressed in L. (L.) amazonensis amastigotes was inhibited by DPPE 1.2 and the most significant inhibition was observed on the cathepsin B activity. However, the drug did not affect the cysteine proteinase activity of mouse macrophages (data not shown). Several studies demonstrated the involvement of cathepsin L-like (cpL) and cathepsin B-like (cpB) in Leishmania growth and virulence in vitro and in vivo [38]–[40]. Furthermore, cysteine proteinase inhibitors have been reported to kill Leishmania in vitro and in vivo [8]–[10]. In the present study we show that in vitro DPPE 1.2 inhibited L. (L.) amazonensis cathepsin B at higher concentrations than those necessary to kill L. (L.) amazonensis. These findings argue against a relationship between the leishmanicidal effect of DPPE 1.2 and the inhibition of L. (L.) amazonensis cathepsin B and suggest that other relevant targets may account for the leishmanicidal effect of the drug. Palladacycle complexes have been associated with organelle-specific effects in tumor cells such as the lysosomal and mitochondrial permeabilization that can trigger apoptosis [41], [42]. The induction of L. (L.) amazonensis apoptosis by DPPE 1.2 has not been investigated.
Leishmania killing is associated to macrophage activation by IFN-γ and TNF-α and the production of nitric oxide [43], whereas TGF-β is an immunosupressor cytokine known to exacerbate visceral and cutaneous leishmaniasis [44]–[47]. Interestingly, Leishmania cathepsin B is involved in the conversion of latent TGF-β to its biologically active form [48]. Since we have shown here that DPPE 1.2 inhibits parasite cathepsin B, killing of L. (L.) amazonensis in infected mice treated with the drug may be associated to protective responses arising from lower expression of the active form of TGF-β. This possibility is now under investigation.
In conclusion, the effectiveness of DPPE 1.2 in destroying L. (L.) amazonensis in vitro and by intralesional administration in vivo at concentrations non toxic to the host support further studies of the leishmanicidal activity of the palladacycle as an additional choice to available chemotherapies.
|
10.1371/journal.pbio.1000465 | IL-6 and IL-10 Anti-Inflammatory Activity Links Exercise to Hypothalamic Insulin and Leptin Sensitivity through IKKβ and ER Stress Inhibition | Overnutrition caused by overeating is associated with insulin and leptin resistance through IKKβ activation and endoplasmic reticulum (ER) stress in the hypothalamus. Here we show that physical exercise suppresses hyperphagia and associated hypothalamic IKKβ/NF-κB activation by a mechanism dependent upon the pro-inflammatory cytokine interleukin (IL)-6. The disruption of hypothalamic-specific IL-6 action blocked the beneficial effects of exercise on the re-balance of food intake and insulin and leptin resistance. This molecular mechanism, mediated by physical activity, involves the anti-inflammatory protein IL-10, a core inhibitor of IKKβ/NF-κB signaling and ER stress. We report that exercise and recombinant IL-6 requires IL-10 expression to suppress hyperphagia-related obesity. Moreover, in contrast to control mice, exercise failed to reverse the pharmacological activation of IKKβ and ER stress in C3H/HeJ mice deficient in hypothalamic IL-6 and IL-10 signaling. Hence, inflammatory signaling in the hypothalamus links beneficial physiological effects of exercise to the central action of insulin and leptin.
| The hypothalamus is a brain region that gathers information on the body's nutritional status and governs the release of multiple metabolic signaling molecules such as insulin and leptin to maintain homeostasis. Overeating and obesity are associated with insulin and leptin resistance in the hypothalamus, and recent studies provide an intriguing link between inflammation and dysfunction of hypothalamic insulin and leptin signaling through activation of IKKβ, a key player in immune response, and endoplasmic reticulum (ER) stress. This means that strategies to reduce the aberrant activation of inflammatory signaling in the hypothalamus are of great interest to improve the central insulin and leptin action and prevent or treat related metabolic diseases. Using a combination of pharmacological, genetic, and physiological approaches, our study indicates that physical activity reorganizes the set point of nutritional balance through anti-inflammatory signaling mediated by interleukin (IL)-6 and IL-10 in the hypothalamus of rodents. Hence, IL-6 and IL-10 are important physiological contributors to the central insulin and leptin action mediated by exercise, linking it to hypothalamic ER stress and inflammation.
| Overnutrition and sedentary lifestyle are among the most important factors that lead to an unprecedented increase in the prevalence of obesity. In mammals, food intake and energy expenditure are tightly regulated by specific neurons localized in the hypothalamus. The hypothalamus can gather information on the body's nutritional status by integrating multiple signals, including potent hormonal signals such as insulin and leptin [1],[2]. The impairment of hypothalamic insulin and leptin signaling pathways is sufficient to promote hyperphagia, obesity, and type 2 diabetes (T2D) in different genetic rodent models with neuronal ablation of insulin and leptin signaling [1],[3],[4]. We and others have proposed that overnutrition induces the central insulin and leptin resistance through the aberrant hypothalamic activation of proinflammatory molecules, including TLR4 and IKK [5]–[7].
IKKβ is a key player in controlling both innate and adaptive immunity. Activation of IKKβ by phosphorylation at S177 and S181 induces phosphorylation, ubiquitination, and subsequent proteosomal degradation of its substrate IκBα. The degradation of IκBα allows NF-κB proteins to translocate to the nucleus and bind their cognate DNA binding sites to regulate the transcription of a large number of genes, including stress-response proteins and cytokines [8]. Growing evidence provides an intriguing link between metabolic inflammation and dysfunction of insulin and leptin signaling via activation of IKKβ and endoplasmatic reticulum (ER) stress [9]–[14]. Examination of ER stress markers in different tissues of dietary (high-fat diet-induced) and genetic (ob/ob) mouse models of obesity demonstrated increased levels of PERK phosphorylation and JNK and IKKβ activity [7],[12]. In addition, a recent study showed the activation of hypothalamic IKKβ/NF-κB, at least in part, through elevated endoplasmic reticulum stress in the hypothalamus and that these phenomena are associated with central insulin and leptin resistance, hyperphagia, and body weight gain in mice [7]. Thus, strategies to reduce the aberrant activation of inflammatory signaling and/or ER stress in hypothalamic neurons are of great interest to improve the central insulin and leptin action and prevent or treat obesity and related diseases.
Physical activity is considered a cornerstone of the treatment for obesity. Exercise has long been reported to reduce body weight and visceral adiposity, increasing the energy expenditure and improving glycaemic control in overweight or T2D patients [15],[16]. Since the discovery of interleukin (IL)-6 releases from contracting skeletal muscle, accumulating evidence indicates that exercise induces metabolic changes in other organs, such as the liver, the adipose tissue, and hypothalamus, in an IL-6 dependent manner. IL-6 is most often classified as a pro-inflammatory cytokine, although consistent data also demonstrate that IL-6 has an anti-inflammatory effect and may negatively regulate the inflammation of acute phase response by increasing IL-10, IL-1 receptor antagonist (IL-1ra), and soluble TNF-receptors (sTNF-R) [17]. Moreover, IL-6 appears to play a central role in the regulation of appetite, energy expenditure, and body composition [18],[19]. However, the effects of physical activity in the metabolic regulatory pathways in the central nervous system (CNS) remain unexplored. Thus, we hypothesized that exercise could exert its effects in the CNS by modulating the specific hypothalamic neurons responsible for the control of food consumption. In the present study, we investigated the effect of the anti-inflammatory response, mediated by IL-6, on hypothalamic IKKβ activation and ER stress, central insulin and leptin sensitivity, and food intake in diet-induced rats after physical activity.
It has been demonstrated that physical activity may contribute to the energy balance by increasing energy expenditure. Although the energy expenditure aspects of such exercise may contribute to the effects of weight loss, the effect of exercise on the control of energy intake remains unclear. To evaluate the impact of physical activity on food consumption, we measured the 12-h total energy intake in lean and diet-induced obese (DIO) rats after one bout of swimming (SW Exe) and treadmill running (TR Exe) exercise. Neither of the exercise protocols changed the energy intake in lean animals; however, exercise suppressed the hyperphagic response, mediated by chronic overnutrition, restoring the energy intake to the levels of lean animals (Figure 1A). To assess whether the effects of exercise on food intake are dependent on the neuropeptides modulation, we performed a real time PCR assay to determine the mRNA levels of Neuropeptide-Y (NPY) and Proopiomelanocortin (POMC). After 9 h of fasting, we found that chronic overnutrition increased NPY mRNA and reduced POMC mRNA levels, while physical activity restored the NPY (Figure 1B) and POMC mRNA levels (Figure 1C) in obese animals; on the other hand, exercise did not change the NPY and POMC mRNA levels in lean rats (Figure 1B and C).
Chronic overnutrition increased body weight, epididymal fat (Figure 1D and E), serum insulin, leptin, triglycerides, and free fatty acid levels (Table 1), compared to age-matched controls. No significant variations were found in body weight, epididymal fat serum leptin, triglycerides, and urinary corticosterone levels between exercised and obese animals under resting conditions (Figure 1D, E and Table 1). The insulin levels were lower in both lean and obese rats after the exercise protocols (Table 1) and exercise increased the free fatty acid in obese animals (Table 1). To determine whether lean and obese rodents were swimming or running in the same fashion, we evaluated lactate production every 15 min during the SW Exe and TR Exe. We did not find any difference in the lactate production between lean and obese rats. Table 1 depicts the final values obtained in this test. These results reinforce the negative relationship between body weight change and stress related with the appetite-suppressive actions mediated by exercise.
To extend our hypothesis, we investigated food intake in leptin-deficient mice (ob/ob) after physical activity. Acute SW Exe and TR Exe did not change the food intake in wild type (WT) mice, however the food consumption was reduced in ob/ob mice (Figure 1F). After 9 h of fasting, we found that NPY mRNA was increased and POMC mRNA levels were reduced in ob/ob mice, while physical activity restored the NPY (Figure 1G) and POMC mRNA levels (Figure 1H) in obese animals; on the other hand, exercise did not change the NPY and POMC mRNA levels in control mice (Figure 1G and H). Exercise did not change the total body weight and epididymal fat pad weight in WT and ob/ob mice (Figure 1I and J). In addition, we observed that the exercise protocols did not change the triglycerides and free fatty acid levels but reduced the insulin levels in WT and ob/ob mice (Table 2). The lactate production was similar between lean and obese mice during the respective exercise protocols (Table 2). These exercise protocols did not evoke any significant stressful effect in these animals, as demonstrated by urinary corticosterone levels (Table 2). Thus, our data demonstrate that exercise modulates hypothalamic neuropeptides (NPY and POMC) and suppresses food intake in obese, but not in lean, rodents without changing the adipose tissue content and corticosterone levels.
Next, we evaluated whether exercise modulates insulin signaling in the hypothalamus. Western blot analysis revealed that IRβ, IRS-1, IRS-2, Akt, and FOXO1 phosphorylation were similar between the groups (Figure 2A and B). Although exercise did not change the basal levels of insulin signaling, we next performed intrahypothalamic insulin (200 mU) or its vehicle injection to evaluated food intake and insulin sensitivity after the SW Exe protocol. Overnutrition markedly reduced the ability of intrahypothalamic insulin infusion to reduce food intake, when compared to chow-fed animals; however, exercise restored the central effects of insulin on reduced food intake (Figure 2C). Using Western blotting analysis, we determined the effects of exercise on the insulin sensitivity in hypothalamic tissue. The high-fat diet impaired insulin-induced tyrosine phosphorylation of insulin receptor β (IRβ), insulin receptor substrate-1 (IRS-1), and IRS-2 in the hypothalamus (Figure 2D). Similar results were observed for the serine phosphorylation of Akt and FOXO1 (Figure 2D). Physical activity was able to restore insulin-induced hypothalamic IRβ, IRS-1, and IRS-2 tyrosine phosphorylation and insulin-induced hypothalamic Akt and FOXO1 serine phosphorylation in DIO rats (Figure 2D). Subcellular fraction of hypothalamic extract was then performed to evaluate the nuclear FOXO1 expression. Intrahypothalamic infusion of insulin reduced the nuclear FOXO1 expression in control rats, but insulin failed to reduce the nuclear FOXO1 expression in rats after overnutrition (Figure 2E). After exercise, insulin reduced the nuclear FOXO1 expression in neuronal cells of obese animals (52%), when compared to DIO at rest (Figure 2E).
We then explored the effects of exercise on hypothalamic leptin action, monitoring Janus Kinase-2 (Jak-2) and STAT-3 tyrosine phosphorylation. Exercise did not change the Jak-2 and STAT-3 phosphorylation in lean animals; however, overnutrition reduced Jak-2 and STAT-3 phosphorylation when compared to lean animals. Interestingly, physical activity was able to increase the neuronal Jak-2 and STAT-3 tyrosine phosphorylation in obese animals (Figure 2F and G). In addition we investigated the effects of exercise on leptin sensitivity. Intrahypothalamic infusion of leptin markedly reduced the 12-h total energy intake in control rats; however, the anorexigenic effects of leptin were attenuated in obese rats. In contrast, exercise restored the central effects of leptin on reduced food intake (Figure 2H). We noted that leptin modestly promoted the hypothalamic tyrosine phosphorylation of Jak-2, IRS-1, IRS-2, and STAT-3 after high-fat diet treatment. Conversely, exercise restored leptin-induced hypothalamic Jak-2, IRS-1, IRS-2, and STAT-3 tyrosine phosphorylation in obese animals (Figure 2I).
We also evaluated nuclear STAT3 expression after intrahypothalamic leptin infusion. After overnutrition, leptin failed to increase the expression of nuclear STAT3 in the hypothalamus. On the other hand, exercise increased the ability of leptin to increase the nuclear expression of STAT3 (48%) in the hypothalamus of obese animals (Figure 2J).
Recently, IL-6 was reported as the first myokine that is produced and released by contracting skeletal muscle fibers, exerting its effects on other organs of the body [20], including the hypothalamus [18],[21]. Thus, we evaluated the central role of IL-6 in the control of food intake. Firstly, the serum level of IL-6 was observed to be slightly up-regulated after high-fat diet treatment and was dramatically increased immediately after SW Exe and TR Exe, but we observed that, in exercised obese animals, the serum levels of IL-6 were higher when compared to exercised lean ones (Figure S1A). Similar results were found when IL-6 protein expression in the hypothalamic tissue was evaluated (Figure S1B). To investigate whether neuronal cells were producing IL-6 in response to exercise, we performed real time PCR to evaluate IL-6 mRNA levels in the hypothalamic tissue. IL-6 mRNA levels were slightly up-regulated after the high-fat diet treatment and were increased by about 53% and 64% immediately after physical activity in lean and obese rats, respectively (Figure 3A). Thus, these data demonstrate that exercise increases the serum and hypothalamic levels of IL-6.
Next, we sought to determine whether exercise requires IL-6 to mediate the anti-hyperphagic response. First we showed that the infusion of recombinant IL-6 into the third ventricle of obese animals under resting conditions reduced the food intake in a dose-dependent manner (Figure 3B) and restored the anorexigenic effects of insulin and leptin (Figure S2A and B). Although we used recombinant IL-6 to mimic the effects of exercise, in obese rats, the dose of recombinant IL-6 used (200 ng) is relatively high and this pharmacological approach does not reflect the same physiological conditions observed after exercise. Thus, we hypothesized that if exercise requires hypothalamic IL-6 activity to reduce food intake, inhibiting the hypothalamic effects of this cytokine, under physiological conditions, should diminish the appetite suppressive action mediated by exercise. To address this hypothesis, we developed an experimental strategy aimed at antagonizing the central action of IL-6 in the presence of a systemic elevation in plasma IL-6 concentration after physical activity. For this, we injected an anti-IL-6 antibody into the third-hypothalamic ventricle in obese animals at 15 min before the exercise protocol. Interestingly, pretreatment with anti-IL-6 antibody blocked the anorexigenic effects of insulin and leptin in exercised DIO rats (Figure 3C and D).
We then explored the mechanism by which IL-6 improves insulin and leptin signaling in the hypothalamus, evaluating the pro-inflammatory pathway. Firstly, we demonstrated that acute exercise did not change the expression or activity of the proteins involved in inflammatory signaling and in an ER stress in the hypothalamus of lean rats, when compared to control animals at rest (Figure 3E). However, high-fat diet consumption induced the aberrant activation of the NF-κB pathway components in the hypothalamic tissue, increasing the TLR4 expression, IKKβ serine phosphorylation, and the IκBα degradation (Figure 3F–H). We also monitored PERK phosphorylation and CHOP protein expression in the hypothalamus to evaluate ER stress. High-fat diet also activated ER stress, increasing PERK phosphorylation and CHOP protein expression in the hypothalamus (Figure 3I and J). In addition, high-fat diet increased IRS-1 serine 307 phosphorylation (Figure 3K). Neither acute exercise nor the single injection of recombinant IL-6 was able to reduce the TLR4 expression in the hypothalamic tissue of obese animals (Figure 3F). On the other hand, exercise and the intrahypothalamic injection of recombinant IL-6, in obese rats at rest, markedly reduced the hypothalamic IKKβ serine phosphorylation (∼60%) and prevented IκBα degradation in obese animals (Figure 3G and H). The recombinant IL-6 injection and exercise reduced PERK phosphorylation by about 60% and CHOP protein expression by about 45% (Figure 3I and J) and IRS-1 serine phosphorylation by about 60% (Figure 3K) in the hypothalamic tissue of hyperphagic animals. In addition, recombinant IL-6 and exercise restored insulin-induced Akt and leptin-induced and STAT-3 phosphorylation in the hypothalamus of obese animals (Figure S3A and B). Interestingly, our results show that the intrahypothalamic injection of anti-IL-6 antibody before the exercise protocol attenuated the ability of exercise to reduce the IKKβ/IκBα pathway, ER stress, and IRS1 serine phosphorylation in the hypothalamus (Figure 3G–K). The pretreatment with anti-IL6 antibody also blocked insulin-induced Akt and leptin-induced and STAT-3 phosphorylation, mediated by exercise in the hypothalamus of obese animals (Figure S3A and B).
Immunohistochemistry with an anti-IL-6 Receptor (IL-6R)-specific antibody showed that IL-6R is expressed in a majority of neurons in the arcuate nucleus (Figure 4A). These data were confirmed when we quantified the positive cells in arcuate (Arc), dorsomedial and ventromedial (DMH/VMH), paraventricular (PVN), and lateral (LH) nuclei of hypothalamus (Figure 4B). The in situ hybridization experiment revealed that IL-6R is expressed in both anorexigenic and orexigenic neurons of rats (Figure 4C).
Since IL-6R is expressed in a majority of neurons in the arcuate nucleus, we dissected this specific hypothalamic region to evaluate the modulation of the neuropeptides in response to exercise in lean and obese rats. We found that exercise did not change the POMC, NPY, and AgRP mRNA in the arcuate nucleus of lean rats but increased the POMC and reduced the NPY mRNA levels in the arcuate nucleus of obese animals (Figure 4D).
Double-staining confocal microscopy showed that most neurons expressing IL-6R in the arcuate nucleus were shown to possess IKKβ, PERK, and IRS-1 in obese rats, showing a possible interaction between these molecules (Figure 4E).
To further support data indicating that IL-6 may modulate ER stress, we performed an acute intrahypothalamic injection of an ER stress inducer, thapsigargin (TG), in lean rats. Acute intrahypothalamic infusion of thapsigargin did not change food intake in lean animals by itself (Figure 5A). However, our results revealed that intrahypothalamic infusion of thapsigargin blocked the anorexigenic effects mediated by insulin and leptin in lean rats and that the injection of recombinant IL-6 and exercise restored the suppressive appetite action of insulin and leptin (Figure 5B and C). In addition, the infusion of anti-IL6 antibody blocked the improvement in insulin and leptin action mediated by exercise (Figure 5B and C).
In accordance with previous studies [7],[14],[22], we observed that thapsigargin markedly activated inflammatory signaling and ER stress in lean rats, as reflected by increased levels of hypothalamic IKKβ and PERK phosphorylation, respectively (Figure 5D and E), and induced central insulin and leptin resistance, increasing IRS-1 serine phosphorylation (Figure 5F) and reducing insulin-induced Akt serine phosphorylation and leptin-induced STAT-3 tyrosine phosphorylation (Figure 5G and H). Intrahypothalamic infusion of recombinant IL-6 and physical activity were sufficient to reverse all these phenomena (Figure 5D–H). Conversely, the infusion of intrahypothalamic anti-IL6 antibody before exercise protocol blocked these effects mediated by exercise (Figure 5D–H). There were no differences in the basal levels of Akt and STAT-3 phosphorylation between the groups (Figure 5I).
Low dose TNF-α has been reported to induce insulin and leptin resistance in the hypothalamus [23]. We injected a low dose of TNF-α into the hypothalamus of lean rats to investigate the effects of IL-6 on low-grade inflammation. First we observed that acute intrahypothalamic infusion of TNF-α did not change the food consumption in lean rats (unpublished data); however, TNF-α infusion blocked the anorexigenic actions of insulin and leptin in these animals (Figure S4A and B). The anorexigenic actions of these hormones were restored with the central infusion of recombinant IL-6 or after exercise in lean rats injected with TNF-α. In addition, the pretreatment with anti-IL6 antibody into the third ventricle blocked the improvement in insulin and leptin action mediated by exercise (Figure S4A and B).
The single injection of TNF-α also induced IKKβ serine, PERK threonine, and IRS-1 serine phosphorylation and reduced insulin-induced Akt serine phosphorylation and leptin- induced STAT-3 tyrosine phosphorylation in the hypothalamus of lean rats (Figure S4C–G). Intrahypothalamic infusion of recombinant IL-6 and physical activity were also sufficient to reverse all these phenomena. On the other hand, the central infusion of anti-IL6 antibody before the exercise protocol blocked the effects of physical activity (Figure S4C–G). There were no differences in the basal levels of Akt and STAT-3 phosphorylation between the groups (Figure S4H).
Next, we sought to determine how IL-6 reduces the inflammatory response and ER stress in the hypothalamus after exercise. Several studies have reported that exercise-induced increases in plasma IL-6 levels are followed by increased circulating levels of well-known anti-inflammatory cytokines such as the IL-1ra and IL-10 [24],[25]. We found that the IL-1ra protein level was not changed in the hypothalamus after chronic overnutrition or after acute exercise protocols (Figure 6A); however, IL-10 protein expression was slightly increased in the hypothalamus in obese animals; both of the exercise protocols increased IL-10 expression in a similar fashion, but the induction of IL-10 expression, mediated by exercise, was higher in the hypothalamus of obese when compared to exercised lean animals (Figure 6B). The increase in hypothalamic IL-10 levels mediated by physical activity was confirmed by real time PCR assay (Figure 6C).
We then investigated whether IL-10 reduced the energy intake in rodents. Intrahypothalamic injection of recombinant IL-10 reduced food intake in obese animals in a dose-dependent manner (Figure 6D). To explore whether IL-6 requires IL-10 expression to improve insulin and leptin action in the hypothalamus, we used an IL-10 antisense oligonucleotide (ASO IL-10) in the hypothalamus of obese rats to keep the expression levels of IL-10 low, even in the presence of high levels of IL-6 in the hypothalamus. Three days after ASO IL-10 treatment, IL-10 protein expression was reduced by about 75% in the hypothalamus of obese animals (Figure 6E). Thereafter, exercise and recombinant IL-6 infusion failed to improve the anorexigenic effects of insulin and leptin in obese animals treated with ASO IL-10 (Figure 6F and G).
IL-10 is a pleiotropic cytokine that controls inflammatory processes by suppressing the production of proinflammatory cytokines and blocking IKK/NF-κB signaling and ER stress [26],[27]. Thus, we investigated whether exercise and IL-6 requires IL-10 expression to reduce IKKβ activation and ER stress in the hypothalamus of obese animals. As demonstrated above, recombinant IL-6 infusion and exercise reduced IKKβ, PERK, and IRS-1Ser307 phosphorylation (Figure 3G, I, and K) and restored insulin and leptin signaling in the hypothalamus of obese animals (Figure S3), but the intrahypothalamic IL-10 ASO treatment abolished all these parameters mediated by recombinant IL-6 and exercise (Figure 6H–L). Conversely, the injection of recombinant IL-10 in the hypothalamus of obese animals at rest markedly reduced IKKβ, PERK, and IRS-1Ser307 phosphorylation and increased insulin-induced Akt and leptin-induced STAT-3 phosphorylation in the hypothalamic tissue of obese rats (Figure 6H–L). There were no differences in the basal levels of Akt (Figure 6M). However, STAT3 tyrosine phosphorylation was reduced in the hypothalamus of obese rats, but neither exercise nor IL-6 intrahypothalamic injection was able to increase the STAT-3 phosphorylation after IL-10 ASO treatment (Figure 6N).
Several studies showed that Toll-like receptor inactivation results in an attenuation of the secretion of several cytokines. TLR4- and MyD88-deficient mice sustain significantly lower levels of serum cytokines such as IL-1β, IL-6, TNFα, and IL-10 after different pro-inflammatory stimuli [28]–[30]. Since TLR4 mediates IL-6 transcriptional responses in myocytes and in the skeletal muscle of C3H/HeJ mice [31], we investigated whether exercise restores insulin and leptin signaling in the hypothalamus of TLR4-deficient mice (C3H/HeJ) injected with thapsigargin (TG, an endoplasmic reticulum stress inducer).
In contrast to WT mice, TLR4-deficient mice were found to sustain significantly lower hypothalamic levels of IL-6 (Figure 7A) and IL-10 (Figure 7B) after exercise. The food consumption was similar between C3H/HeN and C3H/HeJ under basal conditions, and acutely, thapsigargin alone did not affect the food intake in these mice (unpublished data); however, the intrahypothalamic administration of TG impaired the anorexigenic effects of insulin and leptin in WT (C3H/HeN) and in TLR4-deficient mice; while physical activity restored the appetite suppressive actions of insulin and leptin in WT but not in TLR4-deficient mice (Figure 7C and D). Furthermore, the intrahypothalamic injection of either recombinant IL-6 or IL-10 restored the anorexigenic actions of insulin and leptin in both WT and TLR4-deficient mice injected with TG (Figure 7C and D). We also observed that the intrahypothalamic infusion of recombinant IL-6 was able to increase the IL-10 protein expression in the hypothalamus of WT and TLR4-deficient mice (Figure 7E). Moreover, exercise failed to reduce inflammation and ER stress and failed to improve insulin and leptin sensitivity in the hypothalamus of TLR4-deficient mice injected with TG (Figure 7F–J). On the other hand, the intrahypothalamic injection of recombinant IL-6 or IL-10 reduced IKKβ, PERK, and IRS-1Ser307 phosphorylation and restored insulin and leptin signaling in the hypothalamus of TLR4-deficient mice injected with TG (Figure 7F–J). There were no differences in the basal levels of Akt and STAT-3 phosphorylation between the groups (unpublished data). The in situ hybridization experiment revealed that IL-10R is expressed in NPY, POMC, and AgRP neurons of rats (Figure 7K). Finally, immunohistochemistry with anti-IL-6R and anti-IL-10 Receptor (IL-10R)-specific antibodies revealed that IL-6R and IL-10R are expressed in the same specific neuronal subtypes in the arcuate nucleus (Figure 7L).
We then investigated the effects of chronic SW Exe on food intake and body weight in lean and obese rats. As observed in acute exercise, the chronic exercise protocol did not change the food consumption in lean animals; however, we observed that the food intake was reduced in obese animals after onset of the chronic exercise protocol, for 3 d, but thereafter, the food intake returned to basal levels on the sixth day and was maintained similar to that of obese rats at rest (Figure 8A). Exercised obese animals showed a significant reduction of the total body weight between the third and the sixth days, but this phenomenon was not observed in control animals (Figure 8B). We also evaluated the weight gain by analyzing the variation of the body weight between the 1st and 24th days. We observed a slight weight gain in control animals at rest, but the chronic exercise protocol did not attenuate the weight gain in lean animals (Figure 8C). On the other hand, overnutrition induced a great weight gain in the group under resting conditions, while chronic exercise attenuated the weight gain in obese animals (Figure 8C). We did not observe a statistical difference in the absolute values of the epididymal fat mass between the exercised obese animals and the obese animals at rest at the end of chronic exercise protocol (Figure 8D).
Chronic overnutrition increased serum insulin, leptin, triglycerides, and free fatty acid levels, compared to age-matched controls; however, chronic exercise reduced serum insulin, triglycerides, and free fatty acid levels in obese animals (Table 3). To determine whether lean and obese rodents were swimming or running in the same fashion, we evaluated lactate production every 15 min during the SW Exe. We did not find any difference in the lactate production between lean and obese rats. Table 3 depicts the final values obtained in this test. We also determined that this exercise protocol did not change the corticosterone levels in lean and obese animals 3 d after the onset of this exercise protocol (Table 3).
We also evaluated IL-6 and IL-10 mRNA levels in the hypothalamic tissue during the chronic exercise protocol. Interestingly, we observed that the levels of IL-6 mRNA in the hypothalamus were higher on the first day of exercise, when compared to the 15th and 24th days of exercise; this phenomenon was observed in lean and obese exercised rats (Figure 8E). Similar results were found when we analyzed the levels of IL-10 mRNA during chronic exercise (Figure 8F). Finally, the chronic exercise protocol reduced IKKβ phosphorylation and increased IκBα expression in the hypothalamus of obese rats; however, this anti-inflammatory response was more evident on the first day of exercise (Figure 8G). Similar results were found when we analyzed the ER stress markers, such as PERK phosphorylation and CHOP expression (Figure 8H).
Physical activity is a cornerstone in the prevention of obesity and related diseases. Although the energy expenditure aspects of such exercise may contribute to the effects of weight loss, it has been suggested that physical exercise may also contribute to negative energy balance by altering appetite and reducing food intake in rodents [21],[32] and humans [33],[34]. Our study shows that acute exercise per se did not evoke any meaningful effect, in terms of food intake in lean animals, but interestingly, it was crucial for suppressing hyperphagia mediated by overnutrition, reducing hypothalamic IKKβ/NF-κB activation and ER stress, thus improving insulin and leptin action in an IL-6- and IL-10-dependent manner (Figure 9).
In the absence of obesity, exercise does not affect food behavior, as the anorexigenic or orexigenic pathways remain unchanged in rats. Several experimental studies have demonstrated that physical activity does not activate anorexigenic pathways, such as PI3-K or mTOR/p70S6K [18],[21], and does not inhibit the orexigenic pathways, such as AMPK signaling in the hypothalamus of control rodents [35]. On the other hand, the present study provides substantial evidence that physical activity could help to reorganize the set point of nutritional balance and, therefore, aid in counteracting the energy imbalance induced by overnutrition-related obesity. These data are in accordance with Park and colleagues [36], who showed that exercise improved insulin and leptin signaling, increased STAT3, and reduced AMPK phosphorylation in the cerebral cortex and hypothalamus of diabetic rats, contributing to the regulation of body weight and glucose homeostasis. These data demonstrate that exercise increases the anorexigenic pathways and attenuates the orexigenic signals, only in obese and diabetic animals, changing the anorexigenic and orexigenic signaling pathways in the hypothalamus. We also reported that physical activity reduced the hyperphagic response by reducing NPY mRNA and increasing POMC mRNA predominantly in the arcuate nucleus of obese animals. It is important to emphasize that acute exercise did not change the total body weight or epididymal fat pad weight, showing that physical activity can induce the anorexigenic response in the hypothalamus, independently of the body weight change. Our data showed that the reduction on food intake observed in obese animals after both exercise protocols was not related to stress as demonstrated by costicosterone levels. In opposite fashion, it has been demonstrated that NPY mediates stress-induced exacerbation of diet-induced obesity and metabolic syndrome after different stressor agents such as exposure to cold water or aggression in mice [37]. Thus, we hypothesized that some factors, produced during the exercise session, could be involved in this anorexigenic response.
Skeletal muscle is an endocrine organ that, upon contraction, stimulates the production and release of cytokines, also called myokines, which can influence metabolism and modify cytokine production in tissue and organs. IL-6 is the first cytokine present in the circulation during exercise [17]. IL-6 can elicit proinflammatory or anti-inflammatory effects, depending on the in vivo environmental circumstances. Although IL-6 has been associated with low-grade inflammation and insulin resistance, it has been demonstrated that acute IL-6 treatment enhances insulin-stimulated glucose disposal in humans [38].
Centrally acting IL-6 appears to play a role in the regulation of appetite, energy expenditure, and body composition. Wallenius and colleagues elegantly showed that long-term peripheral IL-6 treatment to IL6−/− mice caused a decrease in body weight. In addition to increasing energy expenditure, IL-6 may prevent obesity by inhibiting feeding as obese IL-6−/− mice had increased absolute food intake [39]. In accordance with these data, mice fed on a high-fat diet with sustained circulating human IL-6 secreted predominantly from brain and lung (hIL6tg) had low leptin concentrations, consumed less food, and expended more energy than wild-type mice [40]. In addition, the intercrossing of hIL6tg and ob/ob mice increased the leptin sensitivity in these mice, when compared to ob/ob mice [40]. Recently, we demonstrated that exercise requires IL-6 to increase hypothalamic insulin and leptin sensitivity [18] and increase the effects of leptin on the AMPK/mTOR pathway in the hypothalamus of rodents [21]. Furthermore, IL-6 is also released from the brain during prolonged exercise in humans [41]. In the present study, we showed that the increment of IL-6 expression in the hypothalamus was crucial to exercise for reducing the inflammation and ER stress activation induced by overnutrition. However, these effects, promoted by exercise, were not observed when we used an intrahypothalamic infusion of anti-IL-6 antibody before the exercise protocol. In addition, the infusion of recombinant IL-6 into the third hypothalamic ventricle reduced the energy intake in obese animals under resting conditions, in a dose-dependent manner, and reduced hypothalamic IKKβ and ER stress activation.
In another approach, we used an ER stress inducer in lean rats to evaluate the effects of exercise/IL-6 on hypothalamic ER stress. We demonstrated that acute thapsigargin injection increased IKKβ and PERK phosphorylation and reduced insulin and leptin action in the hypothalamus and that exercise and the infusion of recombinant IL-6 were able to reduce thapsigargin-induced inflammation, ER stress, and insulin and leptin resistance, whereas the IL-6 antibody pretreatment reversed the effects of exercise. Although thapsigargin increased the hypothalamic IKKβ and PERK phosphorylation, we did not observe any difference in the basal levels of Akt serine 473 and STAT3 tyrosine 705 phosphorylation and in food intake in rats injected with thapsigargin alone. These data are in accordance with a previous study that reported that the ER-stress inhibitor, tauroursodeoxycholic acid (TUDCA), acutely reduced the hypothalamic PERK phosphorylation and NF-kB activation but did not change the food intake in mice fed on a high-fat diet [7]. Thus, our data demonstrate that IL-6 plays an important role in the control of the ER stress effects in the hypothalamus of rats.
All these results are significant, since IKKβ and ER stress activation were strongly associated with insulin and leptin resistance in the hypothalamic tissue. Although we showed a consistent anti-inflammatory effect, mediated by IL-6, in the hypothalamus, we cannot exclude the possibility that IL-6 acts directly as an anorexigenic factor.
Although our findings clearly show that IL-6 diminished hypothalamic IKKβ and ER stress activation and restored the central insulin and leptin action in an animal model of obesity, the question remains as to how IL-6 promotes these events in the hypothalamus. Following exercise, the high circulating levels of IL-6 are followed by an increase in two anti-inflammatory molecules, IL-1ra and IL-10 [25]. Therefore, IL-6 induces an anti-inflammatory environment by inducing the production of IL-1ra and IL-10. In our study, we found that exercise increased the hypothalamic levels of IL-10 but did not change IL-1ra expression in this tissue. Thus, we showed that the anti-inflammatory response mediated by IL-6 involves the increase of IL-10 expression in the hypothalamus.
IL-10 is an important immunoregulatory cytokine with multiple biological effects. In the cytoplasm, it has been demonstrated that IL-10 blocks NF-κB activity at two levels: suppressing IKK activity and NF-κB DNA binding activity [26]. Moreover, IL-10 reduced ER stress in intestinal eptithelial cells, whereas IL-10−/− mice demonstrated that the expression of the ER stress response protein grp-78/BiP was increased in intestinal eptithelial cells under conditions of chronic inflammation [27].
In the CNS, the anti-inflammatory role of IL-10 has been extensively studied in experimental autoimmune encephalomyelitis, an animal model of human multiple sclerosis. The increase in IL-10 expression in the CNS during recovery from brain inflammation and the inability of IL-10 null mice to recover from acute CNS inflammation suggests that the presence of IL-10 within this target organ is required for disease remission [42],[43]. However, the role of hypothalamic IL-10 in the control of low-grade inflammation generated during obesity was unknown. Here, we discovered that intrahypothalamic infusion of recombinant IL-10 blocked IKK/NF-κB signaling and ER stress and restored Akt and STAT3 phosphorylation, promoting a re-balance in the energy intake in obese animals. On the other hand, the selective decrease in IL-10 expression in discrete hypothalamic nuclei of obese animals mediated by ASO treatment blunted the effects of both exercise and the intrahypothalamic infusion of recombinant IL-6 in the restoration of central insulin and leptin actions. In addition, we demonstrated that in mice that sustained significantly lower hypothalamic levels of IL-6 and IL-10 after exercise (C3H/HeJ), there was no reduction in pharmacological ER stress activation, in contrast to WT mice. These data are intriguing as IL-10 represents an important cytokine that may reduce both inflammation and ER stress in the hypothalamus. Thus, the modulation of hypothalamic IL-10 expression could be considered the direct target of exercise/IL-6 and constitutes a promising alternative to reduce hypothalamic inflammation and ER stress related to obesity.
The decrease in food intake induced by IL-10 in obese rats is not in accordance with the effects observed in IL-10 KO. It has been reported that mice with combined deficiency of leptin and IL-10 gain less body weight than mice lacking leptin only [44]. However, these discrepancies may be a consequence of methodological differences related to physiological versus genetic approaches and acute versus chronic situation investigated, and most important it may be consequence of IL-10 effects in the regulation of energy expenditure, likewise observed in mice lacking TNF-α receptor [45]; thus, the role of IL-10 in the control of food intake and energy expenditure deserves further exploration.
The long-term reversal effects on body composition, mediated by exercise alone, are controversial. It should be acknowledged that it is often difficult to find long-term reversal effects on body fat in both experimental animals and humans by exercise alone without restrained diet [46]. In the chronic experiments, we observed that the obese animals lost weight during the same period in which a reduction in food intake was observed. After this period, no significant difference was observed in the body weight of exercised animals, although the obese animals presented a significant improvement in metabolic parameters after the chronic exercise protocol.
Since IKKβ/NF-κB inhibition in the CNS represents a potential target therapy to combat obesity and most anti-inflammatory therapies have limited direct effects on IKKβ/NF-κB and a limited capacity for concentration in the CNS, our study provides substantial evidence that physical activity could help to reorganize the set point of nutritional balance and therefore aid in counteracting the energy imbalance induced by overnutrition through the anti-inflammatory response in hypothalamic neurons. Hence, IL-6 and IL-10 are important physiological contributors to the central insulin and leptin action mediated by physical activity, linking it to hypothalamic ER stress and inflammation.
Protein A-Sepharose 6 MB and Nitrocellulose paper (Hybond ECL, 0.45 µm) were from Amersham Pharmacia Biotech United Kingdom Ltd. (Buckinghamshire, United Kingdom). Ketamin was from Parke-Davis (São Paulo, SP, Brazil) and diazepam and thiopethal were from Cristália (Itapira, SP, Brazil). Anti-phospho-JAK2 (rabbit polyclonal, AB3805) antibody was from Upstate Biotechnology (Charlottesville, VA, USA). Anti-JAK2 (rabbit polyclonal, SC-278), anti-STAT3 (rabbit polyclonal, SC-483), anti-phospho-IRβ (rabbit polyclonal, SC-25103), anti-IRβ (rabbit polyclonal, SC-711), anti-phospho-IRS-1 (rabbit polyclonal, SC-17199), anti-IRS-1 (rabbit polyclonal, SC-559), anti-IRS-2 (rabbit polyclonal, SC-1556), anti-phosphotyrosine (mouse monoclonal, SC-508), anti-Foxo1 (rabbit polyclonal, SC-11350), anti-IL-1ra (goat polyclonal, SC-8481), anti-TNF-α (rabbit polyclonal, SC-8301), anti-IKKβ (goat polyclonal, SC-34673), anti-PERK (rabbit polyclonal, SC-13073), anti-phospho-PERK (rabbit polyclonal, SC-32577), anti-CHOP (GADD 153) (rabbit polyclonal, SC-575), anti-IL-10 (goat polyclonal, SC-1783), and anti-IL-6 (rabbit polyclonal, SC-7920) antibodies were from Santa Cruz Biotechnology, Inc. Anti-phospho-STAT3 (rabbit polyclonal, #9131), anti-phospho-Akt (rabbit polyclonal, #9271), anti-phospho-Foxo1 (rabbit polyclonal, #9461), anti-beta tubulin (rabbit polyclonal, #2146), anti-phospho-IKKα/β (rabbit polyclonal, #2687), anti-IκBα (rabbit polyclonal, #9242), anti-TLR4 (rabbit polyclonal, #2219), anti-phospho-IRS-1 307 (rabbit polyclonal, #2381), and anti-Akt (rabbit polyclonal, #9272) were from Cell Signalling Technology (Beverly, MA, USA). Leptin, thapsigargin, and recombinant IL-6 and -10 were from Calbiochem (San Diego, CA, USA). Routine reagents were purchased from Sigma Chemical Co. (St. Louis, MO) unless otherwise specified.
Blood was collected from the cava vein 15 min after the exercise protocols. Plasma was separated by centrifugation (1,100 g) for 15 min at 4 °C and stored at −80 °C until assay. RIA was employed to measure serum insulin. Leptin and IL-6 concentrations were determined using a commercially available Enzyme Linked Immunosorbent Assay (ELISA) kit (Crystal Chem Inc., Chicago, IL). Blood lactate was measured using Accutrend Plus equipment (Roche); sample blood was obtained from the tails every 15 min during the exercise protocols. Serum cholesterol and triglycerides were measured in control and exercised animals after 8 h of fasting using Accutrend Plus equipment (Roche). Serum free fatty acids (FFA) levels were analyzed in rats using the NEFA-kit-U (Wako Chemical GmBH, Neuss, Germany).
Corticosterone levels were determined using urine samples obtained from rats and mice using specific metabolic cage during 24 h after the exercise protocols. The corticosterone level was determined using an EIA kit from Cayman chemical (Ann Arbor, MI).
Male 4-wk-old Wistar rats were obtained from the University of Campinas Breeding Center. The investigation was approved by the ethics committee and followed the University guidelines for the use of animals in experimental studies and experiments conform to the Guide for the Care and Use of Laboratory Animals, published by the U.S. National Institutes of Health (NIH publication no. 85-23 revised 1996). The animals were maintained on 12h∶12h artificial light-dark cycles and housed in individual cages. Rats were randomly divided into two groups: control, fed on standard rodent chow (3,948 kcal.Kg−1), and DIO, fed a fat-rich chow (5,358 kcal.Kg−1) ad libitum for 3 mo. This diet composition has been previously used [47].
Male (10-wk-old) ob/ob mice and their respective controls C57BL/6J background were obtained from The Jackson Laboratory and provided by the University of São Paulo. The mice were bred under specific pathogen-free conditions at the Central Breeding Center of University of Campinas.
Male C3H/HeJ (10-wk-old) mice and their respective controls C3H/HeN were obtained from The Jackson Laboratory and provided by the University of São Paulo. The mice were bred under specific pathogen-free conditions at the Central Breeding Center of the University of Campinas.
The animals were stereotaxically instrumented under intraperitoneal injection of a mix of ketamin (10 mg) and diazepam (0.07 mg) (0.2 ml/100 g body weight) with a chronic 26-gauge stainless steel indwelling guide cannula aseptically placed into the third ventricle at the midline coordinates of 0.5 mm posterior to the bregma and 8.5 mm below the surface of the skull of rats and 1.8 mm posterior to the bregma and 5.0 mm below the surface of the skull of mice.
Animals were acclimated to swimming for 2 d (10 min per day). Water temperature was maintained at 34–35 °C. Rats performed two 3-h exercise bouts, separated by one 45-min rest period. The rats swam in groups of three in plastic barrels of 45 cm in diameter that were filled to a depth of 50 cm. This protocol was conducted between 11:00 a.m. and 6:00 p.m., as previously described [48], and mice performed four 30-min exercise bouts, separated by one 5-min rest period. The mice swam in groups of four in plastic barrels of 40 cm in diameter that were filled to a depth of 20 cm. This protocol was conducted between 3:00 p.m. and 6:00 p.m. Both exercise protocols finished at 6:00 p.m. for evaluation of food intake and analysis of hypothalamic tissue.
The chronic exercise protocol consisted of daily swimming sessions (1 h/d, 5 d/wk, for 4 wk) with an overload (2.0% of the body weight). The hypothalamic tissues and the metabolic parameter were evaluated 36 h after the last exercise session. Rats also performed a single bout of treadmill (Insight LTDA - Ribeirão Preto, SP) running (60 min, speed of 10–15 m/min at a 5% incline) and mice performed a single bout of treadmill running (90 min, speed of 7–10 m/min at a 5% incline).
Rats or mice were deprived of food for 2 h with free access to water and received 3 µl of bolus injection into the third ventricle, as follows:
Intrahypothalamic infusions were performed between 5:00 and 6:00 p.m. Thereafter standard chow or high-fat diet was given and food intake was determined by measuring the difference between the weight of chow given and the weight of chow at the end of a 12-h period. Similar studies were carried out in animals after exercise.
After exercise and/or i.c.v. treatments, the animals were anaesthetized, and the hypothalamus was quickly removed, minced coarsely, and homogenized immediately in a freshly prepared ice-cold buffer (1% Triton X-100, 100 mmol/l Tris pH 7.4, 100 mmol/l sodium pyrophosphate, 100 mmol/l sodium fluoride, 10 mmol/l EDTA, 10 mmol/l sodium vanadate, 2 mmol/l phenyl methylsulphonyl fluoride, and 0.1 mg aprotinin) suitable for preserving phosphorylation states of enzymes, and Western blot was performed, as previously described [1].
Foxo1 and STAT-3 nuclear expression were obtained as described [49]. Fragments of hypothalamic tissue from untreated rats or rats treated with insulin or leptin were obtained 30 min after insulin or leptin infusion and were minced and homogenized in 2 vol. of STE buffer (0.32 M sucrose, 20 mM Tris–HCl (pH 7.4), 2 mM EDTA, 1 mM DTT, 100 mM sodium fluoride, 100 mM sodium pyrophosphate, 10 mM sodium orthovanadate, 1 mM PMSF, and 0.1 mg aprotinin/ml) at 4 °C with a Polytron homogenizer. The homogenates were centrifuged (1,000×g, 25 min, 4 °C) to obtain pellets. The pellet was washed once and suspended in STE buffer (nuclear fraction). The nuclear fraction was solubilized in Triton buffer [1% (v/v) Triton X-100/150 mM NaCl/10 mM Tris/HCl (pH 7.4)/1 mM EGTA/1 mM EDTA/0.2 mM sodium orthovanadate/20 µM leupeptin A/0.2 mM PMSF/50 mM NaF/0.4 nM microcystin LR]. The fraction was centrifuged (15,000 g, 30 min, 4 °C), and the supernatant (nuclear extract) was stored at −80 °C.
Paraformaldehyde-fixed hypothalami were sectioned (5 µm). The sections were obtained from the hypothalami of six rats per group in the same localization (antero-posterior = −1.78 from bregma) and used in regular single- or double-immunofluorescence staining using DAPI, anti-IL6 receptor alpha (rabbit IgG, SC-13947), anti-IL-10 receptor (rabbit IgG, SC-987), anti-IKKβ (goat IgG, SC-34673), anti-PERK (rabbit IgG, SC-32577), anti-POMC (rabbit IgG, FL-267), and rabbit anti-IRS-1 (rabbit IgG, SC-559) (1∶200; Santa Cruz Biotechnology) antibodies. After incubation with the primary antibody, sections were washed and incubated with specific biotinylated anti-rabbit or anti-goat secondary antibodies (1∶150 dilution) for 2 h at room temperature, followed by incubation with Streptoavidin reagent (containing avidin-conjugated peroxidase) and color reaction using the DAB substrate kit (Vector Laboratories, Burlingame, CA, USA), according to recommendations of the manufacturer. Analysis and photodocumentation of results were performed using a LSM 510 laser confocal microscope (Zeiss, Jena, Germany). The anatomical correlations were made according to the landmarks given in a stereotaxic atlas [50]. The frequency of positive cells was determined in 100 randomly counted cells using Analysis software (Version 2.4).
Hypothalamic total RNA was extracted using Trizol reagent (Life Technologies, Gaithersburg, MD, USA), according to the manufacturer's recommendations. Total RNA was rendered genomic DNA free by digestion with Rnase-free Dnase (RQ1, Promega, Madison, WI, USA). Rats were deprived of food for 9 h after for real time PCR analysis. Real time PCR and mRNA isolation were performed using a commercial kit, as follows: IL-6: Rn00561420_m1 IL-10: Rn00563409_m1, POMC: Rn00595020_m1, NPY: Rn00561681_m1, AgRP: Rn01431703_g1, GAPD, #4352338E, for rat and RPS-29 (NCBI: NM012876), sense: 5′-AGGCAAGATGGGTCACCAGC-3′, antisense: 5′-AGTCGAATCATCCATTCAGGTCfG-3′.
After 9 h of fasting, rats were killed by decapitation and hypothalamic nuclei were quickly dissected and homogenized in Trizol reagent (Life Technologies, Gaithersburg, MD, USA), according to the manufacturer's recommendations. Later on, each region of the hypothalamus was dissected from 1 mm thick sagittal sections of fresh brain. Arcuate nucleus was dissected from the first sections from the midline of the brain. Coordinates for the arcuate nucleus is ventral part of the medial hypothalamus with anterior and dorsal margin and posterior margin (border with mammilary body).
For mRNA localization all solutions and materials utilized were RNAse free. The probes were determined and designed using the program Gene Runner 3.05 (Hastings Software, Inc., USA) according to mRNA sequences in NCBI: POMC (NM_139326.2), NPY (NM_012614.1), AgRP (XM_574228.2), IL6ra (NM_017020.1), and IL10ra (AJ_305049.1). Two probes were synthesized for each mRNA and were 5′-end labeled with Alexa Fluor 488 or 546 by Invitrogen Life Technologies (Carlsbad CA, USA). See details in the supplemental data (Table S1). Frozen sections were air dried for 30 min at 37 °C, fixed using cold acetone for 10 min, and washed twice in PBS for 5 min and twice in 2× SSC for 2 min. The sections were incubated with Proteinase K (20 µg/mL) for 10 min at room temperature and then washed twice for 5 min with 2× SSC. The sections were incubated in 0.1 M triethanolamine pH 8 (TEA Buffer) for 10 min and then with 0.25% acetic anhydride in TEA buffer for 10 min under magnetic stirring and then washed with 2× SSC. The pre-hybridization solution was composed by 50% formamide, 5× SSC, Denhardt's solution (1× final concentration), and completed with DEPC-treated water. The sections were pre-hybridized for 4 h without the probe at 50 °C in humidified chamber with 50% formamide in SSC. The probe mix (including two probes for each mRNA; i.e., IL6ra or IL10ra with POMC, AgRP, or NPY) was composed (for each tissue section) of 20 µL of pre-hybridization solution plus 500 µg/mL of torula RNA, 500 µg/mL of salmon sperm DNA, and 50 ng of riboprobe mix (anti-sense or sense). The mixture was placed over the sections and incubated at 52 °C overnight in a humidified chamber. After 18 h hybridization, the sections were washed four times with 4× SSC buffer for 10 and 5 min in PBS. The sections were visualized in Zeiss 510 confocal microscope.
All numeric results are expressed as the means ± SEM of the indicated number of experiments. The results of blots are presented as direct comparisons of bands or spots in autoradiographs and quantified by optical densitometry (Scion Image). Statistical analysis was performed by employing the ANOVA test with Bonferroni post test. Significance was established at the p<0.05 level.
|
10.1371/journal.pcbi.1000007 | Robustness of Learning That Is Based on Covariance-Driven Synaptic Plasticity | It is widely believed that learning is due, at least in part, to long-lasting modifications of the strengths of synapses in the brain. Theoretical studies have shown that a family of synaptic plasticity rules, in which synaptic changes are driven by covariance, is particularly useful for many forms of learning, including associative memory, gradient estimation, and operant conditioning. Covariance-based plasticity is inherently sensitive. Even a slight mistuning of the parameters of a covariance-based plasticity rule is likely to result in substantial changes in synaptic efficacies. Therefore, the biological relevance of covariance-based plasticity models is questionable. Here, we study the effects of mistuning parameters of the plasticity rule in a decision making model in which synaptic plasticity is driven by the covariance of reward and neural activity. An exact covariance plasticity rule yields Herrnstein's matching law. We show that although the effect of slight mistuning of the plasticity rule on the synaptic efficacies is large, the behavioral effect is small. Thus, matching behavior is robust to mistuning of the parameters of the covariance-based plasticity rule. Furthermore, the mistuned covariance rule results in undermatching, which is consistent with experimentally observed behavior. These results substantiate the hypothesis that approximate covariance-based synaptic plasticity underlies operant conditioning. However, we show that the mistuning of the mean subtraction makes behavior sensitive to the mistuning of the properties of the decision making network. Thus, there is a tradeoff between the robustness of matching behavior to changes in the plasticity rule and its robustness to changes in the properties of the decision making network.
| It is widely believed that learning is due, at least in part, to modifications of synapses in the brain. The ability of a synapse to change its strength is called “synaptic plasticity,” and the rules governing these changes are a subject of intense research. Theoretical studies have shown that a particular family of synaptic plasticity rules, known as covariance rules, could underlie many forms of learning. While it is possible that a biological synapse would be able to approximately implement such abstract rules, it seems unlikely that this implementation would be exact. Covariance rules are inherently sensitive, and even a slight inaccuracy in their implementation is likely to result in substantial changes in synaptic strengths. Thus, the biological relevance of these rules remains questionable. Here we study the consequences of the mistuning of a covariance plasticity rule in the context of operant conditioning. In a previous study, we showed that an approximate phenomenological law of behavior called “the matching law” naturally emerges if synapses change according to the covariance rule. Here we show that although the effect of slight mistuning of the covariance rule on synaptic strengths is substantial, it leads to only small deviations from the matching law. Furthermore, these deviations are observed experimentally. Thus, our results support the hypothesis that covariance synaptic plasticity underlies operant conditioning.
| Synaptic plasticity that is driven by covariance is the basis of numerous models in computational neuroscience. It is the cornerstone of models of associative memory [1],[2],[3], is used in models of gradient estimation in reinforcement learning [4],[5],[6],[7],[8],[9],[10] and has been suggested to be the basis of operant conditioning [11]. In statistics, the covariance between two random variables is the mean value of their product, provided that one or both have a zero mean. Accordingly, covariance-based plasticity arises when synaptic changes are driven by the product of two stochastic variables, provided that the mean of one or both of these variables is subtracted such that they are measured relative to their mean value.
In order for a synapse to implement covariance-based plasticity, it must estimate and subtract the mean of a stochastic variable. In many neural systems, signals are subjected to high-pass filtering, in which the mean or “DC component” is attenuated relative to phasic signals [12],[13],[14],[15]. However, it is rare for the mean to be removed completely [16]. Therefore, while it is plausible that a biological synapse would be able to approximately subtract the mean, it seems unlikely that this mean subtraction will be complete. If mean subtraction is incomplete, the synapse is expected to potentiate constantly. Over time, this potentiation could accumulate and drive the synapse to saturation values that differ considerably from those predicted by the ideal covariance rule (see below). Thus, even if neurobiological systems actually implement approximate covariance-based plasticity, the relevance of the idealized covariance models to the actual behavior is not clear.
Here, we study the effect of incomplete mean subtraction in a model of operant conditioning, which is based on synaptic plasticity that is driven by the covariance of reward and neural activity. In operant conditioning, the outcome of a behavior changes the likelihood of the behavior to reoccur. The more a behavior is rewarded, the more it is likely to be repeated in the future. A quantitative description of this process of adaptation is obtained in experiments where a subject repeatedly chooses between two alternative options and is rewarded according to his choices. Choice preference is quantified using the ‘fractional choice’ pi, the number of trials in which alternative i was chosen divided by the total number of trials. The distribution of rewards delivered to the subject is quantified using the ‘fractional income’ ri, the accumulated rewards harvested from that alternative, divided by the accumulated rewards from all alternatives. In many such experiments, choice behavior can phenomenologically be described by(1)where i = 1,2 corresponds to the two alternatives, Dpi≡pi−0.5 and Dri≡ri−0.5. The proportionality constant, k corresponds to the susceptibility of choice behavior to the fractional income and its exact value has been a subject of intense debate over the last several decades. According to the ‘matching law’ k = 1 and thus pi = ri. In this case it can be shown that choices are allocated such that the average reward per choosing an alternative i, is equal for all alternatives [17],[18] (see also Materials and Methods). However, in many experiments the value of k is, in fact, slightly smaller than 1, a behavior that is commonly referred to as undermatching [19],[20],[21]. An alternative phenomenological description of behavior, known as ‘the generalized matching law’ [19] is p1/p2 = (r1/r2)k. Expanding the generalized matching law around ri = 0.5 yields Eq. (1) and thus Eq. (1) is an approximation of the generalized matching law. This approximation becomes equality for k = 1.
In a recent study we showed that the matching law is a natural consequence of synaptic plasticity that is driven by the covariance of reward and neural activity [11]. The goal of this paper is to understand the behavioral consequences of deviations from idealized covariance-based plasticity by investigating the behavioral consequences of incomplete subtraction of the mean in the plasticity rule. By studying an analytically solvable neural decision making model, we show that although the effect of small deviations from the idealized covariance-based plasticity on synaptic efficacies is large, the behavioral effect is small. Thus we demonstrate that matching behavior is robust to the mistuning of the parameters of the covariance-based plasticity rule. Furthermore, we show that the mistuning of the mean subtraction leads to undermatching, in line with experimental observations. Our study also reveals that the mistuning of the mean subtraction in the plasticity rule makes matching behavior sensitive to mistuning of the properties of the decision making network. Thus there is a tradeoff between robustness of matching behavior to changes in the plasticity rule and robustness to changes in the properties in the decision making network.
Decision making is commonly studied in experiments in which a subject repeatedly chooses between two alternative actions, each corresponding to a sensory cue. For example, in many primate experiments, the stimuli are two visual targets, and the actions are saccadic eye movements to the targets [20],[21]. In our model, the responses to the sensory stimuli are represented by two populations of sensory neurons, whose level of activity is denoted by N1 and N2 (Fig. 1A). We assume that the two activities Ni are independently drawn from the same Gaussian distribution with a positive mean and a coefficient of variation σ (standard deviation divided by the mean). We further assume that the level of variability in the activity of Ni is low, σ≪1. This assumption is reasonable if Ni corresponds to the average activity of a large population of uncorrelated neurons. Input from these sensory neurons determines the activities of two populations of premotor neurons via Mi = Wi·Ni where Wi corresponds to the synaptic efficacy of the sensory-to-premotor synapses. Competition between the two premotor populations determines whether the model will choose alternative 1 or 2 in a trial. Unless otherwise noted, alternative 1 is chosen in trials in which M1>M2. Otherwise alternative 2 is chosen. This process of competition between the two premotor populations can be achieved by a winner-take-all network with lateral inhibition [22], which is not explicitly modeled here. Thus, the larger the value of a synapse Wi is, the more likely it is that alternative i will be chosen.
Consider the following plasticity rule, in which the change ΔWi in synaptic efficacy Wi in a trial is described by(2)where η is the plasticity rate, R is the reward harvested in the trial, E[R] is the average of the previously harvested reward, Ni is the activity of sensory population i in the trial, and E[N] is the average activity of the sensory population. The index i is omitted from the latter average because we assume that the activity of the two populations is drawn from the same distribution; α, β are parameters. This plasticity rule corresponds to reward-modulated presynaptic activity-dependent plasticity [23],[24],[25]. If α = 1 and/or β = 1 then Eq. (2) describes a covariance-based synaptic plasticity rule because synaptic changes are driven by the product of two stochastic variables (Ni and R) where the mean of one or both of these variables is subtracted. In order to gain insights into the behavior of Eq. (2), we consider the average trajectory approximation, also known as mean synaptic dynamics [26],[27],[28],[29], which is the dynamics of the expectation value of the right hand side of Eq. (2). If the plasticity rate η is sufficiently small, the noise accumulated over an appreciable number of trials is small relative to the mean change in the synaptic efficacies, called the synaptic drift [26],[27] and(3)where we define a mistuning parameter γ = (1−α)·(1−β). γ = 0 corresponds to the idealized covariance rule. Incomplete mean subtraction corresponds to γ>0. Our analysis focuses on choice behavior when mean subtraction is incomplete (γ>0). Similar results are obtained when mean subtraction is overcomplete (γ<0; see Materials and Methods). In principle, even a small mistuning of the mean subtraction may have a substantial effect on choice behavior for the following reason: Consider the dynamics of Eq. (3) for the simple case in which reward R and neural activity Ni are independent. This corresponds to a case where the neural activity Ni does not participate in the decision making process or to the case where reward is independent of choice. In both cases, Cov[R, Ni] = 0 and therefore Eq. (3) becomes ΔWi≈η·γ·E[R]·E[N]. If E[R]·E[N]>0, the synaptic efficacy Wi is expected to grow indefinitely. The divergence of the synaptic efficacies is also expected in the more general case in which the reward and neural activities are not independent. This is illustrated in Fig. 1B, where we simulated the plasticity rule of Eq. (2) in a concurrent variable-interval schedule (VI; see Materials and Methods) and plotted the efficacy of one of the synapses as a function of the trial number. When the covariance rule is finely tuned such that γ = 0 (here we assumed that α = 0, β = 1), the synaptic efficacy, after a transient period (not shown), is approximately constant (blue line). After 300 trials (red, down-facing arrow), the mean subtraction in the plasticity rule was mistuned by 10% such that γ = 0.9 (α = 0, β = 0.9), resulting in the linear divergence of the synaptic efficacy (red line).
In practice, synaptic efficacies are bounded and such divergence is prevented by synaptic saturation. We model the synaptic saturation by adding a polynomial decay term to the synaptic plasticity rule such that Eq. (2) becomes(4)where ρ>0 is the saturation stiffness parameter. The effect of the decay term on the dynamics of the synaptic efficacy is illustrated in Fig. 1B. After 600 trials (black, left-facing arrow), the plasticity rule of Eq. (2) was replaced with the plasticity rule in Eq. (4) with ρ = 1, resulting in a convergence of the synaptic efficacy to a value that is significantly different from the result of the pure covariance rule (black line).
The synaptic saturation is modeled here using a saturation stiffness parameter, ρ. When ρ = 1, as in Fig. 1B (black line), synaptic efficacies decay linearly. The larger the value of ρ, the stiffer the bound. In the limit of ρ→∞, as long as Wi<Wbound Eq. (4) is equivalent to Eq. (2), but the saturation term prevents Wi from exceeding the value Wbound.
The dynamics of Eq. (4) are stochastic and therefore difficult to analyze. If the plasticity rate η is small then many trials with different realizations of choices and rewards are needed in order to make a substantial change in the value of the synaptic efficacies. Therefore intuitively, the stochastic dynamics of Eq. (4) can be viewed as an average deterministic trajectory, with stochastic fluctuations around it, where we expect that this average deterministic dynamics becomes a better approximation to the stochastic dynamics as the plasticity rate η becomes smaller. The conditions under which this intuitive picture is valid are discussed in [29]. The fixed point of the average trajectory of Eq. (4) is(5)and we study choice behavior when synaptic efficacies are given by Eq. (5). Assuming that p1, p2≠0, and γ>0, we show (Materials and Methods) that in the limit of low noise σ≪1, the model undermatches [19]; that is, when pi<0.5 then pi>ri whereas when pi>0.5 then pi<ri. Furthermore, the level of deviation from matching scales with the product of the mistuning and synaptic saturation parameters,(6)Finally, expansion of Eq. (6) around Dpi = 0 yields Eq. (1) with(7)Importantly, we show that overcomplete mean subtraction γ<0 also leads to undermatching with the same scaling of the deviations from matching with the mistuning and synaptic saturation parameters (Materials and Methods).
Consider Eq. (7). When γρ = 0, k = 1 and the fractional choice is equal to the fractional income yielding matching behavior. Note that when the mistuning of mean subtraction is small, γ≪1, the deviation of the susceptibility index k from 1 is small. This occurs despite the fact that such mistuning has, in general, a substantial effect on the values of the synaptic efficacies (Fig. 1B). Thus, matching behavior is robust to the mistuning of the mean subtraction, even though the synaptic efficacies are not.
Eq. (7) is derived assuming that the stochastic dynamics, Eq. (4) has converged to the fixed point of the average trajectory, Eq. (5) and that σ≪1 (Materials and Methods). In order to study the validity of this approximation, we numerically simulated the decision making model with σ = 0.1 and a stochastic synaptic plasticity rule, Eq. (4) in a concurrent VI reward schedule (Materials and Methods). These simulations are presented in Fig. 2. Each symbol in Fig. 2A corresponds to one simulation in which the baiting probabilities of the two targets were kept fixed. The fraction of trials in which action 1 was chosen is plotted against the fractional income earned from action 1. As predicted by Eq. (7), the dependence of the fractional choice on the fractional income is linear, and susceptibility depends on the values of both γ and ρ (red squares, γ = 0.05, ρ = 1; blue diamonds, γ = 0.5, ρ = 1; gray triangles γ = 0.5, ρ = 4; colored lines are the analytical approximation, Eq. (7); the black line is the expected behavior according to the matching law). In order to better quantify the relation between the stochastic dynamics and the analytical approximation, we simulated Eq. (4) for different values of γ and ρ and measured the susceptibility of behavior. The results of these simulations appear in Fig. 2B (blue dots, ρ = 5; red dots, ρ = 1; black dots, ρ = 0.2) and show good fit with the expected behavior from Eq. (7) (lines).
In the previous section we analyzed the behavioral consequences of mistuning of the plasticity rule in a particular network model. The question of robustness is equally applicable to the parameters of the decision making network as it is to the parameters of the synaptic plasticity rule. Therefore, in this section we study the robustness of matching behavior to the mistuning of the parameters of the network.
There are various ways in which the decision making network can be mistuned. We chose to study the effect of a bias in the winner-take-all network, because this is a generic form of error that is likely to significantly affect choice behavior. It is plausible that a winner-take-all network will be able to choose the alternative that corresponds to the larger activity of the two premotor populations in trials in which M1 and M2 are very different. However, if M1 and M2 are similar in their level of activity it is likely that a biological implementation of a winner-take-all mechanism, which is not finely tuned, will be biased to favoring one of the alternatives. Formally we assume that alternative 1 is chosen in trials in which (M1−M2)/(M1+M2)>ε where ε is a bias. The unbiased case studied in the previous section corresponds to ε = 0. In contrast, ε>1 or ε<–1 correspond to a strong bias such that choice is independent of the values of M1 and M2. With the same assumptions as in the derivation of Eq. (7), p1, p2≠0 and σ≪1, we show (Materials and Methods) that a bias in the winner-take-all mechanism results in a bias in choice that is O(ργ·ε/σ). Furthermore, analyzing choice behavior for small value of |Dpi| yields(8)where k is given by Eq. (7) and(9)is the offset. The offset b1 is proportional to the deviation of the susceptibility of behavior from unity, 1−k. As discussed in the previous section, this deviation depends on the level of incomplete mean subtraction as well as the synaptic saturation term (Eq. (7). If γ = 0 then k = 1 and the offset term vanishes, b1 = 0 for any value of bias ε. This robustness of matching behavior to bias in the winner-take-all network is due to the fact that the idealized covariance based plasticity rule can compensate for the bias in the decision making network in almost any neural architecture [11]. In contrast, if γ>0 then the offset b1 is proportional to the bias ε. The larger the deviation of the plasticity rule from the idealized covariance rule, the larger the proportionality constant. Thus, there is a tradeoff between the robustness of matching behavior to changes in the plasticity rule and robustness to changes in the parameters of decision making. The larger the mistuning of the plasticity rule, the smaller the robustness of matching behavior to mistuning of the parameters of the decision making network. Importantly, the level of noise in the sensory populations strongly affects the bias in behavior through ε/σ. This contrasts with the independence of the susceptibility parameter k of σ. To understand the reason for this result it is useful to note that as discussed in the previous section, the magnitude of trial to trial fluctuations in the activity of the sensory neurons determines the magnitude of the fractional income signal stored in the synaptic efficacies (the difference in the two synaptic efficacies). The smaller the value of σ is, the weaker the fractional income signal and therefore the stronger the relative contribution of the bias in the winner-take-all network to choice. If Ni corresponds to the average activity of a large population of uncorrelated neurons, σ is expected to be small and therefore the effect of even small bias in the winner-take-all network on behavior is expected to be large.
To study the validity of Eq. (8) numerically, we simulated the synaptic plasticity rule of Eq. (4) in the decision making model of Fig. 1A with a bias ε in the winner-take-all network. Similar to Fig. 2A, Fig. 3A depicts the fraction of trials in which alternative 1 was chosen, which is plotted against the fractional income earned from that alternative. The level of deviation from matching behavior (solid black line) depends on the value of ε (red squares, ε = −3σ; blue diamonds, ε = 0; gray triangle, ε = 3σ; γ = 0.05, ρ = 1). Colored lines are the analytical approximation, Eq. (8). In order to better quantify the relation between the stochastic dynamics and its deterministic approximation, we numerically computed the value of p1 that corresponds to δr1 = 0 for different values of ε and γ (Fig. 3B; red, γ = 0.05; blue, γ = 0.5). The results are in line with the expected behavior from Eq. (8) (solid lines).
In this study we explored the robustness of matching behavior to inaccurate mean subtraction in a covariance-based plasticity rule. We have shown that (1) although this deviation from the idealized covariance rule has a substantial effect on the synaptic efficacies, its behavioral effect is small. (2) The direction of the behavioral effect of incomplete mean subtraction is towards the experimentally observed undermatching. (3) When the plasticity rule is mistuned, matching behavior becomes sensitive to the properties of the network architecture. Thus, there is a tradeoff between the robustness of matching behavior to changes in the plasticity rule and robustness to changes in the parameters of the decision making network.
Covariance-based, Hebbian synaptic plasticity dominates models of associative memory. According to the popular Hopfield model, the change in the synaptic efficacy between pairs of neurons is proportional to the product of their activities in the training session, measured relative to their average activity [1],[2],[3]. If the mean subtraction is not finely tuned in this model, the synaptic efficacies diverge with the number of patterns stored. If this divergence is avoided by adding a saturation term to the plasticity rule, the capacity of the network to store a large number of memory patterns is lost [2],[30]. Thus, fine tuning of the mean subtraction in the plasticity rule is crucial for covariance-based associative memory models. This contrasts with the robustness of matching behavior to the mistuning of the mean subtraction demonstrated here. The difference in robustness stems from the difference in the solution space of the two tasks. Consider a general decision making network model consisting of n synapses. If n>1 the decision making model is expected to be redundant. There are many possible combinations of synaptic efficacies that yield the same probability of choice and thus are behaviorally indistinguishable. The dimension of the hyperspace of synaptic efficacies that corresponds to a single probability of choice is, in general, n−1. Consider now the hyperspace of synaptic efficacies that corresponds to the matching solution p1 = r1. Any set of synaptic efficacies that resides within this hyperspace is a fixed point of the family of synaptic plasticity rules that is driven by the covariance of reward and neural activity (in the average trajectory approximation) [11]. In contrast to this manifold of solutions, the approximate covariance plasticity rule with saturation is expected to have a single fixed point. In order for this fixed point to correspond to an approximate matching solution, it should reside near the matching hyperspace. The distance of the fixed point solution from the matching hyperspace depends on the decision making model and the level of mistuning of the covariance plasticity rule. However, because of the high dimensionality of the matching solution, there is a large family of decision making models in which the solution to the approximate covariance plasticity rule resides near the matching hyperspace for that model, for example, the model analyzed here with ε = 0. In contrast, in associative memory models, the volume in the synaptic efficacies hyperspace that can retrieve a large number of particular memories is small [31] and therefore even small deviations from the covariance plasticity rule will lead to a solution that is far from the memory retrieving hyperspace, resulting in a large reduction in the performance of the network.
Several studies have reported stochastic gradient learning in a model in which changes in the synaptic efficacy are driven by the product of the reward with a measure of past activity known as the ‘eligibility trace’ [4],[5],[6],[7],[8],[9],[10]. The mean of the eligibility trace is zero and therefore synaptic plasticity in these models can be said to be driven by the covariance of reward and a measure of past activity. Violation of the zero mean condition is expected to produce a bias in the gradient estimation and could potentially hinder learning. The consequences of mistuning of the mean subtraction in the estimation of the eligibility trace have not been addressed. We predict that the relative volume in the model parameter hyperspace that corresponds to the maximum reward solution will be an important factor in determining whether these gradient learning models are robust or not to the mistuning of the mean subtraction.
The level of fine-tuning required for normal brain functioning is unknown and robustness represents a major open issue for many models of brain systems. For example, the fine-tuning of neural parameters involved in the short term memory of analog quantities such as eye position in the oculomotor neural integrator [32],[33],[34],[35] or the frequency of a somatosensory stimulation [36],[37] have been studied extensively. It has been suggested that synaptic plasticity keeps the synaptic efficacies finely-tuned [38],[39]. However, in those models it is assumed that the parameters of the plasticity rule are finely tuned. In this study we demonstrated a tradeoff between the robustness of behavior to changes in the parameters of the network architecture and the robustness to changes in the parameters of the plasticity rule. This tradeoff is likely to be a property of many models of brain function.
Undermatching in our model is the outcome of inaccurate mean subtraction, whether it is incomplete or overcomplete. This result is expected to hold in other symmetrical decision making models: when the mean subtraction is inaccurate, synaptic efficacies are determined by a combination of a covariance term, and bias and saturation terms. The bias and saturation terms are not influenced by the correlation between the neural activity and the reward. Therefore they drive the synaptic efficacies to values that are independent of the fractional income. If the architecture of the decision making network is symmetrical with respect to the two alternatives (as is the case in our model for ε = 0), they will drive the synaptic efficacies in the direction of a symmetrical solution for which the two alternatives are chosen with equal probability, which corresponds to k = 0. In contrast, the covariance term drives the efficacies to the matching solution, k = 1. The combined effect of the covariance term and a small bias and saturation terms is expected to be a behavior for which the susceptibility index k is slightly smaller than 1, in line with the experimentally observed slight undermatching. Importantly, the experimentally observed undermatching is consistent with approximate covariance-based synaptic plasticity but does not prove it. Undermatching is also consistent with other models that do not assume this particular synaptic plasticity rule (see below).
We hypothesize that the observed matching behavior results from a synaptic plasticity rule that is driven by an approximation to the covariance of reward and neural activity. In this case, behavior adapts because synapses in the brain perform a statistical computation and ‘attempt’ to decorrelate the reward and the fluctuations in neural activity. However, a very different class of matching models has been proposed, in which the brain performs computations that are “financial.” According to these models, subjects keep track of financial quantities such as return or income from each alternative and make choices stochastically according to the difference or ratio of the financial quantities between the two alternatives leading to matching [20],[40],[41], or undermatching [42],[43]. A common feature of these models is the implicit assumption that financial computations and probabilistic choice are implemented in two separate brain modules. One brain module records past reward and choices to calculate quantities such as income and return and the other brain module utilizes these quantities to generate stochastic choice. A covariance-based plasticity rule can be distinguished experimentally from the financial models by making the reward directly contingent on fluctuations in the stochastic neural activity. This could be done by measuring neural activity in a brain area involved in decision making, using microelectrodes or brain imaging, and making reward contingent on these measurements, as well as on actions. This sort of contingency has previously been employed by neurophysiologists, though not in the context of operant matching [44],[45]. If, by the construction of the reward schedule, reward directly depends on fluctuations in neural activity, then it would be impossible to decorrelate the reward and the neural activity. According to our covariance hypothesis, the ‘attempt’ of the synaptic plasticity rule to do just this will lead to a change in the dependence of choice on the financial quantities (formally, this will lead to violation of Eq. (21) in Materials and Methods). In contrast, in the financial models, neural fluctuations and learning are mediated through different modules and therefore this contingency will not alter the dependence of choice on financial quantities (see also [11]).
As was described above, the identity of choice in the network of Fig. 1 is determined by a competition between two premotor neurons Mi = Wi·Ni. In the Incomplete mean subtraction section we assume that alternative 1 is chosen in trials in which M1>M2. Otherwise alternative 2 is chosen. Thus, the fraction of trials in which alternative 1 is chosen, or the probability that it is chosen is given by(10)where A∈{1,2} denotes the alternative chosen, or(11)where Zd≡(δN1−δN2)/(2·E[N]), Zs≡(δN1+δN2)/(2·E[N]), δNi = Ni−E[N], T≡Wd/Ws, Ws≡(W1+W2)/2, Wd≡(W1−W2)/2. Because N1 and N2 are independent Gaussian variables with a coefficient of variation σ, Zd and Zs are two independent Gaussian variables with zero mean and standard deviation. Therefore, Zd+T·Zs is a Gaussian variable with zero mean and standard deviation and(12)Note that the assumption that p1,p2≠0 implies that in the limit of σ→0, T = O(σ).
Next we use Eq. (11) to compute two quantities that will become useful later:(13)and similarly(14)Assuming that T = O(σ),(15)and(16)
In this section we compute the dependence of deviations from matching behavior on γ, assuming that synaptic efficacies are given by the fixed point of the average trajectory, Eq. (5). The precise conditions for the correctness of the approach are discussed in details in [29]. We further assume that synaptic saturation is linear, ρ = 1. The latter assumption is relaxed in the Incomplete mean subtraction and saturation stiffness section below.
According to Eq. (11), the probability of choice depends on the ratio of the synaptic efficacies; thus the scaling of the synaptic efficacies by a positive number does not change the probabilities of choice. For clarity we scale the synaptic efficacies of Eq. (5) (assuming ρ = 1) such that,(17)Rewriting Eq. (17) in terms of Wd and Ws yields(18)(19)where the asterisk corresponds to the value at the fixed point. Next we separate the covariance terms into trials in which alternative 1 was chosen and trials in which alternative 2 was chosen(20)The reward R is a function of the actions A and the actions are a function of the neural activities Zs and Zd. Therefore, given the action, the reward and the neural activities are statistically independent and the average of the product of reward and neural activity is equal to the product of the averages, E[R/E[R]·Zx|A = i] = E[R/E[R]|A = i]·E[Zx|A = i]. Hence, Eq. (20) becomes(21)Next we separate E[Zx] to trials in which alternative 1 was chosen and trials in which alternative 2 was chosen and use the fact that E[Zx] = 0(22)Substituting Eq. (22) in Eq. (21) yields(23)In order to evaluate the second term in the right hand side of Eq. (23) we note that by definition, ri = pi·E[R|A = i]/E[R] and therefore,(24)where we assumed that p1,p2≠0 and used the fact that p1+p2 = 1 and r1+r2 = 1. Substituting Eqs. (13), (14), (23) and (24) in Eqs. (18) and (19) yields(25)and(26)where . Combining Eqs. (25) and (26),or(27)Eq. (27) is central to this manuscript. Together with Eq. (12) which relates the probability of choice p1 with T it determines the level of deviations from matching behavior at the fixed point, (The relation between r1 and p1 is determined by the reward schedule). Next we use Eq. (27) to show that:
(1) As was discussed above, the assumption that p1,p2≠0 in the limit of σ→0 implies that T = O(σ)and therefore 1−T2>0. Thus, . Using Eq. (12) and the notations of Eq. (1),(28)(Dp1 and Dr1 in Eq. (28) are the values at the fixed point and therefore a more accurate notation would have included an asterisk. However, in order to keep notations in the text simple and notations in the Materials and Methods section consistent with the text we omitted the asterisk). When , whereas when , . Thus we have shown that in the limit of σ→0 the model undermatches.
(2) Taking the dominant terms in σ in Eq. (27) yields(29)T* = O(σ) and thus the second term in the right hand side of Eq. (29) is O(1); therefore, the level of deviations from matching behavior is O(γ), Eq (6).
(3) In order to obtain a closed form approximation to Eq. (29) we expand Eq. (12) around Dp1 = 0 yielding(30)Expanding Eq. (29) around Dpi = 0 and using Eq. (30) yields Eq. (7).
In order to study the effect of bias in the winner-take-all network on choice behavior, we assume that that alternative 1 is chosen in trials in which (M1−M2)/(M1+M2)>ε where ε is a bias. Formally,(31)Rewriting Eq. (31) in terms of Zs and Zd yields(32)where(33)or(34)The assumption that p1,p2≠0 implies in the limit of σ→0 T′ = O(σ). As in the derivation of Eqs. (13) and (14)(35)and(36)Assuming that T′ = O(σ),(37)and(38)From here we follow the same steps as in the derivation of Eq. (27) yielding(39)or(40)Assuming that T′* = O(σ) and taking the limit σ→0 yields(41)Because r1−p1 = O(1), the assumption that implies that γ·ε/σ = O(1). Thus in the limit of σ→0, ε≪1. Taking O(ε) terms in Eq. (41) yields(42)The first term in the right hand side of Eq. (42) is equal to the right hand side of Eq. (29) and yields O(γ) deviations from matching behavior in the direction of undermatching. The bias in the decision making process, ε affects choice preference through the second term in the right hand side of Eq. (29). For T′ = O(σ), and the contribution of the bias term ε to deviations from matching is O(γ·ε/σ).
Expanding Eqs. (34) and (42) around Dpi = 0 yields Eq. (8).
Rewriting Eq. (5),(43)Next we show that in the limit σ→0 and assuming that , Cov[R/E[R],Ni/E[N]]/γ≪1 and therefore the second term in the right hand side of Eq. (43) can be expanded around 1. In order to see this, we follow the same route as in the derivation of Eq. (23) and separate the covariance term into trials in which alternative 1 was chosen and trials in which alternative 2 was chosen(44)As before, the reward R is a function to the actions, which in turn, are a function of the neural activity. Therefore, given the action A, R and δNi are statistically independent and therefore(45)By construction, E[δNi/E[N]] = 0 and therefore,(46)Substituting Eq. (46) in Eq. (45) yields(47)Note that(48)Substituting Eqs. (16) and (15) in Eq. (48) yields,(49)Using Eq. (24), the assumption that and taking the limit σ→0, such that σ/γ≪1 yields Cov[R/E[R],Ni/E[N]]/γ≪1. In fact, substituting Eq. (6) in Eq. (24), Cov[R/E[R],Ni/E[N]]/γ≪1 even when σ/γ↛0 as σ→0. Therefore, using self consistent arguments, the derivation of Eq. (50) is valid even when γ scales like σ. Expanding the second term in the right hand side of Eq. (43) yields,(50)According to Eq. (11), the probability of choice depends only on the ratio W1/W2. Therefore, the first term in the right hand side of Eq. (50) does not affect the probabilities of choice. The saturation stiffness parameter ρ affects the probability of choice through the second term and this effect is equivalent to the scaling of the mistuning parameter γ by ρ. Thus, assuming that synaptic efficacies converge to the fixed point of the average trajectory, Eq. (5), the effect of deviations of the saturation stiffness parameter from unity on choice is equivalent to the scaling of γ by ρ.
The synaptic saturation term also changes the effective plasticity rate, which will change the conditions of applicability of the average trajectory approximation. This analysis goes beyond the scope of this manuscript and will be discussed elsewhere. In short, changing the value of ρ changes the effective plasticity rate to . Therefore in the simulations in Fig. 2 we used(51)
According to Eq. (3), when γ<0, Wi is expected to depress until it becomes negative. In reality, synaptic efficacies are bounded and synaptic saturation prevents them from changing their sign. We model the synaptic saturation by replacing the synaptic plasticity rule of Eq. (2) by(52)where ρ>0 is the saturation stiffness parameter. The larger the value of ρ, the stiffer the bound. In the limit of ρ→∞, as long as Wi>Wlow Eq. (52) is equivalent to Eq. (2), but Wi is bounded from going below Wlow.
The fixed point of the average trajectory of Eq. (52) is(53)Following the same steps as in the derivation of Eq. (50), the limit σ→0 with the assumption that yields(54)Thus, assuming that synaptic efficacies converge to the fixed point of the average trajectory, Eq. (5), the behavior of a model with overcomplete mean subtraction is similar to that of a model with incomplete mean subtraction. In both cases the synaptic efficacies are given by(55)where |
10.1371/journal.pcbi.1004396 | Manifold Based Optimization for Single-Cell 3D Genome Reconstruction | The three-dimensional (3D) structure of the genome is important for orchestration of gene expression and cell differentiation. While mapping genomes in 3D has for a long time been elusive, recent adaptations of high-throughput sequencing to chromosome conformation capture (3C) techniques, allows for genome-wide structural characterization for the first time. However, reconstruction of "consensus" 3D genomes from 3C-based data is a challenging problem, since the data are aggregated over millions of cells. Recent single-cell adaptations to the 3C-technique, however, allow for non-aggregated structural assessment of genome structure, but data suffer from sparse and noisy interaction sampling. We present a manifold based optimization (MBO) approach for the reconstruction of 3D genome structure from chromosomal contact data. We show that MBO is able to reconstruct 3D structures based on the chromosomal contacts, imposing fewer structural violations than comparable methods. Additionally, MBO is suitable for efficient high-throughput reconstruction of large systems, such as entire genomes, allowing for comparative studies of genomic structure across cell-lines and different species.
| Understanding how the genome is folded in three-dimensional (3D) space is crucial for unravelling the complex regulatory mechanisms underlying the differentiation and proliferation of cells. With recent high-throughput adaptations of chromosome conformation capture in techniques such as single-cell Hi-C, it is now possible to probe 3D information of chromosomes genome-wide. Such experiments, however, only provide sparse information about contacts between regions in the genome. We have developed a tool, based on manifold based optimization (MBO), that reconstructs 3D structures from such contact information. We show that MBO allows for reconstruction of 3D genomes more consistent with the original contact map, and with fewer structural violations compared to other, related methods. Since MBO is also computationally fast, it can be used for high-throughput and large-scale 3D reconstruction of entire genomes.
| Understanding genomes in three dimensions (3D) is a fundamental problem in biology. Recently, the combination of chromosome conformation capture (3C) methods with next-generation sequencing, such as 5C [1], Hi-C [2], TCC [3], and GCC [4], has enabled the study of contact frequencies across large genomic regions or entire genomes. These methods consist in crosslinking a large sample of cells followed by restriction enzyme digestion and ligation. Ligated DNA molecules are isolated, and sequenced using massively parallel paired-end sequencing. The end-result is typically a large matrix containing interaction (ligation) frequencies between all regions of the genome under study in the cell population. While such matrices can be visualized and analyzed directly [2], determining the 3D structure corresponding to the interaction frequency matrix has been of steady increasing interest in the fields of computational biology and genomics. However, such 3D genome reconstruction is challenging due to the sparse and noisy nature of the data, the fact that the matrices typically contain aggregated interaction frequencies across millions of cells [5], and the dynamic nature of chromatin [6]. These limitations constitute an obvious problem with respect to reconstructing a “consensus” 3D structure.
Several approaches have been proposed to take into account the dynamic nature of chromatin and the aggregated nature of the data. Baù et al. [7] used the Integrative Modelling Platform (IMP) [8, 9] and a Markov Chain Monte Carlo (MCMC) method to simulate a large set of 50,000 independent structural models from 5C data. A subset of the resulting structural ensemble consisting of the 10,000 structures with the best scores was then clustered, such that the different clusters arguably represent the variability of chromatin conformation in the population-averaged data. An MCMC approach for structural ensemble determination from 5C data was also utilized in a study by Rousseau et al. [10], leading to a probabilistic model of the interaction frequency data. This allows for sampling from the posterior distribution of structures after a sufficient number of Monte Carlo steps. IMP has also been used to simulate an ensemble of 10,000 structures, that simultaneously encounter the restraints, assuming that the ensemble represents the dynamic nature of chromatin [3].
Another class of methods for identifying 3D chromatin structure from chromosomal contact data relies on reconstructing a “consensus” 3D structure from a (possibly incomplete and noisy) Euclidean distance matrix (EDM) consisting of pairwise distances (in 3D) between different regions in the genome. In general, this EDM is not known, but is typically estimated from the interaction frequency matrix. Given an EDM various optimization approaches that fall under the general topic of multidimensional scaling (MDS) (see e.g. [11] for an overview) can be used to find an optimal 3D structure. Methods based on MDS are often simpler and can handle larger problems, such as multiple chromosomes or single chromosomes on finer scales, than many of the more complex probability based methods. On the other hand, such methods often ignore the dynamic nature of chromatin and the aggregated nature of the Hi-C data.
The most basic form of MDS is the so-called classical (or metric) MDS, where the optimal coordinate reconstruction from a given EDM is found directly by eigen decomposition of the so-called Gram matrix (see Methods for details). An early application of classical MDS to determine 3D structure from chromosome contact data was presented by Dekker et al. [12]. In general, however, when the EDM has been inferred from interaction frequencies, the MDS approaches consider the reconstruction as a nonlinear and non-convex optimization problem using some iterative optimization method. For example, the EDM has been inferred by assuming simple transformations of genomic distances to Euclidean distances, and an iterative optimization method has been applied to reconstruct the coordinates best corresponding to the EDM [13, 14].
Other optimization methods applied on MDS problems to find coordinates from incomplete distances exploit the rank constraints on the EDM (or corresponding Gram matrix) to find an optimal EDM for the relevant spatial dimension. One successful method in this respect is based on convex semidefinite programming [15, 16], which relaxes the problem to a convex optimization problem. These approaches are applicable to model 3D chromosome configurations [17]; however they cannot handle large problems, due to computational limitations.
Technological improvements have also facilitated the reconstruction of 3D genome structures. In particular, adjustments to the Hi-C protocol have been introduced to enable identification of interactions between chromosome regions in single cells [18]. Single-cell Hi-C, however, inevitably suffers from sparse sampling of chromosomal interactions and a general lack of information on non-local distances between genomic regions with no mutual contacts. Nonetheless, mapped interactions are found in individual cells, potentially enabling a more robust determination of the underlying 3D structure [18].
One way to handle these limitations is to replace missing distances with their ‘shortest-path’ equivalence; that is, considering the existing (observed) entries in the EDM as weighted edges in a graph, and replacing each missing edge weight with the smallest possible sum of weights traversing the graph along the observed edges [19]. One drawback of completing the EDM with the shortest-path distances, however, may be that the imputed distances introduce noise which dominates over the more accurate local distances.
An application of single-cell like contact maps coupled with missing-value imputation using the shortest-path method and classical MDS to find 3D coordinates, was recently proposed [20]. This approach offers an efficient way of establishing 3D genome structures. However, accuracy may be limited both by the noise introduced by the shortest-path procedure as well as from the limitations of the classical MDS approach.
Another approach proven to be effective on many optimization problems relies on optimization on manifolds. The problem of finding optimal coordinates from an EDM can be formulated as an optimization problem on the manifold of the set of positive semidefinite matrices of fixed rank [21, 22]. The Riemannian quotient geometry of the manifold can be exploited to yield efficient algorithms for the optimization problem [23]. However, this strategy has, to our knowledge, not been applied to 3D genome reconstruction in previous studies.
In this paper, we show that the manifold based optimization (MBO) approach can be successfully applied to 3D genome reconstruction. MBO significantly outperforms the simpler methods based on classical MDS in terms of consistency with the original contact map and structural violations, while remaining sufficiently efficient to handle large-scale problems.
Using both simulated and real single-cell Hi-C data, we show that, by combining the shortest-path derived distances with appropriate weights to reduce the influence of noise, MBO can efficiently reconstruct 3D structures consistent with the chromosome contact maps, despite the noisy and sparse nature of the data. Our implementation of the manifold optimization method is based on the Manopt software [24] that provides a Matlab interface for optimization on manifolds.
In the following sections, we apply MBO to reconstruct the 3D structure of genomes in two types of settings, and compare to two other popular approaches. First, to evaluate the method’s ability to reconstruct a known 3D structure, we consider a given a priori 3D structure, and sample contact frequencies from this structure. Then, we apply the methods to recently published single-cell Hi-C data [18], and evaluate the ability of the resulting structural models to reconstruct the original contact maps.
Given a matrix of interaction frequencies, typically from a Hi-C or single-cell Hi-C data set, we seek to reconstruct the corresponding 3D coordinates of the genome structure. In classical MDS (CMDS), this reconstruction is performed by converting the contact frequencies into an EDM (Fig 1B), and uses singular value decomposition for direct coordinate reconstruction. Crucially, such approaches assume that all Euclidean distances in the EDM are of equal importance and equally accurate. This is problematic, since it is known that short genomic distances are sampled much more frequently than long genomic distances. Also, in single-cell Hi-C, contacts are restricted to only two interactions per restriction fragment, for autosomal chromosome pairs, resulting in a large number of missing values.
In our method, which relies on manifold based optimization (MBO) [22], the low-rank property of the EDM, and the resulting redundancy in the distances, are exploited to infer the missing distances. We consider the completion of the EDM while simultaneously allowing for missing distances. We do this by combining the shortest-path completed distances with weights, such that imputed (and typically long) distances are weighted less in the subsequent optimization procedure (Fig 1C). This allows for flexibility in the reconstruction of uncertain regions of the final 3D structure, while enforcing distances in more reliable sections of the structure. The Methods section provides an in-depth description of the full algorithm.
As a first validation of the method, we have considered an in silico test case where a known chromosome structure was employed to test the ability of different methods to reconstruct the original structure from incomplete and noisy distance information. Here, MBO is compared to the classical MDS (CMDS) approach recently presented in Lesne et al. [20], where the graph shortest-path method is utilized to replace missing distances. This method is generally known as Isomap [19], while the adaptation to 3D genome reconstruction was named ShRec3D in Lesne et al. [20]. In the following we will refer to this method simply as CMDS. In addition, we present comparison with the ChromSDE method of Zhang et al. [17], which is based on semidefinite programming and is significantly more computationally demanding than both the CMDS method and MBO.
The structure considered in this validation is a 3D model of mouse haploid chromosome X generated from single-cell Hi-C data by Nagano et al. [18]. The 3D model represents chromosome X using a 50 kilo base pair (kbp) resolution. However, for the current test, the structure was re-sampled at 600 kbp, by taking the average spatial position of groups of bins, this due to the computational limitation of the ChromSDE method. Additionally, we evaluate different levels of noise (σ), added to the final contact matrix, as well as different levels of contact scarcity (see Methods section). The results from these tests are shown in Fig 2. The data shows the structural similarity between original distances and reconstructed distances for the different methods, for different noise levels (σ) and ratios of missing distances.
For the weakly noisy case (Fig 2; σ = 0.1) MBO and ChromSDE still reconstruct structures more consistent with the orignal structure than CMDS. For the two cases with higher noise levels, however, MBO performs markedly better, and produces structures more similar to the original, compared to the two other methods (Fig 2; σ = 0.5 and 1.0).
In the noiseless case (σ = 0) both MBO and ChromSDE are able to reconstruct the original structure exactly as long as a sufficient number of the pair-wise distances are known. This would be expected for ChromSDE, since the semidefinite programming approach is convex in this case. That MBO also recovers the original coordinates exactly is not a priori obvious. Naturally, the ratio of distances needed for an exact reconstruction will vary with the size n of the problem. In fact, it has been shown that knowledge of m ≥ Cn6/5 r log n (for some positive contant C) random entries of an n × n matrix of rank r is sufficient for an exact completion of the matrix in most cases [25].
We inspected the ability of MBO to reconstruct the considered orignal structure when the missing distances approach this limit. The original structure can be exactly reconstructed with up to ∼ 90% missing data (Fig 3A–3C). With 95% missing data, the structure is still similar to the original structure, with an RMSD of ∼ 610 nm. At levels of missing data above 98%, however, the structure collapses into a compact globule, due to missing interactions between distal bins (Fig 3E–3F).
To inspect this dependency further, we calculated the minimum ratio of observed distance values needed for complete reconstruction ([1-ρ]<1e-10) and partial reconstruction ([1-ρ]<0.1)), for a range of different sampled structures with varying number of bins (n) (see Fig 3G). The required percentage of observed interactions is dependent on the total number of bins in the system considered. We compared the structures from Fig 3B–3F with these estimated curves, and indeed found that the compact globular structures correspond to a ratio of observed values crossing the boundary of partial reconstruction. Furthermore, we compared these curves to the sets of all chromosomes from the single-cell Hi-C data from [18]. As can be seen in Fig 3G, the datasets are distributed around the curve of partial reconstruction ([1-ρ]<0.1)). This could indicate that the current single cell Hi-C data sets are generally too sparse for high confidence structure reconstruction. Note, however, that the single-cell Hi-C data for chromosome X (cell 1 and cell 2) are between the partial and complete reconstruction curves, and are therefore likely to be among the more reliable data sets for structural reconstruction and method comparisons.
Typical computation times for the methods considered in the validation performed above are shown in Fig 4, as a function of the problem size n (i.e. n is the number of bins in the reconstructed structure). As expected, CMDS (excluding the shortest-path algorithm) is fastest, while ChromSDE is slowest. Note, however, that MBO has the same asymptotic behavior as CMDS for large n. Further, when the input EDM has missing values, the shortest-path distances must be calculated before application of CMDS. Hence, for n larger than about 500, MBO is actually the fastest of the three methods. In practice, using stringent settings, reconstruction of e.g. chromosome X using MBO at 50kbp resolution takes less than 5 minutes.
Next, we examined the ability of MBO and CMDS to reconstruct contact maps for the full set of chromosomes, based on single-cell Hi-C data [18]. We therefore applied MBO and CMDS to all mouse chromosomes individually, for two different single cells (named “cell 1” and “cell 2” in [18]), and evaluated the resulting structures. We evaluated and compared the ability of the methods to reconstruct structures with resulting contact maps consistent with the input data, by inspecting the percentage of contacts established in the reconstructed structure that were also present in the original contact map (% correct contacts). Additionally, we evaluated the occurrence of structural inconsistencies in the inferred structures, i.e. the percentage of bins being too close to each other (% min distance violation), and the percentage of consecutive bins that are too far away from each other (% connectivity violation). See the Methods section for details.
We started by considering chromosome X, where only one copy is present in the data. For chromosome X, we found that MBO was able to reconstruct the original contact map of the haploid X nearly completely (both cases > 95% reconstructed). CMDS, on the other hand, was not able to reconstruct the contact matrix of chromosome X at more than ∼50–60% correct contacts (Figs 5C and 6A). Similar results were found for all 10 individual cells from [18] (see S1 Fig), even though the percentage of correct contacts was closer to 80% for some of the cells with the fewest number of input contacts (cells 9 and 10).
Interestingly, for homologous chromosome pairs, where two chromosome copies are present, reconstruction was not as consistent with the input contact maps as for chromosome X, as only ∼20% of the contacts in the original maps could be reconstructed (Fig 6A). In other words, the presence of two chromosomal copies affects the ability to reconstruct structures that reflect the original contact matrix. This indicates that the structures of the two homologous copies may contain mutually exclusive contacts, making full reconstruction of the contact maps difficult.
We were interested in investigating the effect of having possibly mutually exclusive contact information from two separate chromosome X structures from cell 1 and cell 2. We therefore randomly sampled 50 new datasets consisting of an equal number of contacts from the matrices from these two cells and inspected the ability of MBO to reconstruct structures corresponding to the resulting contact maps. As S2 Fig shows, the mixed datasets produce structures with a significantly lower percentage of correct contacts, and structures with higher connectivity violations. It should be noted that 3D reconstruction from mixed populations of contact data has no guarantee of reliably estimating a correct structure.
For homologous chromosome pairs, MBO and CMDS performed similarly in terms of percentage of successfully established interactions (Fig 6A). However, when looking at minimum distance violations (chromosomal bins closer than 30 nm, Fig 6B, or violations of the connectivity of consecutive regions (consecutive bins further away than 200 nm, Fig 6C), it is clear that MBO is more successful in positioning the regions in 3D, without imposing obvious violations.
Since MBO, like most optimization strategies for structural reconstruction, is non-convex, optimized structures might depend on the random starting configuration of the optimization. We wanted to study this effect by running 100 independent optimizations of chromosome X using different random initialization of the starting configurations. We then calculated the root-mean-square deviation (RMSD) between the resulting superimposed structures, and found a high degree of similarity between all the 100 chromosome X structures, with an average RMSD of ∼ 322 nm, similar to what was reported in [18]. Furthermore, we clustered the RMSD values using hierarchical clustering, and the resulting clusters are visualized in Fig 7. As the figure shows, 4–5 large clusters are found, where the structural similarity within the clusters is clearly higher than between clusters, probably reflecting different local optima in the cost function. By inspecting example structures within each of the clusters, overall the similarity between the structures is high. This indicates that the MBO method gives robust results, with similar structures regardless of starting configuration. Nevertheless, it is advisable to run several independent optimizations, to inspect whether the different local optima in the cost function represents disparate structures.
In S3 Fig, the reconstructed 3D structure from chromosome 1 based on MBO is displayed. We note that, despite the presence of two copies, the reconstructed structure shows few structural violations, with minimum distance violation < 0.01% and connectivity violations below 10%. By performing 100 independent reconstructions, as for chromosome X, (see S4 Fig), the average RMSD was found to be ∼ 262 nm. However, for chromosome 1, the resulting clusters were not as clear as for chromosome X, possibly due to the two separate copies of chromosome 1.
For comparison reasons, we applied MBO using a weighting scheme where the shortest-path completed matrix was used directly without accompanying weights. In S5 Fig, the results from this analysis is shown. As the figure shows, using no weights results in a reduced fraction of correct contacts, and additionally, a higher fraction of connectivity violations. The latter point can be explained by considering that all genomic distances are weighted equally when no weights are used. However, when weights are used, as in the MBO method that we present here, short genomic distances will be weighted more, since these will typically contain more contact information. And as a result, connectivity violations will be reduced.
All in all, we have shown that MBO reconstructs 3D structures consistent with the input chromosomal contact data, at the same computational speed as the popular CMDS approach. Additionally, MBO imposes fewer violations relating to the connectivity of the chain, as well as fewer violations from placing regions too close to each other. We have shown that MBO can be used for routine reconstruction of 3D structures from sparsely sampled data, such as single-cell Hi-C.
In contrast to methods such as MCMC and molecular dynamics, methods aiming at reconstructing a single consensus 3D structure can be utilized quickly and in a high-throughput fashion. One challenge with such approaches, however, has been the lack of possibilities for handling the sparse and noisy interaction frequency matrices in a flexible and robust way. In this paper, we have shown that combining weights with manifold based optimization (MBO) allows for reconstructing 3D structures of genomes, even when data are sparse and noisy, such as for single-cell Hi-C. We have shown that the weights allow for prioritization of interactions where information about spatial positioning is found, while allowing the remaining regions to be positioned in a consistent fashion. Specifically, by comparing the reconstructed and original contact maps, we have shown that the single copy of chromosome X in male mouse cells can be reconstructed in a fashion consistent with the input data. For homologous chromosome pairs, however, reconstruction was not complete, most likely due to considerable structural difference between the two chromosome copies.
We note that it is also possible to run MBO on ensemble Hi-C datasets, since the weighing scheme is equally applicable in this case. However, the assumption of a consensus structure would in this case probably be less justifiable, due to the known inherent variability in chromosome interactions across cells in a large population.
As chromosome conformation capture data are becoming increasingly available [26], quick and robust methods for reconstructing chromosomal 3D structure from chromosomal interaction data, are needed. Additionally, for a complete understanding of the mechanisms involved in gene regulation, cell differentiation, DNA replication and repair, genome organization needs to be studied in its correct dimensions. Efficient and robust 3D genome reconstruction tools such as MBO, are likely to play an increasingly important role for such studies in the future.
A fundamental problem relevant for many applications in various disciplines is to find some coordinates, xi ∈ ℝr, i = 1, ⋯, n in an r-dimensional Euclidean space, given some information about the pair-wise distances between the points. The pairwise distances can be represented by the Euclidean distance matrix (EDM), D ∈ ℝn×n, q
D i j = | | x i - x j | | 2 , (1)
which is an n × n matrix containing the squared distances between the n points. By construction the EDM is a symmetric matrix with zero diagonal and non-negative entries satisfying the triangle inequality D i j ≤ D i k + D k j. Note also that D is invariant to arbitrary rotations and translations of the set of coordinates xi.
If the EDM is known exactly (without noise or missing entries), the coordinates xi can be uniquely determined up to arbitrary rotations and translations by introducing the matrix B ∈ ℝn×n,
B = - 1 2 ( I - 1 n e e T ) D ( I - 1 n e e T ) , (2)
where I ∈ ℝn×n is the identity matrix and e ∈ ℝn is a vector of all ones. If D is a true EDM in an r dimensional space, B is a symmetric positive semidefinite matrix of rank r. That is, B has maximum r nonzero eigenvalues, and B = V Λ VT, where Λ ∈ ℝr×r is the diagonal matrix with the r nonzero eigenvalues of B on the diagonal, and V ∈ ℝn×r is the matrix with the r eigenvectors of B as its columns. It can then be shown that X = V Λ is an n × r matrix with the coordinates xi as its rows. It is easy to see that B = XXT, thus B contains the inner product of the coordinates and is often called the Gram matrix for the set of coordinate vectors.
In many practical applications, however, the EDM may contain noisy and missing values. In this case, finding optimal coordinates xi must be treated as an optimization problem of finding coordinates that minimize some cost function based the known distances. If all pair-wise distances between points are known, but not necessarily accurately, one solution to the optimization problem is given in terms of classical multidimensional scaling (CMDS). CMDS basically solves the optimization problem of finding a matrix B ^ that solves
min B ^ ∈ 𝓢 + n ( r ) | | B ^ - B | | 2 , (3)
where 𝓢 + n ( r ) is the set of positive semidefinite n × n matrices of rank r or less, and B is the matrix derived from the EDM by using Eq (2). This problem has a closed-form solution in terms of the r largest eigenvalues and corresponding eigenvectors of B, namely B ^ = V Λ V T, where Λ is now the diagonal matrix with the r largest eigenvalues of B on the diagonal, and V is the matrix with the corresponding eigenvectors of B as its columns. Consequently, the corresponding coordinates are given by X ^ = V Λ. Obviously, if D is a true EDM for the relevant dimension r, there will be exactly r nonzero eigenvalues and the procedure reduces to the one described in the previous paragraph, and the coordinates are recovered exactly up to arbitrary rotations and translations. However, if D is not close to a true EDM, CMDS is often not robust since the nearest distances are measured through B rather than on D directly.
A manifold based optimization approach for the completion of Euclidean distance matrices was recently presented in Mishra et al. [22]. They solved a minimization problem in the form
min D ^ ∈ 𝓔 n ( r ) 1 2 | | H ⊙ ( D ^ - D ) | | 2 , (4)
where 𝓔n(r) is the set of EDMs with embedding dimension r or less, H is a symmetric weight matrix with binary entries (i.e. a matrix whose elements are either 0 or 1) and where ⊙ denotes the element-wise (Hadamard) product between matrices.
For the application of this approach to the case of the 3D genome reconstruction we have applied a slightly more general framework where the weights are allowed to take any non-negative values (not restricted to 0 and 1). In addition, we choose to minimize the differences between the ordinary Euclidean distances rather than the squared distances used in Eq (4). This choice is motivated by the fact that the longer genomic distances will be weighted less in the final optimization, and results in improved performance compared to using squared distances (see S6 and S7 Figs). Thus, we consider the minimization problem
min D ^ ∈ 𝓔 n ( r ) 1 2 | | H ⊙ ( D ^ - D ) | | 2 , (5)
where square roots here and in the following denote the element-wise square root of the matrix. Following Mishra et al. [22], Eq (5) can alternatively be formulated as an optimization problem on the set of positive semidefinite matrices with fixed rank, denoted 𝓢 +n ( r ), by using the mapping from the set 𝓢 +n ( r ) to the set of EDMs 𝓔n(r) given by
D = κ ( B ) = b e T + e b T - 2 B , (6)
where b is the vector with the diagonal entries of B, i.e b = diag(B) = (B ⊙ I)e. As discussed above a positive semidefinite matrix of rank r admits the factorization B = XXT, where X ∈ ℝn×r and rank(X) = r. Thus, the cost function that we wish to minimize may be written
f ( X ) = 1 2 | | H ⊙ ( κ ( X X T ) - D ) | | 2 . (7)
Note that the X that minimizes Eq (7) is in fact the coordinates that we wish to find.
To minimize Eq (7) we have implemented a solver for the optimization problem in Matlab using the Manopt toolbox [24] using the symfixedrankYYfactory(n, r) manifold, which provides the geometry for the manifold of n × n positive semidefinite matrices with rank r.
Manopt includes a number of different solvers for the optimization problem. Here we will employ a trust-region solver which, unlike steepest descent, utilizes information about both the gradient and the Hessian of the cost function, and has been shown to have good convergence rates. The gradient of f(X) can be written
grad f ( X ) = κ * ( H ( 2 ) ⊙ ( e e T - K ) ) X , (8)
where H(2) = H ⊙ H is the matrix with the squared weights and the matrix K is the symmetric matrix with zero diagonal and off-diagonal entries given by
K i j = D i j κ ( X X T ) i j , i ≠ j . (9)
κ*(B) is the adjoint operator of κ defined by
κ * ( B ) = 2 ( Diag ( B e ) - B ) , (10)
where Diag(v) = (veeT) ⊙ I is the function that returns the n × n matrix with the n × 1 vector v on the diagonal and zeros elsewhere.
In addition to the gradient the trust-region algorithm also requires the Hessian in a given direction U, Hess f(X)[U]. One can show that the Euclidean Hessian of f(X) takes the form
Hess f ( X ) [ U ] = κ * ( H ( 2 ) ⊙ ( e e T - K ) ) U + 1 2 κ * ( H ( 2 ) ⊙ G ⊙ κ ( X U T + U X T ) ) X , (11)
where G is the symmetric matrix with zero diagonal and off-diagonal entries
G i j = D i j ( κ ( X X T ) i j ) 3 , i ≠ j . (12)
The conversion from the Euclidean to the Riemannian Hessian, needed for the optimization algorithm, is performed internally in Manopt. For additional details about the manifold based algorithm, see [22, 24].
From the known 3D structure. a true EDM was constructed containing the pair-wise squared distances between all the 600 kbp sized bins. To model the uncertainty and possible sparsity of distance information inferred from chromosomal contact data such as Hi-C, the original distance matrix was contaminated by adding random noise as well as randomly removing a given percentage of the distances. That is, from the original Euclidean distance matrix D (containing the squared pair-wise distances), a noisy and incomplete set of distances δij is generated as
δ i j = δ j i = D i j | 1 + σ ϵ i j | , for ( i , j ) ∈ 𝓝 (13)
where ϵij are sampled randomly from a standard normal distribution and where 𝓝 is the set of entries (i, j) for which the distances are available.
Tests were run for different values for the noise level σ and ratio of missing distances (size of 𝓝).
The raw results from a single-cell Hi-C experiment typically lists a number of observed contacts between specific genome positions. From the raw results, the contacts were aggregated into equally spaced bins along the chromosomes. For the results presented here a bin size of 50 kbp was used. Then all observed contacts were assigned to their corresponding bins. In the case that multiple contacts fell into the same bin, the duplicate entries were ignored so that a binary contact matrix Cij was obtained for each chromosome. Hence, Cij = 1 represents a Hi-C contact between bins i and j, while Cij = 0 represents the absence of a contact.
To use the MBO approach, the binary contact map must be converted into a distance matrix Dij. First a target distance dc is assigned to all bins with an observed Hi-C contact. Secondly, the connectivity along the chromosome is taken into account by assigning a distance dn to neighboring bins. Hence, as a first step the following matrix is constructed
D i j = { d c if C i j = 1 , d n if C i j = 0 and | i - j | = 1 , 0 elsewhere . (14)
Since the MBO method works also for incomplete distance matrices, the optimization could in principle be run directly on Eq (14), letting the weights Hij be nonzero only for the nonzero entries of Dij. However, since only the local distances (contacts and neighboring bins) are known, a direct optimization of Eq (14) would lead to a very compact structure where all bins are located close together. Hence, for a consistent 3D structure some information about the large distances must be included. One possible method is to assign large distances and small weights to the non-interacting bins (see e.g. [27, 28]). The large distances will then act as a repulsive force and counteract the formation of a compact state. Another possibility is to apply the shortest-path method to fill the missing entries of the distance matrix. In this way the missing distances may take more realistic values since they are deduced directly from the known distances. However, these shortest path-distances still introduce noise that may seriously influence the result. Motivated by the fact that the shortest-path derived distances are more noisy than the ‘original’ contact-distances that we wish to satisfy, we have adopted a slightly more flexible approach where we combine the shortest-path completed matrix with weights so that the shortest-path inferred distances are weighted less in the optimization procedure.
Thus, we first replace the zero entries in Dij with the shortest-path derived distances. We then introduce the weight matrix Hij whose elements are chosen to be inverse proportional to the number of edges traversed in the shortest path, i.e H i j = n i j − q where nij is the number of edges that is needed to connect node i and j. That is, the original distances will have weights equal to one, while the shortest-path derived distances will have smaller weights. The value q is a factor that specifies the relative magnitude of the weights for the non-observed edges compared to the observed ones, and was found by maximizing the percent correct contacts and minimizing distances violations (see S8A Fig for an example). In our case this value was always found to be between 1 and 3 (see S1 File), but in theory, for other data, the optimal value may be outside this range. Here, we have used a simple optimization scheme by trying out a range of values for q. This is likely sufficient in most cases, since the effect of using different values for q on the final structures is not very large. For example, on chromosome X for cell 1, using a range of values of q between 0 and 3, the structures all had RMSD<300nm compared to the structure with optimized q (see S8B Fig).
MBO is initialized by starting with a random initial configuration (a random point on the manifold), and convergence is considered obtained if the cost function or the norm of the gradient drops below a small value (1e-20 and 1e-08, respectively). After a successful convergence of the optimization algorithm the resulting coordinates xi are scaled to best agree with the original contact map. That is, we search for a scaling constant cl so that D ^ i j = ∣ ∣ c l x i − c l x j ∣ ∣ contains exactly nc pair-wise distances smaller than the contact distance dc, where nc is the number of contacts in the original contact matrix. Note that in the case of perfect agreement, the contact matrix derived from the coordinates cl xi will be identical to the original contact matrix, since the number of entries are the same. The optimal value for cl is found by a simple binary search method.
The percent correct contacts was calculated by direct comparisons of original and reconstructed contact matrices. Minimum distance violations were defined as the percent fraction of pairwise distance below 30 nanometers. Connectivity violations were defined as the percent fraction of neighboring (connected) bins with a distance above 200 nanometers. In Eq 14, dc = 60nm, dn = 120nm.
MBO is implemented in Matlab, and is based on the Manopt software [24]. Code is freely available at http://folk.uio.no/jonaspau/mbo/.
|
10.1371/journal.ppat.1005450 | Resistance to Bacillus thuringiensis Mediated by an ABC Transporter Mutation Increases Susceptibility to Toxins from Other Bacteria in an Invasive Insect | Evolution of pest resistance reduces the efficacy of insecticidal proteins from the gram-positive bacterium Bacillus thuringiensis (Bt) used widely in sprays and transgenic crops. Recent efforts to delay pest adaptation to Bt crops focus primarily on combinations of two or more Bt toxins that kill the same pest, but this approach is often compromised because resistance to one Bt toxin causes cross-resistance to others. Thus, integration of Bt toxins with alternative controls that do not exhibit such cross-resistance is urgently needed. The ideal scenario of negative cross-resistance, where selection for resistance to a Bt toxin increases susceptibility to alternative controls, has been elusive. Here we discovered that selection of the global crop pest, Helicoverpa armigera, for >1000-fold resistance to Bt toxin Cry1Ac increased susceptibility to abamectin and spineotram, insecticides derived from the soil bacteria Streptomyces avermitilis and Saccharopolyspora spinosa, respectively. Resistance to Cry1Ac did not affect susceptibility to the cyclodiene, organophospate, or pyrethroid insecticides tested. Whereas previous work demonstrated that the resistance to Cry1Ac in the strain analyzed here is conferred by a mutation disrupting an ATP-binding cassette protein named ABCC2, the new results show that increased susceptibility to abamectin is genetically linked with the same mutation. Moreover, RNAi silencing of HaABCC2 not only decreased susceptibility to Cry1Ac, it also increased susceptibility to abamectin. The mutation disrupting ABCC2 reduced removal of abamectin in live larvae and in transfected Hi5 cells. The results imply that negative cross-resistance occurs because the wild type ABCC2 protein plays a key role in conferring susceptibility to Cry1Ac and in decreasing susceptibility to abamectin. The negative cross-resistance between a Bt toxin and other bacterial insecticides reported here may facilitate more sustainable pest control.
| The soil bacterium Bacillus thuringiensis (Bt) produces proteins that kill insect pests but do not harm most other organisms including people. Extensive use of Bt proteins in sprays and genetically engineered crops selects for rapid evolution of resistance in pests, reducing economic and environmental advantages of this alternative to conventional insecticides. We discovered that resistance to Bt toxin Cry1Ac in the invasive crop pest Helicoverpa armigera increased susceptibility to abamectin and spineotram, insecticides derived from two other soil bacteria. Both resistance to Cry1Ac and increased susceptibility to abamectin are linked with the same mutation in a gene encoding a transporter protein ABCC2. The results imply that negative cross-resistance occurs because the wild type ABCC2 protein plays a key role in conferring susceptibility to Cry1Ac and in decreasing susceptibility to abamectin. The negative cross-resistance between a Bt toxin and other bacterial insecticides reported here may facilitate more sustainable pest control.
| Insecticidal proteins from the bacterium Bacillus thuringiensis (Bt) are used widely in sprays and transgenic plants to control insects that attack crops and vector diseases [1,2]. These Bt proteins are especially valuable because they kill some devastating pests, but are not toxic to humans and most other organisms [1,3–6]. Farmers planted corn, cotton and soybean genetically engineered to produce Bt proteins on 78 million hectares worldwide in 2014, with a cumulative total of 648 million hectares of Bt crops planted since 1996 [2]. In the United States, transgenic Bt plants accounted for 80% of the corn and 84% of the cotton grown in 2014 [7]. Although Bt crops have provided substantial economic and environmental benefits [1,8–12], evolution of pest resistance to Bt proteins can diminish or even eliminate these advantages [13–17].
To delay pest adaptation, many farmers have switched from transgenic crops producing only one Bt toxin to newer ones producing two or more Bt toxins that kill the same pest [18]. This “pyramid strategy” aims to use toxins sufficiently different so that evolution of resistance to one toxin does not confer cross-resistance to the others [18,19]. Unfortunately, cross-resistance is common between Bt toxins, often strong between closely related toxins and weak, yet generally positive, between more distantly related toxins [16,18].
Because cross-resistance occurs between many Bt toxins, increasing the sustainability of Bt crops requires integration of more diverse pest management tactics that are not undermined by cross-resistance. The ideal scenario is negative cross-resistance, where selection for resistance to a Bt toxin increases susceptibility to alternative controls [20,21]. When negative cross-resistance occurs, the alternative control imposes a fitness cost associated with Bt resistance that selects against Bt resistance [22]. Despite recognition that negative cross-resistance could greatly enhance sustainability, identifying practical, alternative controls that show negative cross-resistance with Bt toxins has remained elusive [20,21].
Among the diverse mechanisms causing resistance to Bt toxins [23,24], mutations disrupting the ATP-binding cassette (ABC) transporter protein named ABCC2 are genetically linked with resistance in strains of at least seven species of Lepidoptera including Helicoverpa armigera [25–28]. This insect recently invaded the New World and is one of the world’s most damaging crop pests. Previous work revealed a mutation in the gene encoding ABCC2 tightly linked with >1000-fold resistance to Bt toxin Cry1Ac in a laboratory-selected strain (LF60) relative to its unselected parent strain (LF) [28]. These results support the conclusion that the wild type ABCC2 plays an essential role in the mode of action of Cry1Ac against lepidopteran larvae [29,30]. Because the mutation in LF60 is expected to cause the loss of 143 amino acids, it is also likely to disrupt the normal function of ABCC2
In addition to recent reports implicating ABCC2 in susceptibility to Bt toxins, extensive evidence shows that many members of the superfamily of ABC transporter proteins protect cells by excreting xenobiotics, including ABC transporters that confer resistance to drugs and chemotherapy agents in humans and resistance to insecticides in arthropods [30–32]. Because some ABC transporter proteins protect cells from insecticides, we hypothesized that the disruption of ABCC2 conferring resistance to Cry1Ac would also increase its susceptibility to other insecticides. To test this hypothesis, we evaluated responses of the LF60 and LF strains of H. armigera to three conventional insecticides and two bacterial insecticides, abamectin (an avermectin) and spineotram (a spinosyn). Abamectin and spineotram are widely used neurotoxic insecticides derived from the soil bacteria Streptomyces avermitilis and Saccharopolyspora spinosa, respectively [33–35]. Abamectin acts via various ligand-gated ion chloride channels, such as glutamate-gated chloride channels [33,35,36] and spineotram acts on a subgroup of nicotinic acetylcholine receptors [34].
We discovered that laboratory selection for resistance to Cry1Ac increased susceptibility of the LF60 strain by 9.0-fold to abamectin and 2.6-fold to spinetoram, but did not affect susceptibility to endosulfan (a cyclodiene), phoxim (an organophosphate), or cyhalothrin (a pyrethroid). Analyses of inheritance, transcription, and abamectin concentration in larvae and in transfected Hi5 cells support the hypothesis that ABCC2 mediates the observed negative cross-resistance between Cry1Ac and abamectin.
Selection for resistance to Cry1Ac significantly increased susceptibility to abamectin and spineotram of the LF60 strain of H. armigera relative to its unselected, Cry1Ac-susceptible parent strain LF (Tables 1 and 2). Relative to LF, the concentration killing 50% of larvae (LC50) for LF60 was 9.0 times lower for abamectin (Table 1) and 2.6 times lower for spineotram (Table 2). Selection of LF60 with Cry1Ac did not affect susceptibility to endosulfan, phoxim, or cyhalothrin (Table 2). The LC50 of abamectin was not higher for LF (1.23) than for the independently derived susceptible strain 96S (1.31), confirming that LF was not resistant to abamectin.
We conducted reciprocal crosses between LF and LF60 and tested the F1 progeny to evaluate inheritance of increased susceptibility to abamectin. The responses were similar between the F1 progeny from the two reciprocal crosses (Table 1), indicating that inheritance of susceptibility to abamectin is autosomal (i.e., no sex linkage or maternal effects). Based on the LC50 values for the F1 relative to the parent strains (Table 1), we calculated the dominance parameter (h), which varies from 0 for completely recessive to 1 for completely dominant [37]. For the two reciprocal crosses, h was 0.34 and 0.36, indicating that increased susceptibility to abamectin was a partially recessive trait.
Analysis of F2 progeny shows that increased susceptibility to abamectin is genetically linked with the ABCC2 mutation that confers resistance to Cry1Ac (Fig 1 and S1 Table). We tested progeny from five single-pair F2 families that fed on either untreated diet or diet containing abamectin and sequenced genomic DNA individually for 20 survivors from each diet type. On untreated diet, the observed genotype frequencies at the HaABCC2 locus were 0.24 ss: 0.52 rs: 0.24 rr (Fig 1 and S3 Table), which do not differ significantly from the frequencies expected under Mendelian inheritance (0.25 ss: 0.50 rs: 0.25 rr, Chi-squared = 0.08, df = 2, P = 0.96). However, on abamectin-treated diet, the genotype frequencies were 0.73 ss: 0.27 rs: 0.00 rr, which differ significantly from the expected frequencies (Chi-squared = 55, df = 2, P < 0.0001) (Fig 1 and S3 Table). Excluding the rr, which had 0% survival on treated diet, the results also show that relative to ss, the rs genotype was significantly lower than expected on treated diet (Chi-squared = 39, df = 1, P < 0.0001). Overall, the low survival on treated diet of rs and rr relative to ss demonstrates a strong association between increased susceptibility to abamectin and the r allele of HaABCC2 that confers resistance to Cry1Ac.
Reversion transcription-polymerase chain reaction (RT-PCR) analysis of the LF strain showed that transcription of HaABCC2 in 5th instar larvae was much higher in the foregut, midgut and hindgut than in Malphigian tubules or cuticle (S1 Fig). Transcription of HaABCC2 did not differ significantly between 4th and 5th instars, and was significantly higher in 4th instars than in 1st to 3rd instars, pupae, and moths (S1 Fig).
For the Cry1Ac-susceptible strain LF, HaABCC2 transcription was suppressed more than 50% for larvae fed droplets of water containing dsHaABCC2 relative to larvae fed control droplets with water only or dsGFP (Fig 2). However, this treatment of LF larvae with dsHaABCC2 did not affect transcription of two other genes, HaABCC3 and HaCAD (S2 Fig). This suppression of HaABCC2 transcription in LF larvae by RNAi significantly decreased susceptibility to Cry1Ac and increased susceptibility to abamectin (Fig 3 and S3 Fig). On untreated diet (control), treatment with dsHaABCC2 did not affect survival (S4 Fig).
We hypothesized that the mutant HaABCC2 in LF60 increased susceptibility to abamectin by interfering with removal of abamectin. Consistent with this hypothesis, after larvae fed on diet containing abamectin, the concentration of abamectin in midgut tissues was significantly higher for LF60 (with mutant HaABCC2) than LF (with wild type HaABCC2), with a 2-fold difference after 24 h and a 4-fold difference after 48 h (Fig 4).
To test effects of HaABCC2 directly, we transformed Hi5 cells with hybrid genes containing the GFP gene fused at the C-terminus of either the wild type HaABCC2 gene from LF or the mutant HaABCC2 from LF60 (HaABCC2-GFP and mHaABCC2-GFP, respectively). We also transfected Hi5 cells with only the GFP gene as a control. Western blots confirmed production of HaABCC2-GFP, mHaABCC2-GFP, and GFP respectively, in cells transfected with each of the three genes (S5 Fig).
Transfection of Hi5 cells with HaABCC2-GFP, but not with mHaABCC2-GFP or GFP, conferred susceptibility to Cry1Ac (Figs 5 and S6). We also found that after treating Hi5 cells with abamectin for 12 h, the concentration of abamectin was significantly lower in cells transfected with HaABCC2-GFP than in cells transfected with either mHaABCC2-GFP or GFP (Fig 6). These results are consistent with the results summarized above for the LF and LF60 strains implying that, relative to mutant HaABCC2, wild type HaABCC2 increased susceptibility to Cry1Ac and the efflux of abamectin.
The results reported here show that selection for resistance to Bt toxin Cry1Ac in the LF60 strain of H. armigera increased its susceptibility to two other bacterial insecticides, abamectin and spineotram, but not to three conventional insecticides (endosulfan, a cyclodiene; phoxim, an organophosphate; and cyhalothrin; a pyrethroid) (Tables 1 and 2). Whereas previous work showed that the resistance to Cry1Ac in LF60 is tightly linked with a mutant allele that disrupts the ABC transporter protein ABCC2 [28], the new results here show that increased susceptibility to abamectin in LF60 is autosomal, partially recessive (mean h = 0.35), and linked with the same mutant allele (Fig 1 and S3 Table). Moreover, suppressing transcription of HaABCC2 with RNAi both decreased susceptibility to Cry1Ac and increased susceptibility to abamectin (Fig 3). In addition, susceptibility of Hi5 cells to Cry1Ac was conferred by transfecting them with the wild type HaABCC2 gene from the Cry1Ac-susceptible LF strain, but not with the mutant HaABCC2 gene from LF60 (Figs 5 and S5). Finally, after exposure to abamectin, the concentration of abamectin was higher in midgut tissues of LF60 than LF, and higher in cells transfected with mutant HaABCC2 from LF60 than in cells transformed with wild type HaABCC2 (Figs 4 and 6). Collectively, these results imply that negative cross-resistance occurs in this case because the wild type ABCC2 protein in LF plays a key role in conferring susceptibility to Cry1Ac and decreasing susceptibility to abamectin by reducing the abamectin concentration in the midgut. Conversely, the evidence also suggests that the mutant ABCC2 protein in LF60 confers resistance to Cry1Ac by disrupting the mode of action of this Bt protein [28], while increasing susceptibility to abamectin by interfering with removal of abamectin by ABCC2.
An alternative hypothesis is that the decreased susceptibility to abamectin in LF relative to LF60 was caused by increased detoxification of abamectin in LF relative to LF60. We evaluated this alternative hypothesis because the results of Chen et al. [38] suggest that increased activity of some general detoxification enzymes accounted for a small portion of the 822-fold resistance to abamectin generated by their laboratory selection of H. armigera with abamectin. However, LF was not resistant to abamectin relative to the susceptible 96S strain (Table 1). Unlike the abamectin-selected strain analyzed by Chen et al. [38], LF had been reared in the laboratory without exposure to any insecticide for more than 15 years and was unlikely to have increased detoxification activity. In addition, relative to LF60, the LF strain did not have significantly decreased susceptibility to endosulfan, phoxim, or cyhalothrin (Table 2), which does not support the idea of increased general detoxification activity in LF relative to LF60. Most importantly, our analyses of genetic linkage, suppression of transcription by RNAi, and residual abamectin concentration in larvae and in transfected Hi5 cells provide strong evidence that ABCC2 mediates the observed negative cross-resistance between Cry1Ac and abamectin.
In related work from Australia, resistance to Bt toxin Cry2Ab in H. armigera was associated with a 1.5-fold increase in susceptibility to emamectin benzoate, an insecticide derived from abamectin, and a five-fold increase in susceptibility to an organophosphate (chlorpyrifos) and a carbamate (methomyl) [39]. Similarly, resistance to Cry2Ab in Helicoverpa punctigera was associated with increases in susceptibility of 1.2-fold to abamectin, 1.8-fold to chlorpyrifos, and 3.9-fold to methomyl [39]. In both species, resistance to Cry2Ab is genetically linked with mutations in the ABC transporter protein ABCA2 [40]. These results suggest that, similar to disruption of ABCC2 conferring resistance to Cry1Ac and negative cross-resistance to abamectin, the disruption of ABCA2 conferring resistance to Cry2Ab may cause the observed weak negative cross-resistance to avermectins (abamectin and emamectin benzoate) and stronger negative cross-resistance to chlorpyrifos and methomyl.
In a related study from China, susceptibility to emamectin benzoate ranged from 4.5 to 11-fold higher in 16 field populations sampled in 2011 relative to the Cry1Ac-susceptible laboratory strain named SCD [41]. Many of these field populations had been exposed to Bt cotton producing Cry1Ac, and several had minor, but statistically significant resistance to Cry1Ac relative to SCD [17,42].
Because disruption of ABCC2 is only one of the many mechanisms of resistance to Cry1Ac in H. armigera [17,24,28,43–45], the extent of negative cross-resistance between Cry1Ac and avermectins such as abamectin or emamectin benzoate may vary among populations or even within a population over time. Nonetheless, avermectins are considered effective as complements or alternatives to Bt toxins for control of this global crop pest because of their unique mode of action [46] and the typical absence of positive cross-resistance to avermectins caused by resistance to Bt toxins, organophosphates, carbamates, or pyrethroids [39,41,47,].
More generally, it will be important to determine if the ABCC2-mediated resistance to Cry1Ac in other major pests [25,26,48,49] causes negative cross-resistance to abamectin or other insecticides. It will also be useful to find out if the disruption of ABC transporters that confers resistance to Bt toxins increases insect susceptibility to other xenobiotics, such as plant defensive compounds. Because ABC transporters play a vital role in protecting insects from xenobiotics, their evolution may be constrained to preserve this function. The negative cross-resistance seen here between Bt toxin Cry1Ac and two other insecticides derived from soil bacteria, abamectin and spineotram, raises the intriguing possibility that Bt bacteria have exploited this constraint by targeting ABC transporter proteins and thereby delaying evolution of resistance to Bt toxins in their hosts. This delay could occur because the increased susceptibility to poisons from other bacteria such as Streptomyces avermitilis and Saccharopolyspora spinosa might be a major fitness cost associated with resistance to Bt toxins [22]. In any case, the negative cross-resistance reported here may facilitate more sustainable control of H. armigera and other pests because selection for resistance to Bt crops could increase their susceptibility to some insecticides.
The LF strain was started with larvae collected from Bt cotton in Langfang, Hebei Province, China in 1998 and was reared in the laboratory on artificial diet without exposure to Bt toxins or insecticides [50]. LF60 was generated by selecting insects from the susceptible LF strain with MVPII (Dow AgroSciences), a commercial formulation of Cry1Ac protoxin incorporated in the diet [45]. Selection was conducted for more than a decade with progressively increasing concentrations: 1, 5, 10, 30 and 60 μg Cry1Ac protoxin per g diet [45]. As an internal control, we also tested the susceptible strain 96S, which was started in 1996 from adults collected from conventional cotton in Xinxiang, Henan Province, China and reared in the laboratory on artificial diet without exposure to Bt toxins or insecticides [28, 50]. Rearing and bioassays were conducted at 27 ± 2°C, a photoperiod of 14L:10D, and 75 ± 10% relative humidity.
We used diet incorporation bioassays to test 2nd instars against five insecticides [45] (purity and source given in parentheses): abamectin, cyhalothrin, and endosulfan (95%, 96%, and 90%, respectively, the pesticide factory of Institute of Plant Protection, Chinese Academy of Agricultural Sciences, Beijing, China), spineotram and phoxim (6% and 90%, BaySystems, Leverkusen, Germany). Spineotram was dissolved in distilled water and the other four insecticides were dissolved first in dimethyl sulfoxide (DMSO), and then in distilled water. For controls, we incorporated distilled water into diet for spineotram, and DMSO dissolved in distilled water for the other insecticides.
Larvae from the LF and LF60 strains were tested against a series of concentrations of each of the five insecticides. In addition, the F1 progeny from a cross between LF and LF60 were tested against abamectin. Mortality was recorded after 3 days and analyzed using probit methods (for details see Statistical Analysis below).
We obtained 40 males and 40 female virgin moths from the LF and LF60 strains respectively, and set up two reciprocal mass crosses: 40 ♂LF x 40 ♀LF60 and 40 ♂LF60 x 40 ♀LF in plastic crates (5 L). Adults were allowed to mate and deposit eggs onto oviposition gauze. Newly hatched F1 neonates were transferred to individual 24 well-plate for bioassay.
We tested progeny from five single-pair F2 families, each generated by crossing a female F1 moth and a male F1 moth in paper cup (350 mL). We reared F2 larvae on untreated diet until they were early 2nd instars and split them into two groups. For the next 3 days, one group of 50 larvae was provided untreated diet and the other group of 200 larvae was provided diet treated with 3 μg abamectin per ml. For 20 survivors from each type of diet, we sequenced genomic DNA of each individual (primers GP-F: TACTCTGCGATACAATTTGGA and GP-R: AGTACCACCTTCAGCTACTTT) at the HaABCC2 locus to distinguish between the previously identified r allele harboring a mutation that confers resistance to Cry1Ac and the s allele that confers susceptibility to Cry1Ac [28].
We compared transcription of HaABCC2 among tissues and development stages in the susceptible LF strain. For 5th instar larvae, we compared the following five types of tissues: foregut, midgut, hindgut, malphigian tubules, and cuticle. For each biological replicate of each tissue type, we pooled tissues dissected under a dissecting microscope from 20 larvae. We also compared whole insects at the following five developmental stages (with age and sample size for each biological replicate in parentheses): 1st instar (1 day, n = 80), 2nd instar (3 days, n = 40), 3rd instar (6 days, n = 30), 4th instar (9 days, n = 15), 5th instar (12 day, n = 10), pupae (4 days, n = 10), and adults (5 days, virgins, n = 10). All dissected tissues from 5th instars and whole insects representing different developmental stages were quickly frozen in liquid nitrogen and stored at -80°C for subsequent RNA extraction. We used three biological replicates for each tissue type and development stage.
Total RNA was extracted from the sample homogenates according to the standard TRIzol reagent protocols (Invitrogen). RNA purity and concentration was evaluated by 260/280 and 260/230 ratios measured in a NanoDrop 3300 (Thermo). Total RNA (4 μg) was used for reverse transcription using SuperScript III First-Strand Synthesis for RT-PCR (Invitrogen), according to the manufacturer’s instructions in a final volume of 20 μl. The cDNA was diluted in nuclease free water for immediate use in qPCR or stored at -20°C.
We used quantitative real time PCR (qRT-PCR) to evaluate relative transcription of ABCC2 [28]. The thermal program of qPCR was 40 cycles of 95°C for 15 sec and 60°C for 34 sec. Gene expression was normalized relative to the reference genes of β-actin (Accession no. EU527017.1) and glyceraldehyde-3-phosphate dehydrogenase (GAPDH) (Accession no. JF417983.1). We calculated the relative copy number of mRNA using the method of 2-ΔΔCt [51]. Primers and probes are shown in S4 Table.
We used RNA interference (RNAi) technique to test the role of HaABCC2 in susceptibility to Cry1Ac and abamectin in larvae from the susceptible LF strain [24]. We fed early third instars with dsRNA twice in 48 hours (At 0 hours and 48 hours) and determined transcript levels using qRT-PCR as described above at 24, 72 and 120 h after dsRNA-feeding at the second time. Feeding with water and dsRNA of green fluorescent protein (dsGFP) gene were used as controls.
The dsRNA was prepared using PCR products as template by in vitro transcription for RNAi. The primers for HaABCC2 gene were: ABCC2 RNAiF 5’-TAA TAC GAC TCA CTA TAG TGG GCG ACT TTG GTG ATT TG’-3, ABCC2 RNAiR 5’-TAA TAC GAC TCA CTA TAT TTG ATG CTG CCG CTT ATG T’-3, and the green fluorescent protein (GFP) gene were: GFP RNAiF 5’-TAA TAC GAC TCA CTA TAG TCA AAG ATG ACG GGA ACT AC’-3, GFP RNAiR 5’-TAA TAC GAC TCA CTA TAC AAA CTC AAG AAG GAC CAT G’-3. T7 primer sequence was placed in front of both forward and reverse primers. In vitro transcription to produce dsRNA of HaABCC2 and GFP genes were performed with T7 RNA polymerase using the HiScribe RNAi T7 In Vitro Transcription Kit (New England Biolabs) as reported by the manufacturer.
The dsRNA was used to feed 3rd instar larvae of H. armigera as follows: larvae were individually placed into each well of 24-well plates to avoid cannibalism and starved for 12 h. Seventy two larvae were then fed with a 2 μl DEPC water drop containing 2.5 μg dsRNA from either HaABCC2 and GFP respectively. After 2 h, droplet-fed larvae were placed back individually into 24-well plates provided with artificial diet. Two days later the dsRNA oral delivery was done one more time with a water drop containing 5 μg dsRNA as reported above. After feeding the dsRNA-fed larvae were placed individually into a 24-well plate provided with artificial diet with either 0 or 60 μg Cry1Ac/ml diet. To evaluate the silencing method, we assessed transcription of HaABCC2 after 1, 2, and 3 days of dsRNA feeding. One insect was collected per sample, and quantitative RT–PCR was performed as described above. Each experiment and treatment consisted of five biological replicates. For these bioassays the 24 larvae previously fed with dsRNA were transferred individually into each vial containing diet either untreated diet or diet treated with 60 μg Cry1Ac protoxin/ml diet or 3 μg abamectin/ml diet and maintained under the rearing conditions described above. Survival, pupation and eclosion were scored every day.
We provided individual middle 4th instar larvae from LF and LF60 with 0.125g portions of diet containing 3 μg abamectin/ml. After all of the treated diet was eaten 6 h later, each larva was transferred to untreated diet for 24 h to make the larvae ingest the abamectin. 50 larvae were dissected and midgut of each larvae was taken out and washed in normal saline. After drying with absorbent paper, midgut were put in glass homogenizer (type76, Haimen BoTai, China) for fully homogenization, one min per homogenate and then put on ice cooling for one min. Homogenization was repeated until the midguts were completely smooth. 0.5g homogenated midgut tissue were transferred to a centrifuge tube (10 ml). We added 1 ml ultra-pure water and 2 ml acetonitrile, vortexing for 2 min added 0.5g NaCl and 1g anhydrous MgSO4 to the centrifuge tube, vortexing for 2 min. High speed (3500 RCF) centrifugation at 4°C for 5 min. The upper layer of the prepared sample was filtered using a 0.22 μm nylon syringe filter and transferred to an autosampler vial for injection. Area of abamectin in each injection were collected and used to calculate the concentration of abamectin in midgut of different insect lines and times according to the method of Du et al. [52] under the direction of Dr. Du in his laboratory. The analytical standard abamectin (95%) were from the pesticide factory of Institute of Plant Protection, Chinese Academy of Agricultural Sciences, Beijing, China. Chromatography grade acetonitrile and methanol were purchased from Honeywell International Inc. (New Jersey, USA). Acetonitrile for pesticide residue analysis was of analytical grade and purchased from Beijing Chemical Reagent Company (Beijing, China). Analytical grade sodium chloride (NaCl) and anhydrous magnesium sulfate (anhydrous MgSO4) were purchased from Beijing Chemical Company (Beijing, China). Ultra-pure water was obtained from a Milli-Q system (Bedford, MA, USA) [52]. The data was repeated in triplicate. Abamectin with the concentration of 1μg/ml was used for standard sample.
Trichoplusia ni BTI-Tn-5B1-4 (Hi5) cells were grown in Grace’s insect cell culture medium (Life Technologies Co., Grand Island, NY, USA) with 10% fetal bovine serum (FBS) (Life Technologies Co., Australia), 100 unit/ml penicillin 100 μg/ml streptomycin (Life Technologies Co., Grand Island, NY,USA) at 28°C [53].
Both HaABCC2 from the susceptible larvae and the resistant strains were amplified by PCR using the specific primers (HaABCC2F: 5’-CTC AAG CTT CGA ATT CGC CAC CAT GGA AAA CGG TAC TAG TCC-3’; HaABCC2R: 5’-CCG CGG TAC CGT CGA CTG ACC GCC TCC GCC ACC GCC GTG GTG GTG GTG GTG GTG CT- 3’) and the corresponding pGEM T vectors inserted with the two genes as template. The fragments digested with EcoR I and Sal I were then inserted into plasmid pie2-EGFP-N1 at the corresponding sites, to generate plasmids pHaABCC2-GFP and pmHaABCC2-GFP. Both plasmids contained the pie2 promotor and the gene encoding GFP; pHaABCC2-GFP had the gene from LF encoding wild type ABCC2 and pmHaABCC2-GFP had the gene from LF60 encoding the mutant ABCC2. Hi5 cells were seeded into well of 6-wells culture plates at the density of 1× 106 cells/well and grown overnight in Grace´s insect cell culture medium. We transfected each plasmid into Hi5 cells at 2μg/well using cell-fectin reagent (Life Technologies). At 24 h after transfection, the cells expressing recombinant proteins were washed with phosphate buffered saline (PBS), fixed in 4% paraformaldehyde for 20 min and stained with Hoechst 33342 (1 μg/ml) for 20 min at room temperature. The images were taken with a Nikon fluorescence microscope (E400).
The Hi5 cells were seeded into 6-wells culture plate and transfected with plasmids pHaABCC2-GFP, pmHaABCC2-GFP and pGFP as described above, respectively. Cells were harvested by centrifugation and subjected to protein extraction at 24 h post transfection. The proteins were separated on 8% SDS-PAGE gel. After electrophoresis, the proteins were transferred onto PVDF membrane (Millipore Corporation, Billerica, MA, USA). The membrane was blocked with 5% non-fat milk in TBS-T buffer for 2 h at room temperature, and then incubated with mouse anti-GFP antibody (Abcam, Cambridge, UK) 1: 3000 dilution in TBS for 2 h at room temperature. The membranes were then washed for three times with TBS-T and then incubated with fluorescent secondary antibody 1:5000 dilution (Earthox, San Francisco, CA, USA). Finally, the membranes were washed for three times with TBS-T and bands were visualized using the Odyssey system (LI-COR Bioscience, Lincoln, NE, USA).
Hi5 cells cultured into 96-well plates were transfected with plasmids pGFP, pHaABCC2-GFP and pmHaABCC2-GFP, respectively, according to the method [53]. The cells were washed twice with PBS at 24 h post transfection, then treated with the activated Cry1Ac toxin at the indicated concentrations for 1 h. The activated toxin was obtained by digesting protoxin (The State Key Laboratory for Biology of Plant Disease and Insect Pests, Institute of Plant Protection, Chinese Academy of Agricultural Sciences) [54] with trypsin (Sigma) at 37°C for 2 h at 20:1 (protoxin: trypsin) mass ratio.
The treated cells were observed and photographed under an inverted fluorescence microscope. The percentage of aberrant (swollen) cells were calculated according to the method of Tanaka et al. [55].
Hi5 cells were seeded into T75-flask and grown over night. Then the cells were transfected with plasmids pGFP, pHaABCC2-GFP and pmHaABCC2-GFP, respectively, and grown over night. Abamectin dissolved in DMSO was added into the flask at the final concentration of 5 μg abamectin/ml and incubated with the cells for 12 h. Then the cells were washed three times with PBS and collected by centrifugation at 500 g and 10 min. 0.05g cells were thawed five times and then transferred to a centrifuge tube (2 ml), we added 0.1 ml ultra-pure water and 0.2 ml acetonitrile, vortexing for 2 min. added 0.05g NaCl and 0.1g anhydrous MgSO4 to the centrifuge tube, vortexing for 2 min. High speed (3500 RCF) centrifugation at 4°C for 5 min. The upper layer of the prepared sample was filtered using a 0.22 μm nylon syringe filter and transferred to an autosampler vial for injection. The residuals of abamectin were then detected with the method of residuals of abamectin of insect midgut as described above. Area of abamectin in each injection were collected and used to calculate the concentration of abamectin in insect cells. These data were repeated in triplicate.
We used the SPSS Statistics (version 20.0) software (SPSS Inc.) to estimate the concentrations of insecticide killing 50% of larvae (LC50) and its 95% fiducial limits, and the slope of the concentration-mortality line and its standard error (Tables 1 and 2) [48]. Two values of LC50 were considered significantly different if there was no overlap between their 95% fiducial limits, which is a standard, but conservative criterion [56]. Because this criterion is conservative and this approach depends on fit of the data to the Probit model, we also analyzed the bioassay data using Fisher’s exact test (http://graphpad.com/quickcalcs/contingency2/), which is not conservative and does not rely on the Probit model. For each pairwise comparison, the conclusion about statistical significance did not differ between these two statistical approaches (S2 Table). We calculated the dominance parameter h from LC50 values as described previously [32]. In the genetic linkage analysis, we used chi-squared tests (http://vassarstats.net/newcs.html) to determine if the observed genotype frequencies differed significantly from the expected genotype frequencies (0.25 rr, 0.50 rs, and 0.25 ss) on untreated diet and on diet treated with abamectin. We used one-way analysis of variance (ANOVA) followed by Duncan’s multiple range test Statistica 6 (Statistica, SAS Institute Inc., Cary, NC, USA) to determine if means differed significantly among treatments in several experiments: among water, dsGFP, and dsHaABCC2 treatments for relative transcription of HaABCC2 (Fig 2) and survival on diet treated with Cry1Ac, treated with abamectin, or untreated diet (control) (Figs 3, S3 and S4); among genes used for transfecting Hi5 cells (wild type HaABCC2, mutant HaABCC2, and GFP) in their effects on survival and abamectin concentration (Figs 5 and 6), among tissue types and developmental stages for transcription of HaABCC2 (S1 Fig).
|
10.1371/journal.pntd.0002761 | The Effect of Multiple Rounds of Mass Drug Administration on the Association between Ocular Chlamydia trachomatis Infection and Follicular Trachoma in Preschool-Aged Children | To examine the relationship between ocular Chlamydia trachomatis infection and follicular trachoma (TF) in children prior to and following multiple rounds of annual mass drug administration (MDA) with azithromycin.
Thirty-two communities with endemic trachoma in Kongwa District, Tanzania, were offered annual MDA as part of a district-wide trachoma control program. Presence of ocular C. trachomatis infection and TF were assessed in 3,200 randomly sampled children aged five years and younger, who were examined prior to each MDA. Infection was detected using the Amplicor CT/NG assay and TF was identified by clinical examination using the World Health Organization (WHO) simplified grading system. The association between chlamydial infection and TF in children was evaluated at baseline prior to any treatment, and 12 months after each of three annual rounds of mass treatment. Factors associated with infection were examined using generalized estimating equation models.
At baseline, the overall prevalence of chlamydial infection and TF was 22% and 31%, respectively. Among children with clinical signs of TF, the proportion of those with infection was 49% prior to treatment and declined to 30% after three MDAs. The odds of infection positivity among children with clinical signs of TF decreased by 26% (OR 0.74, 95% CI 0.65 to 0.84, p = <0.01) with each MDA, after adjusting for age. For children aged under one year, who did not receive treatment, the relationship was unchanged.
The association between ocular C. trachomatis infection and TF weakened in children with each MDA, as both infection and clinical disease prevalence declined. However, there was still a significant proportion of TF cases with infection after three rounds of MDA. New strategies are needed to assess this residual infection for optimal treatment distribution.
| Trachoma, which is caused by infection by the bacterium Chlamydia trachomatis, is the leading preventable cause of blindness worldwide. Annual mass drug administration with azithromycin is recommended for trachoma control; however, monitoring the impact of azithromycin, which targets C. trachomatis, relies on the clinical assessment of follicular trachoma. If the relationship between chlamydial infection and the presence or absence of follicular trachoma were to remain unchanged with each round of treatment, we would be able to predict the level of residual infection, and the need for additional treatment, from the prevalence of follicular trachoma. In this study, we examined the association between infection and presence or absence of follicular trachoma in children prior to and following multiple rounds of treatment. Findings suggest that with increasing rounds of treatment, the prevalence of infection declines in children both with and without signs of follicular trachoma. Newer strategies, including tests that can rapidly detect infection under field conditions, may be needed to assess residual infection in treated communities.
| Trachoma, the most common preventable cause of blindness in the world [1], [2], is caused by repeated and/or prolonged episodes of ocular infection by the bacterium Chlamydia trachomatis. The disease disproportionately affects individuals living in rural and resource-poor settings, and children are the primary carriers of ocular C. trachomatis infection and active disease [3]. We and others have previously shown that clinical disease persists longer than infection in children such that in any prevalence survey, a significant proportion of children will have active trachoma without demonstrable signs of C. trachomatis infection [4], [5], [6].
Trachoma control programs rely on a package of interventions developed by the World Health Organization (WHO) that comprises Surgery for trichiasis, Antibiotics for reducing infection, Facial cleanliness, and Environmental improvement, which is referred to as the SAFE strategy [7], [8]. Annual mass drug administration (MDA) with single-dose azithromycin is recommended for trachoma control in communities in which the prevalence of follicular trachoma (TF) is 10% or greater in children aged 1–9 years, with the aim of reducing TF prevalence to under 5% [8]. At least three years of MDA are recommended before reassessing the need for further MDA; however, monitoring the impact of antibiotic intervention, which targets C. trachomatis, relies on the clinical assessment of TF. If the relationship between chlamydial infection and TF remained constant with each MDA, one should be able to predict the level of residual infection in treated communities, and thus the need for further antibiotics, from the prevalence of TF.
As part of a randomized community trial, we evaluated ocular chlamydial infection and TF in communities located in the Kongwa District of Tanzania, where the prevalence of active trachoma ranged from 20% to 50% prior to any treatment [9], [10]. These communities have since undergone several years of annual MDA with azithromycin. The aim of the study described here was to examine the relationship between ocular C. trachomatis infection and TF in these communities prior to and following multiple rounds of MDA.
A total of 32 communities in Kongwa District, Tanzania, were selected for the community trial, which included annual mass treatment [9], [10]. A census of all households in the communities enrolled in the study was conducted to collect basic demographic information on each family member. Based on the census list for each community, a sample of 100 index children aged five years and younger was randomly selected for each survey, for a total of approximately 3,200 children examined prior to each round of mass treatment [9], [10].
Individual written informed consent from each child's parent or guardian was obtained prior to ocular examination. All study protocols and procedures were approved by the institutional review boards at the Johns Hopkins University and the National Institute for Medical Research in Tanzania. The community trial is registered on ClinicalTrials.gov under NCT00792922.
Ocular examinations and assessments were conducted prior to any treatment, and 12 months after each of three annual rounds of mass treatment. Standardized graders examined and graded both eyelids of each index child using the WHO simplified grading system [11], which assesses the presence or absence of TF, intense trachoma (TI), trachomatous scarring (TS), trachomatous trichiasis (TT), and corneal opacity (CO). For this study, the relevant signs of trachoma were TF and TI. For quality control purposes, a subset of the index children in each community was randomly selected to have ocular photographs taken using a Nikon D-series camera, and these photographs were graded by a single grader. No drift in grading over time was detected, and kappa coefficients for intergrader agreement were above 0.7 for assessment of TF at all times [12].
Ocular swabs were collected from the left upper eyelid of each index child. A Dacron swab (Fisher HealthCare, Houston, TX) was rotated and swiped across the upper conjunctiva three times and placed dry in a vial. Vials were placed in a cooler in the field, transferred to a refrigerator at the end of the day, and temporarily stored there until shipped within 30 days of collection to the International Chlamydia Laboratory at the Johns Hopkins University. Field protocols to avoid contamination of specimens were strictly followed, including changing gloves between examinations and collecting “air” controls to monitor field contamination. No evidence of field contamination was identified during the entire study [12].
All ocular specimens were processed for detection of C. trachomatis in the laboratory using the AMPLICOR CT/NG test (Roche Molecular Diagnostics, Indianapolis, IN) according to manufacturer's specifications. Each ocular swab was eluted by vortexing in lysis buffer in polypropylene tubes, after which diluent was added. Using a known positive sample, two positive and two negative processing controls were run with each batch of specimens. After the hybridization reaction, the optical density (OD) for each specimen was measured. Samples with ODs of 0.8 or greater were recorded as positive for C. trachomatis and evidence for infection; samples with ODs less than 0.2 were recorded as negative, while samples with ODs between 0.2 and less than 0.8 were considered equivocal. Samples with equivocal results were retested in duplicate; samples that retested equivocal or repeated as negative on two occasions were considered negative. Less than 0.1% of specimens were equivocal. Further details on quality assurance are discussed elsewhere [10], [13].
After ocular examinations and assessments were completed, treatment was offered to all community residents, and comprised a single dose of oral azithromycin at 20 mg/kg for adults and children over six months of age. A 1% topical tetracycline eye ointment was provided for children under six months of age with instructions to apply twice daily for four to six weeks. Antibiotic coverage exceeded 80% for children aged under 10 years in all communities at each MDA [12]. Treatment coverage at baseline, one year, and two years in children aged 0–5 years was 94% (SD 5.4%), 90% (SD 5.5%), and 90% (SD 4.9%), respectively. Compliance at six weeks with topical tetracycline among infants was not measured. Data are reported as number of rounds of community treatment offered, and are not reflective of the specific number of treatments given to each child.
All statistical analyses were conducted using SAS 9.2 software (SAS Institute Inc., Cary, NC). Contingency table analysis was used to examine the relationship at the individual level between infection and TF (with or without TI) and TI alone (with no sign of TF), stratified by child's age, gender, and the community disease prevalence prior to the most recent round of treatment. The association between ocular chlamydial infection and TF and TI alone (with no sign of TF) was evaluated at baseline prior to any treatment, and 12 months after each of three rounds of MDA. There was no statistically significant difference by treatment arm in the original trial [12], so all children from these communities were combined for these analyses. Among children with clinical signs of TF, factors associated with the presence of infection were examined using single and multiple logistic regression modeling, and a generalized estimating equation was employed to account for clustering of infection within communities. The adjusted model included age as well as number of rounds of MDA in the community.
At baseline, there was an average of 1,457 residents in each of the 32 communities included in the study (Table 1). Overall, 76% of households reported having a primary water source during the dry season that was over 30 minutes away, and 65% of households were observed to have a latrine. Heads of household reported having completed an average of 3.3 years of formal education.
The baseline mean prevalence of ocular chlamydial infection was 22%, and the baseline mean prevalence of TF was 31% (Figure 1) [12]. After each successive round of MDA, the overall prevalence of chlamydial infection declined, from 22% at baseline to 5% at 36 months, for a reduction of 79%. The prevalence of TF declined from 31% to 8%, for a reduction of 75%. At baseline, the prevalence of TI alone (without TF) was 1.4%, and this remained low over time; at 36 months, the prevalence of TI alone was 0.5%.
Among children with no signs of active trachoma, 9.0% had infection at baseline prior to administration of any treatment (Table 2). After three rounds of MDA, the proportion of children with infection among those with no signs of active trachoma declined to 2.5%, for a reduction in infection of 72% (p = <0.001). Among children with TF, 49% had infection at baseline and 30% had infection after three rounds of MDA, for a reduction in infection of 40% (p = <0.001). Among children with TI alone (without TF), 49% had infection at baseline and 35% had infection after three rounds of MDA, for a reduction in infection of 28% (p = 0.186).
At baseline, infection increased with increasing age and with community disease prevalence of over 20%, in both children with TF and children with no signs of trachoma (Table 3). There were no significant differences in infection by gender.
In unadjusted analyses, the odds of infection in children with TF decreased with each MDA (OR 0.74, 95% CI 0.65 to 0.85, p = <0.001) (Table 4). In the age-adjusted model, the odds of infection in children who had TF significantly decreased with each MDA round (OR 0.74, 95% CI 0.65 to 0.84, p = <0.001) (Table 4). For example, if we take a child aged two years, the model suggests that at baseline, the conditional probability of infection given TF is about 0.40; after two rounds of MDA, the probability of infection given TF is 0.26. Based on our model, after four rounds of MDA, the conditional probability of infection given TF is 0.16. An additional analysis was conducted to assess whether the relationship between clinical signs of TF, C. trachomatis infection, and number of MDAs changed depending on age. Stratified analysis was performed for children aged 0–1 year and children aged 2–5 years (Table 5). Among children aged 0–1 year with TF, the odds of infection were unchanged with each MDA (OR 0.93, 95% CI 0.77 to 1.12, p = 0.425). Among children aged 2–5 years with TF, the odds of infection significantly decreased with each MDA (OR 0.71, 95% CI 0.61 to 0.83, p = <0.001) (Table 5).
This study indicated that the association between ocular C. trachomatis infection and trachoma in children decreased with each successive round of mass treatment, as the overall prevalence of chlamydial infection and TF declined in the communities treated. In our statistical model, the odds of infection in children with TF were reduced by 26% with each round of MDA.
We and others have previously evaluated the cross-sectional association between ocular chlamydial infection and active trachoma at both the community level [6], [14], [15] and individual level [5], [16], [17], [18]. Most have found low levels of infection in individuals without clinical disease and higher levels of infection in those with clinical disease, which we also observed in this study. Higher levels of infection have been generally observed in individuals with TI alone (without TF) than those with TF, and we observed this same trend, though we had few TI cases in our sample.
Keenan et al. [14] observed a reduced association between active trachoma and chlamydial infection following two to three years of biannual treatment in hyperendemic areas of Ethiopia, though the association was studied at the village level and not at the individual level, so it is not possible to evaluate whether the decline was among children who had TF, as we did here. We also observed a decline in the proportion of children with infection who did not have signs of trachoma, which may reflect a decline in pre-clinical disease as trachoma declines with each successive round of MDA.
While TF and infection declined following each round of MDA, we also observed a growing discrepancy between infection and TF, reflecting an increasing proportion of children with TF and no C. trachomatis infection. There are several factors that potentially explain the presence of TF in the absence of C. trachomatis infection. There may have been cases of ocular inflammation that were recorded as “TF” when in fact the cases were caused by infectious pathogens other than C. trachomatis or by mechanical irritants, which has been previously observed [19]; however, the presence of other pathogens was very low in that study and could not entirely explain the cases of TF without chlamydial infection that were observed. Additionally, individuals who live in endemic settings and are continually exposed to C. trachomatis may exhibit persistent clinical signs of active trachoma regardless of current infection status [5]. The growing discrepancy may also relate to a likely distribution of resolution periods for follicles following infection, from very fast to very slow. With stable transmission prior to introduction of antibiotics, a prevalence survey would reflect this mix of resolution periods, with a proportion of children with TF and infection as well as a proportion of children with TF and no infection. However, if infection is drastically lowered under antibiotic pressure, then there are fewer opportunities to acquire infection; the result would be a decline in the overall TF prevalence, as was observed here. In addition, a prevalence survey would disproportionately reflect TF cases with longer resolution periods and no infection.
We do not believe that the decline in infection among cases of TF reflects a growing inability to detect organism. While insensitive laboratory tests, such as culture or Giemsa staining, might be sensitive to declining infectious loads following treatment, we used a highly sensitive nucleic acid amplification test. We did not use pooling strategies to detect infection, which can reduce sensitivity due to dilution as infectious load drops. Furthermore, all specimens were processed similarly across the trial with the same number of freeze/thaw cycles, which may also affect test sensitivity.
Finally, it is worth noting that as the prevalence of TF declines in a community, both the prevalence of infection in cases of TF and in those without disease declines, though an increasing proportion of infections overall would be seen in those without disease due to the reduced positive predictive value of TF for infection.
The odds of being infected with C. trachomatis increased with age and in those with TF. It has also been observed in previous research that infection and active disease prevalence increases by age among children under age five, and then declines by age once they start school [20]. We studied children aged five years and younger, as that is the age group of children at highest risk of trachoma, as well as being easiest to study in the community. In this population, the Pearson correlation of TF between children aged under five years and those aged under ten years was very high (0.96). Moreover, the odds of infection among children aged 0–<1 year with clinical signs of trachoma remained essentially unchanged with each community MDA. Of note, children aged under one year had not received any treatment at the time of the survey, as they were born after the previous round of MDA was provided, so the lack of change is not surprising. The finding gives further support to the idea that the rounds of MDA were associated with the decline in the association of TF and infection, and not some secular trend that would also affect children aged under one year.
We also note that there is a small risk of infection when transmission is ongoing, and in our study, 2.5% of children without TF had infection after three rounds of MDA. Even though the prevalence of infection in these children is low, in communities with very low TF prevalence, these children can become the primary source of infection. As TF prevalence declines, even fewer infections will be associated with disease, as has been observed [21].
Our model was based on communities with an average trachoma prevalence of 30% at baseline, and with very high treatment coverage rates in children. It is not clear whether findings would differ in communities with a baseline prevalence of trachoma of under 20% or over 50%, or with low treatment coverage at the community level. The association between ocular chlamydial infection and active trachoma may also be impacted by other individual or community factors favoring transmission of infection that were not measured in this study.
These study findings suggest that with increasing rounds of MDA at high coverage, the prevalence of infection decreases in children with TF. In this setting, we can model the change in the relationship between infection and disease, and it seems to hold even when TF prevalence drops below 10%, with the odds of infection in children with TF reduced by about 26% with each round of MDA. Note that some chlamydial infection still exists when TF prevalence is below 10%, indicating the potential ongoing need for MDA. Therefore, it may be worthwhile to investigate newer strategies, including tests that can rapidly detect infection under field conditions [22], [23] or possibly tests for antibodies in young children, to determine whether infection transmission is still occurring, or whether transmission has been interrupted and treatment efforts can be stopped.
|
10.1371/journal.pcbi.1000937 | Semantic Similarity for Automatic Classification of Chemical Compounds | With the increasing amount of data made available in the chemical field, there is a strong need for systems capable of comparing and classifying chemical compounds in an efficient and effective way. The best approaches existing today are based on the structure-activity relationship premise, which states that biological activity of a molecule is strongly related to its structural or physicochemical properties. This work presents a novel approach to the automatic classification of chemical compounds by integrating semantic similarity with existing structural comparison methods. Our approach was assessed based on the Matthews Correlation Coefficient for the prediction, and achieved values of 0.810 when used as a prediction of blood-brain barrier permeability, 0.694 for P-glycoprotein substrate, and 0.673 for estrogen receptor binding activity. These results expose a significant improvement over the currently existing methods, whose best performances were 0.628, 0.591, and 0.647 respectively. It was demonstrated that the integration of semantic similarity is a feasible and effective way to improve existing chemical compound classification systems. Among other possible uses, this tool helps the study of the evolution of metabolic pathways, the study of the correlation of metabolic networks with properties of those networks, or the improvement of ontologies that represent chemical information.
| Among the existing systems capable of computationally comparing chemical compounds, the majority use only structural and physicochemical properties. However, with the emergence of ChEBI and other chemical compound databases, it has become feasible to create a system that can use the relevance of compounds in a biological context as well. This setting enables the distinction of molecules with different roles in nature but similar structures, or similar roles and different structures. ChEBI is organized as an ontology that classifies chemical compounds, which we use to derive a semantic similarity measure that reflects the biological relevance of molecules. In an effort to use as much information as possible, we introduce Chym, a system that integrates structural and semantic information in a single hybrid metric, and we show the accuracy of the system in three distinct classification problems, which consist in deciding whether a compound crosses the blood brain barrier, is a P-glycoprotein substrate or an estrogen receptor ligand. Chym outperforms the previous attempts to solve these three problems, with a maximum accuracy of 90.0%.
| The recent publication of large-scale chemical information, made available by PubChem, ChEMBL and ChEBI, for instance, increased the focus of the scientific community on the problem of chemical comparison. With the amount of chemical data being published and produced today, it has become increasingly necessary to devise automatic systems capable of handling this information. The creation of an effective and accurate system that can compare and classify chemical compounds is useful in a number of different applications. For instance, it can help the understanding of the evolution of metabolic pathways, [1]; it can improve the information retrieval of disease, phenotype, and other models that contain references to chemical compounds; it enhances the study and development of pharmacophores [2], [3]; and it can also aid in toxicology, e.g. to estimate whether a given compound is or has the potential to be harmful to animals or humans without attempting a potentially harmful in vivo experiment [4].
The best approaches existing today are based on the structure-activity relationship premise (SAR), which states that biological activity of a molecule is strongly related to its structural or physicochemical properties. While the existing methods prove that this assumption generally holds, it is not always true. For instance, while L-amino acids are used to synthesize proteins, their stereo-isomers, D-amino acids, are much less frequent in nature and their role is totally different [5]. From a biological point of view, they are distinct; however, to capture their structural differences, one needs to use three-dimensional methods (like optical methods [6]), and even with that consideration the structural similarity will be high, because both molecules have the same atoms and bonds. A possible solution involves simulating the docking between molecules and a protein pocket to determine whether they should interact in vivo [7], but this method needs the three-dimensional structure of the protein, and is only valid when the property of interest is caused by a protein binding mechanism (an example of a binary classification where no protein is involved is, although a simple one, the determination of liposolubility of chemical compounds). On the other hand, both clavulanic acid and 3-carboxyphenyl phenylacetamidomethylphosphonate are -lactamase inhibitors, despite their different structures (see Figure 1). To address this problem, we propose the use of the semantics of a chemical compound in the context of biological relevance, which we used to improve the existing methods, through the development of a novel hybrid metric that takes into account both structural and semantic information. We dubbed the novel approach Chym, for Chemical Hybrid Metric. We extract semantic information from ChEBI, the Chemical Entities of Biological Interest ontology, an ontology containing more than 23,000 terms, which can be used at the base of semantic similarity [8]. Our proposal states that considering semantic similarity improves the performance of classification algorithms.
Most automatic classification methods implemented currently use either (i) the chemical structure as the foundation of the comparison [9], [10], or (ii) physicochemical properties like the molecular weight, the octanol-water partitioning coefficient etc. [11]–[14].
One of the main advantages of approach (i) is its ability to compare two or more molecules on demand, i.e., one can theoretically draw an arbitrary molecule and compare it to a whole database of structures without any prior knowledge about its function or properties.
There have been attempts to use graph comparison algorithms applied to the chemical structure of two molecules. One way of doing this is to restrict the similarity problem to the search for the maximum common sub-graph [15]. The general topology of the molecule can be used as the base of chemical similarity measures as well, where, for instance, a molecule can be represented as the matrix of the number of bonds between any two atoms and compared based on those matrices [16].
More often, though, structural similarity is calculated with the aid of fingerprints. A fingerprint, in this context, is a bitstring, a sequence of 0's and 1's, where each bit represents the presence or absence of a given feature or substructure. There are several ways to construct the fingerprint. For instance, for Daylight fingerprints, all the distinct linear fragments, up to a certain size, are identified from the graph and then converted into numbers (usually, a hash function is applied to the fragment followed by a modulo function, effectively obtaining a number in the required range). The bits in the fingerprint are then set to 1 [17], [18]. Other methods assign a particular substructure to each one of the bits of the fingerprint. Two molecules can then be quickly compared based on the number of common bits in the fingerprints, for example, through the Jaccard-Tanimoto coefficient [19].
For approach (ii), one has to compute the describing properties (if possible), to gather them from literature or to conduct experiments to obtain them.
For example, in [11], the authors used Artificial Neural Network (ANN) and Support Vector Machine (SVM) to distinguish compounds capable of crossing the blood-brain barrier (BBB) from those that do not cross it. Each compound is described as a 9-dimensional vector, where each element is a physicochemical property of the molecule (molecular weight, volume, total surface area etc.). An ANN is composed of a number of artificial neurons (a conceptual object that receives several input values and combines them non-linearly to produce a single output) arranged in layers, where the first layer gets as input the descriptors of the molecule and the last layer outputs the classification; the SVM method consists of finding the hyper-surface that best separates the active compounds' vectors from the inactive compounds' [20].
In [2], the authors used a three-dimensional representation of molecules and applied an approach named “four-point pharmacophore”. This approach builds millions of descriptors, each being a different spatial arrangement of 4 features with the respective distances between them, and then determines whether the compound contains each of the descriptors, effectively constructing a bitstring which can be used like fingerprints, as previously described. In their work, the four-point pharmacophore model was used to predict whether compounds are substrates of the P-glycoprotein (P-gp). A SVM approach was also attempted on this set [21].
The work of [13] applies the concept of decision forests to predict whether a chemical compound binds to an estrogen receptor. A decision tree consists of several if-then statements, operating over the descriptors, which ultimately come together to create a tree with several branches. The last limbs of the tree classify the compound as active or inactive. A decision forest is then an ensemble of several decision trees, where each tree is constructed from the set of descriptors still not used in previous trees, so as to minimize the fraction of misclassifications, and the final output is a combination of the outputs of the trees [13].
Random forests also use decision trees as its basis, as shown by [12]. In their work, they used random forests to classify compounds as active or inactive in several sets, including the BBB, P-gp and estrogen sets above. Unlike the decision tree approach, however, the descriptors used in each tree are randomly drawn from the set of all descriptors, rather than drawn from the set of unused descriptors.
These previous works (as well as the present study) validate their approaches by using the comparison algorithms as classification systems and consistently report performance as the fraction of correctly classified compounds: . Table 1 presents the accuracy values obtained from those systems. To evaluate the effectiveness of our approach, we took the data of these previous studies and compared the outcome of our measure to those results.
The semantic information of an object, i.e., its meaning in a predetermined context, is not easily handled by computers, mainly because meaning is mostly described in terms of natural language. For this reason, comparing the semantics of two objects (in this case, two chemical compounds), is not a straightforward task, and is only possible if the semantics of both objects are described under a common schema [22]. In this work, we used the ChEBI ontology (see below) to semantically describe chemical compounds, and under that common schema, we were able to derive a semantic similarity metric.
An ontology is a representation of terms and the relationship between them, and is usually visualized as a directed graph where nodes are the terms and the directed edges are the relationships [23]. A common type of relationship in ontologies is the “is a” relationship. It expresses the fact that one term's meaning subsumes the other's meaning, or, in other words, one term (the child) is a subclass of another term (the parent). Thus, some ontologies can be interpreted as directed acyclic graphs (DAG), where a term can have several parents and children; in such a graph, the deeper a term is, the more specific is its meaning. In the context of ontologies, a semantic measurement between two terms measures their proximity in the ontology. One of the simplest ways to compare two terms is to count the minimum number of relations that must be crossed to get from one compound to the other [24]. Another approach, used in DAGs, is to find the closest common ancestor of both terms; the distance between them is then the maximum number of relations from one of the two terms being compared to the common ancestor. It is worth noting that a measure can be a distance (as the terms get closer, the distance decreases) or a similarity (as the terms get closer, the similarity increases). Here we will consider only similarity measures.
In this work, we used both the ontology as a graph and a concept known as information content. The information content is an abstract concept that reflects the specificity of a particular object [25]. From information theory, the information content of an object can be evaluated as the negative logarithm of the probability of finding that object [24]. When calculating information content, it should be noted that a function is only meaningful if each term's occurrence contains all its children's occurrences too. In an ontology like ChEBI, this means that for more abstract terms the probability embraces many terms, decreasing its information content, which, in turn, reflects its low specificity. The probability function we will use is based on the number of pathways each compound participates in. The reason behind this choice is that counting the number of pathways gives a measure of specificity (compounds or chemical classes that are more specific will be found in less pathways), but it is not biased against the problems that Chym tries to solve.
To validate the effectiveness of Chym as a classification tool, we tested it on the sets presented in Table 1 and compared our results with the ones in that table. Since the results of chemical classification algorithms are usually reported in terms of accuracy (the fraction of correctly classified compounds), we report accuracy of Chym. However, for binary classification problems, Matthews Correlation Coefficient (MCC) is a better performance indicator [26]. Therefore, we use this coefficient as the main measure of Chym's performance.
In the three sets retrieved from the previous works presented in the introduction, the compounds were listed by name only, with no information on structure. The first step in the assessment of Chym was, therefore, to translate that list of names into ChEBI identifiers. The task of getting the identifiers was accomplished by string matching techniques, since there was no structural information to make the search. We split the names into bags of words, where a word is a sequence consisting of only letters or only numbers, to determine whether two names refer to the same chemical entity. We used not only the preferred names of the compounds but also the synonyms stored in the ChEBI database. Only compounds present in the ontology and with a described molecular structure in the ChEBI database were considered. Because ChEBI is continually growing, we estimate that older compounds in the ontology are usually more correctly annotated and tend to have lower identifiers. So, in case of more than one possibility, we chose the lowest ChEBI id.
Since the ontology does not contain all the possible molecules, we were not able to get a full mapping between names and ChEBI compounds, which means that our sets were shorter versions of the original ones. We refer to our smaller sets as purged versions and denote them as BBB, P-gp and estrogen. Table 2 shows the fraction of compounds in each of the three sets that are present in the ontology.
The results of this table show a significant reduction in the size of all three sets after converting the names into ChEBI identifiers. Facing these values, we chose to directly compare our results only to the ones obtained with the blood-brain barrier, because (i) it is the set with higher percentage of ChEBI coverage, (ii) after purging, it remains the biggest set, and as such is more fit to be broken into testing and training sets without losing too much information, and (iii) it is the set with a more balanced distribution of active vs. inactive compounds. We will also apply Chym to the two other sets, but the analysis will not be as deep.
The BBB set is first described in [11], where the authors use an artificial neural network (ANN) and a support vector machine (SVM) to classify several chemical compounds as either able to cross the blood-brain barrier (active) or unable to do so (inactive). The paper showed that SVMs are more effective in this particular classification problem than ANNs (see Table 1). The set was further used by [12], where a random forest was grown to classify the compounds. The authors of this work showed the effectiveness of this system in several chemical compound sets, but the results obtained for the BBB set in particular were not better than the ones obtained with SVM. For this reason, we compared our results to the 81.5% accuracy reported by [11] (cf. Table 1). It is worth mentioning again that we report accuracy only for comparison purposes, but the real performance indicator should be MCC, which is also reported in the tables with the results.
In order to make an unbiased comparison between Chym and SVMs, we addressed the validation process in three steps, which were devised so that only one specification of the process changed in each step:
It must be mentioned here that Chym is actually a collection of 24 metrics, each having a real parameter, , that balances the metric between structural and semantic information. We used values from 0 to 1 in steps of 0.01, making a total of metrics (to understand the reason for these numbers, refer to the section Methods). The metric that yields the highest Matthews Correlation Coefficient is reported on the tables, alongside the MCC and accuracy values achieved with that same metric.
For the SVM approach, we retrieved the compounds' properties from the article as 9-dimensional vectors and used the SVMlight [27] software with a radial basis function kernel, as described in [11].
Moreover, to decrease the potential bias in our analysis, we implemented two different validation methods. The first one is a leave-multiple-out process, described in [11] and here dubbed “LMO25”. LMO25 follows this algorithm:
The second validation approach is -fold cross-validation, which is more widely used and well documented [28]. It starts by partitioning the compounds in the original set into approximately equal-sized, stratified sets, meaning that the proportion of active inactive compounds is maintained in the partitions. Then each partition is used as testing group once, with the other partitions being used to train the model. Accuracy and MCC values are recorded for each partition and averaged in the end of the iterations. To remove the noise coming from the initial partition, we performed this method times and averaged the accuracy and MCC obtained. This validation approach was also applied to the P-gp and estrogen sets. We used and throughout the whole analysis [28].
The last step in the assessment of Chym was to predict some new active compounds in each of the three sets. We calculated an activity coefficient for all compounds in the ChEBI ontology annotated with a structure, based on the active compounds in the respective purged sets, and the best metric for each problem, and retrieved the ones whose coefficient was higher. For a discussion about the methods used to calculate this value, refer to section Methods.
Table 3 shows the main results of the validation process, including the attempt to replicate the results of [11]. Given that we have 24 different metrics, each one tuned with a real parameter , Chym had to select one of the possibilities. The best combination for this problem with the 10-fold cross-validation approach was FP3 fingerprint format with semantic similarity calculated for all the ontology with a simGIC method, with 29% of weight to structure and 71% to semantics (). The same metric was pre-chosen for the LMO25 approach, even though Chym could have found that the best metric with this approach was not this one. The parameters of Chym's best metric (in this case FP3, simGIC, the whole ontology and the value of ) are explained in more detail in the Methodology section below. The results presented in the table show the superiority of Chym when compared with the SVM approach. Moreover, when we compare the two sections of the table with each other, it is possible to see that the validation method does not significantly affect the results. Since the 10-fold approach is more widely used, at least when compared to the LMO25, we performed the main analysis of our results with this method.
In its second part, Table 3 shows the results of using 10-fold cross-validation instead of LMO25. Here we show that the accuracy of the SVM method used previously decreases significantly when some of the compounds in the set are removed. However, the same purged set can be used by Chym, and still achieve an accuracy 10% superior to the one originally reported, with an associated Matthews Correlation Coefficient increase of almost 0.2 units. One possible explanation for this is the effect of the retention of chemotypes. Information on the chemotypes is implicitly contained in the ontology, which may buffer the effect of removal of individual molecules. It would be interesting to make an analysis of chemotypes, for example through Murcko scaffolds [29], [30]. However, the fact that the data sets were retrieved by name and not by structure invalidates this approach. But the SVM approach's accuracy decreases 6–8% when used on the purged set, which seems to suggest that the chemotype retention is not prominent.
Table 4 shows the performance of Chym when applied to the data sets used. The “Chym parameters” column specifies the parameters of the best metrics in terms of which fingerprint format, semantic method, ontology and value are best suited for that set (cf. Table S1, Table S2, Table S3 in the Supporting Material, available online for the results obtained for all metrics). This table reinforces the prediction power of Chym, since its performance with the P-gp and estrogen sets, which are also about 60% smaller than the original ones, is still higher than (for the P-gp set) or comparable to (for the estrogen set) the value obtained with the random forest approach, the best method applied so far to those sets (cf. Table 1). Although there does not seem to be any improvement in the estrogen problem, we must underline that we have used a smaller set, and we believe the performance would increase with a more complete ChEBI ontology. On the other hand, using the values reported in the work about random forests [12], it is possible to recalculate the MCC value obtained with that method, 0.647. The value of 0.673 achieved with Chym represents a slight increase.
Table 5 and Figure 2 show the MCC of three Chym systems against values. For each set, the parameters used with the Chym system are the ones which reached maximum accuracy for some value of . It is visible that, in the three Chym systems, the accuracy starts by increasing at first, reaching a maximum, and decreasing again. This shows that using the hybrid measure is better than using only purely structural or semantic metric. When this same analysis is applied to other Chym parameters, we can observe the same behavior, which confirms the idea that, even if one system is not very accurate, the crossing of structural and semantic information increases the prediction power of Chym.
Finally, Table 6 shows the most active ChEBI compounds, as defined by the activity coefficient, retrieved for each problem. In each problem, we retrieved all the ChEBI compounds in the ontology and with a molecular structure (more than 15000) and ranked them by activity coefficient. The table reports the first three whose classification has been previously determined in a publication and shows that those compounds are, in fact, active compounds (they cross the blood-brain barrier, are substrates to P-glycoprotein or ligands to the estrogen receptor), which also contributes to the idea that the Chym method is effective. The entire ranked lists of compounds are available as Table S4, Table S5, Table S6.
The work presented in this paper shows compelling evidence that using semantic information in chemical classification algorithms improves their performance. To show that, we used three sets of compounds previously described and used as input in other classification methods. On those sets, Chym achieves higher performance for class prediction when compared to previously existing methods, with Matthews Correlation Coefficient as high as 0.810, corresponding to an accuracy of 90.0%. Parallel to this result, we also showed that the use of a hybrid metric that uses both structural and semantic information is better suited for this kind of problems than a system which uses only one of these types of information. Some issues should, however, be discussed in order to complete the analysis of this tool.
The properties that are relevant to decide whether a molecule should be classified as active or inactive depend obviously on the problem being solved. As such, the best metric for a problem is not necessarily the same for other problems. Thus, selecting the best metric is not much different than selecting the appropriate descriptors for SVM, random forest or the other approaches presented before. While it may be argued that the value of is inherent to Chym's method, it does reflect the relative amount of structural and semantic information that must be used to correctly classify compounds. Choosing the appropriate value for this and the other parameters can be seen, from the point of view of usage, as a task similar to choosing the appropriate descriptors that better reflect the important characteristics of the molecules (those that yield the best results). For instance, as specified in Table 4, while the FP3 fingerprint format is good at detecting some substructures that are important in the BBB problem, it misses the relevant structures in the other problems. Furthermore, the BBB problem is better solved with a stronger focus on the semantic information, and the value of the best metric reflects this, as evident in Figure 2.
On another note, our high performance could be due to a possible term in the ontology that classified compounds as able to cross the blood-brain barrier, as substrates to the P-glycoprotein or as estrogen receptor ligands. Admittedly, if there were such terms in the ontology, Chym would be biased and would report high accuracy values because it would be using the information it was trying to validate as a means to prove its effectiveness. As it turns out, no term in the ontology refers to the words “brain”, “barrier”, “P-glycoprotein” or “permeability” (the meaning of the P in P-glycoprotein). “Estrogen receptor” appears twice, in “estrogen receptor modulator” and “estrogen receptor antagonist”, but these two terms have only a total of 5 descendants in the ontology, and none of them is present in the set estrogen. This fact suggests that Chym can be used in many classification and similarity problems, even if they are not well represented in the ontology.
The reason for this fact is that, although the information to solve the classification problem is not explicitly stated in the ontology, the proximity of terms in the ontology (their semantic similarity) is a good indicator that they should behave similarly. For instance, both the compounds ChEBI:8069, phenobarbital, and ChEBI:49575, diazepam, cross the blood brain barrier. Moreover, they share many of their ancestors. Their semantic similarity, as measured with a simGIC method in the whole ontology, is 0.324, and their structural similarity, as measured with the FP3 format, is 0.667. With an (these are the parameters chosen for the BBB problem, cf. Table 4), this results in a similarity of 0.423, well above the mean similarity between the active compounds in the BBB set, 0.238. The compounds are both annotated as sedative drugs, and Chym was then able to determine that ChEBI:51137, mianserin, another sedative drug, also crosses the BBB (see Table 6).
Still in respect to the results presented in Table 6, a further analysis showed that ChEBI:5078, flavonol, was ranked in the list of estrogen receptor ligands (activity coefficient, higher than the threshold calculated for that problem), but [31] showed that this compound is not an estrogen receptor ligand. However, the class of compounds named flavonoids, into which flavonol is classified, is known to contain several compounds that bind to the estrogen receptor [32], [33]. Moreover, this compound shares most of galangin's ancestry, 58 common ancestors out of 61 total ancestors (galangin is also on that table). This means that the ChEBI ontology is not yet able to differentiate between these two compounds, and so it produces a false positive. As a matter of fact, the similarity between these two compounds in the metric chosen for the estrogen problem is 0.716, while the mean similarity between all the active compounds is 0.216, demonstrating that ChEBI assigns high similarity to these molecules.
As discussed in the Methodology, the ChEBI ontology contains three partially overlapping branches. One concern raised by this fact is that the molecular structure more or less reproduces the structural information used in the first part of the metric. Although the information being used is indeed the same, the ontology explores the structural properties from a totally new perspective (namely, a semantic perspective), that would be otherwise unusable in a similarity measure: purely structural comparison methods are probably unable to use the fact that both glucose and fructose are monosaccharides to compare them. So, even if there seems to be a duplication of information, the different approaches used yield similarity values that can be combined to produce a more robust score (as Chym does).
Another concern raised about the use of ChEBI ontology is the subatomic branch. This branch was never chosen by itself as the best branch of the ontology, which is not surprising, for two reasons. First, it is not much richer than the molecular structure or role branches, since only 35 ChEBI terms are unique to this branch of the ontology. Secondly, each of these 35 terms is either an ancestor to all chemical compounds used in the input set (as happens with electron, for instance, which is part of the atom, which is part of every molecular structure) or ancestor to none of the chemical compounds (photon, for instance). This means that this branch does not offer any kind of resolution.
However, like any other classification algorithm, Chym has its limitations. The most important drawback of this method is that it can only compare structures that are annotated in the ChEBI ontology. Of course that any chemist or other scientist wishing to use Chym may annotate the compound they are trying to study in ChEBI by creating a “non-official” node. There is, however, a large number of classes, which could potentially introduce a difficulty in selecting the most appropriate position for the compound; this annotation is also unfeasible for a large number of compounds. This severely impairs applications like drug discovery, or toxicology analysis.
In spite of this limitation, Chym introduces the comparison of chemical compounds through their semantics, which is an important technique that can be used in projects where comparison and or classification of known chemical compounds is needed. One instance of such project is the search for a possible correlation between strains of bacteria and their virulence. One could be interested in determining differences in metabolic networks of said strains and compare the differences with the different amount of virulence of those strains; the comparison of metabolic networks would benefit from the metrics explored here. Other applications include the comparison of models, for instance models of diseases containing references to molecules responsible for the disease or to drugs known to improve the condition of patients. On the other hand, the semantic similarity applied to ChEBI (developed and explored in this work) can also be useful in ontology managing, as happens in GO [34], where semantic similarity is used to automatically annotate other molecules in the ontology and automatically improve the ontology. This would in turn be useful in information retrieval and automatic reasoning methodologies.
In the future, it would be interesting to try other hybrid metrics, especially other structural comparison algorithms. For instance, since SVM and random forests seem to perform well, perhaps a system where the structural part of the comparison is done through one of these methods would outperform the actual version of Chym.
In order to develop and validate our hybrid similarity for chemical compounds, the Chemical hybrid metric (Chym), we built a model based both on fingerprints and on the semantic similarity measures developed for the Gene Ontology (GO) [35].
To calculate the structural similarity between two molecules, we need a representation of their structures. Because ChEBI contains a list of structures in SMILES, MDL and InChI chemical file formats, these are the formats used. For each distinct molecule, we prefer a SMILES representation of the structure. If one does not exist, we use MDL. The rationale for this choice is the wide use of SMILES over MDL. InChI was not used since every molecule with a structure in this format had at least one of the other formats as well.
For each structure, three fingerprints were calculated. These formats were computed with the OpenBabel software [36], [37], and as such we used the names and files provided by it:
Given two molecules and the corresponding fingerprints and , the similarity score between them is calculated according to the Jaccard-Tanimoto coefficient [17], [38], [39]:(1)where and are the bit in each of the fingerprints. Obviously, comparison of fingerprints is only valid if the fingerprints were obtained by the same method. This equation is valid only if the denominator is different from 0. It was verified that all fingerprints calculated had at least one bit set to 1, thus making the denominator always positive.
From equation 1, it can be seen that the structural similarity will run from 0, when no bit is 1 for both molecules (total disparity), to 1, when the 1-bits in the two molecules are the same (equal fingerprints).
Following the application of semantic measures for the GO [35], we developed a similar approach but instead of proteins, we work with chemical compounds. As has been stated above, there are a number of ways to measure semantic similarity based on an ontology. We chose to use the same ones as [35]. In the next paragraphs, consider , and as chemical compounds and as the set of ancestors of the chemical compound , including itself.
simUI is a graph-based measure, which means that it considers the compounds and all their ancestors in the graph of the ontology. It is defined as follows [40]:(2)
It is known, however, that for ontologies where term specificity is not well correlated with term depth, methods based on information content (IC) are preferable [35]. Let be the frequency of usage of the term in some corpus. The information content of a term can be given by [41]:(3)Intuitively, equation 3 means that a very frequent term conveys less information and vice-versa. Notice that the frequency of a term subsumes the frequency of the terms that derive from . This means that the frequency of the term amino acid includes the frequency of terms L-serine or carnitine, a -amino acid. Therefore, less specific terms are less informative. Chym makes use of this equation, where the terms are the nodes of the ChEBI ontology, viz chemical compounds or chemical classes.
simGIC is a combination of the graph-based simUI metric with the information content properties of compounds. The concept behind the equation is the same as the one behind simUI, but now each ancestor is weighted according to its information content, which reflects its specificity. simGIC is calculated through equation 4 [35].(4)
It is worth underlining here that the concept of information content is just a method to give weight to the compounds in the ontology. If two compounds share many ancestors, simUI will attribute a high similarity between them, but, for example, if most of those ancestors are unspecific, the similarity should be lowered accordingly; by weighting the ancestors, simGIC achieves this effect. For example, compounds ChEBI:17802, pseudouridine, and ChEBI:31747, kanosamine, share 30 or their 37 ancestors, but the most specific of those is ChEBI:23008, carbohydrate, already a very abstract term in the ontology. simGIC takes into account this fact. Considering the similarity values between all pairs of compounds that appear in the corpus at least once, the mean similarity measured with simUI is 0.431 and the mean similarity with simGIC is 0.048. Those two compounds share a simUI similarity value of 0.811, about twice the mean value, but by weighting the ancestors, simGIC assigns a similarity of 0.023, about half of the mean value.
For both metrics, the similarity value is between 0 and 1 because an intersection of two sets is always a subset of their union.
Until this point, we presented two orthogonal metrics to measure the similarity between two chemical compounds. Our intent, however, is to join them together to produce a hybrid metric that takes into account both structural and semantic information.
Since both measures explained above always fall in the closed interval , we propose the following definition for our hybrid similarity:(5)where is a real number from 0 to 1. When , the identity degenerates into pure structural similarity and with , into pure semantic similarity.
One of the possible uses of Chym is the application of this similarity metric to classify compounds. Ideally, we want to be able to get a set of chemical compounds that possess a common property as input, and then determine whether other chemical compounds also possess that property. This is also the approach used in SVM and random forests, for example, where the input serves as a training set that is used to create a classification model. In Chym, the model consists of a threshold that is used to decide whether a compound is active or inactive.
Given a training set of compounds, some sharing a common property (which we call active compounds), and some lacking that property (inactive compounds), the following algorithm is used to predict whether a compound in the validation set is active or inactive:
From the algorithm above, it can be seen that the inactive compounds are only used to adjust the value of the threshold, while the active compounds are used both in the adjustment of that value and in the determination of the activity coefficient of the validation compounds.
Chemical Entities of Biological Interest (ChEBI) is a freely available database of small molecular entities (distinct isotopes, atoms, ions, molecules etc.). These entities may be products of nature or synthetic products used to intervene in biological processes [8].
The ontology also includes classes of molecular entities and partial molecular entities, enabling ChEBI to be organized as an ontology, structuring molecular entities into classes and defining the relations between them. Several relationship types exist in ChEBI, with a number of them reciprocal in nature. The ontology is subdivided into three separate sub-ontologies:
As of the time of the computations (January 2010, release 64), the graph of this ontology contained 23,545 nodes representing chemical compounds, which represents approximately 4% of the whole ChEBI database. As stated above, some terms are not chemical compounds but parts of compounds, such as functional groups, that make the ontology structure possible. Also, for each individual chemical compound, there may be several identifiers, which come from different annotations that were later identified as the same compound.
Chym's branches are partially overlapping. For instance, the term glucose is classified as a molecular structure, as having the role of macronutrient and as having part electron, which means that it is present in three branches. Including glucose, 21676 nodes (92%) are part of the three branches.
Besides the ontology, the ChEBI database is enriched with an extensive list of synonyms and manually curated cross-references to other non-proprietary databases, as well as a list of chemical structures.
Kyoto Encyclopedia of Genes and Genomes (kegg) is a collection of databases categorized into systems information, genomic and chemical information. The different kegg databases are highly integrated in an effort to constitute a computer representation of the biological system [42].
One of the main components of kegg is the kegg pathway database, which contains a collection of graphical representations of known pathways. Each metabolic entry integrates information from other databases in kegg, such as the intervening enzymes (kegg enzyme), chemical reactions (kegg reaction) and chemical compounds present in the pathway (kegg compound).
kegg compound is a chemical structure database for metabolic compounds and other chemical substances that are relevant to biological systems. We use entries in the kegg compound database as annotations of the compounds present in the metabolic pathways (kegg pathway entries). The existence of a mapping between ChEBI and kegg compound allows us to integrate information from both databases.
The methods used to structurally compare compounds are implemented by the software, OpenBabel [36], [37]. We used version 2.2.3, which was downloaded and installed on December 2009.
The semantic similarity was not as straightforward. As in [43], we had to reorganize the ChEBI ontology so that it could fulfill our purposes. All cyclic relationships (“is tautomer of” etc.) were removed, and the other relationships were merged into a single “is a”-like relationship. Also, ChEBI identifiers pointing to the same chemical compounds were merged into a single node. Thus, we produced three independent DAGs, one for each branch of the main ontology, and a forth DAG resulting from merging the other three. With this modification, we can directly calculate simUI similarities with equation 2.
To calculate the IC-based metric (simGIC), we had to find a corpus where the compounds are referenced. We chose kegg pathway because it is not connected to any of the problems solved by Chym. This has the advantage of avoiding a potential bias that could boost Chym's results. To map a ChEBI identifier into a kegg identifier, we used the ChEBI cross-references. Sometimes, however, these references were ambiguous (one ChEBI id pointing to two or more kegg compound ids). For this reason, whenever a ChEBI id had more than one kegg compound reference, we used them all to determine the number of pathways in which participates. With this corpus, the value of from equation 3 is the fraction of pathways where the compound or any of its descendants appear.
Since there are 3 fingerprint formats, and semantic similarity can be calculated based on 4 different DAGs and with 2 different methods, the approach we are presenting here is able to use different similarity metrics, each with a real parameter .
|
10.1371/journal.ppat.1006779 | A parapoxviral virion protein targets the retinoblastoma protein to inhibit NF-κB signaling | Poxviruses have evolved multiple strategies to subvert signaling by Nuclear Factor κB (NF-κB), a crucial regulator of host innate immune responses. Here, we describe an orf virus (ORFV) virion-associated protein, ORFV119, which inhibits NF-κB signaling very early in infection (≤ 30 min post infection). ORFV119 NF-κB inhibitory activity was found unimpaired upon translation inhibition, suggesting that virion ORFV119 alone is responsible for early interference in signaling. A C-terminal LxCxE motif in ORFV119 enabled the protein to interact with the retinoblastoma protein (pRb) a multifunctional protein best known for its tumor suppressor activity. Notably, experiments using a recombinant virus containing an ORFV119 mutation which abrogates its interaction with pRb together with experiments performed in cells lacking or with reduced pRb levels indicate that ORFV119 mediated inhibition of NF-κB signaling is largely pRb dependent. ORFV119 was shown to inhibit IKK complex activation early in infection. Consistent with IKK inhibition, ORFV119 also interacted with TNF receptor associated factor 2 (TRAF2), an adaptor protein recruited to signaling complexes upstream of IKK in infected cells. ORFV119-TRAF2 interaction was enhanced in the presence of pRb, suggesting that ORFV119-pRb complex is required for efficient interaction with TRAF2. Additionally, transient expression of ORFV119 in uninfected cells was sufficient to inhibit TNFα-induced IKK activation and NF-κB signaling, indicating that no other viral proteins are required for the effect. Infection of sheep with ORFV lacking the ORFV119 gene led to attenuated disease phenotype, indicating that ORFV119 contributes to virulence in the natural host. ORFV119 represents the first poxviral protein to interfere with NF-κB signaling through interaction with pRb.
| Poxviruses have evolved multiple strategies to subvert signaling by NF-κB, a crucial regulator of host innate immune responses. Viruses often encode multiple inhibitory proteins, which largely target cytoplasmic activation events of NF-κB signaling. The retinoblastoma protein (pRb), a multifunctional protein best known for its tumor suppressor activity, has been suggested to affect NF-κB signaling during virus infection however, viral effectors and mechanisms of actions are unknown. Here, we identified a virion-associated orf virus NF-κB inhibitory protein, ORFV119, which interacts with pRb. ORFV119 was shown to inhibit IKK complex activation in a pRb-dependent manner early in infection. Results show that ORFV119 interacted with both pRb and TRAF2 and that a ORFV119-pRb complex likely is required for efficient interaction with TRAF2 and inhibition of NF-κB signaling. ORFV119 represents the first poxviral protein to interfere with NF-κB signaling through interaction with pRb.
| Orf virus (ORFV), the type member of the genus Parapoxvirus of the Poxviridae, is the causative agent of orf or contagious ecthyma, a ubiquitous disease of sheep and goats characterized by proliferative lesions affecting muco-cutaneous tissues of the mouth and muzzle [1,2]. Orf is a zoonotic disease affecting humans in close contact with infected animals [3–5].
Like other parapoxviruses (PPV), ORFV is highly epitheliotropic, and keratinocytes and their ontogenetically related counterparts in the oral mucosa are the most important if not the only cell type to support ORFV replication in vivo [6]. In addition to producing the essential protective stratum corneum of the epidermis, keratinocytes function as immune sentinels and instigators of inflammatory responses in the skin, representing a specialized branch of the innate immune system. Keratinocytes are well equipped for pathogen sensing as they express a broad spectrum of pattern recognition receptors (PRRs), including surface and endosomal toll-like receptors (TLRs), NOD-like receptors (NLRs), and retinoic acid-inducible gene (RIG-I)-like receptors, and rapidly respond to cell injury and infection by releasing critical pro-inflammatory chemokines and cytokines such as tumor necrosis factor α (TNFα) and interleukin 1 (IL-1) [7,8]. Engagement of these receptors initiates downstream pro-inflammatory cascades, including the NF-κB signaling pathway, which mediates innate immune responses and contributes to skin homeostasis by regulating keratinocyte proliferation and differentiation [9].
The NF-κB family of transcription factors consists of five members, NF-κB-p65 (RelA), RelB, c-Rel, NF-κB-p50/p105, and NF-κB-p52/p100, which contain an N-terminal Rel homology domain (RHD) responsible for homo- and heterodimerization and for sequence specific DNA binding [10]. In unstimulated cells, NF-κB dimers are sequestered in the cytoplasm through binding to the inhibitor kappa-B alpha (IκBα). Following cell stimulation, IKK complex-mediated phosphorylation of IκBα results in proteasomal degradation of IκB and nuclear translocation of p65/p50 dimers, which bind κB-responsive DNA elements, interact with transcription co-regulators, and activate or repress gene expression [11,12]. The critical IKK complex consists of two kinases, IKKα and IKKβ, and the regulatory subunit IKKγ/NF-κB essential modulator (NEMO) [13,14]. Various stimuli, including those initiated by proinflammatory cytokines TNFα and IL-1, lead to IKK activation. Engagement of the TNF receptor 1 (TNF-R1) results in sequential recruitment of TRADD (TNF-R1-associated death domain), TRAF2 (TNF receptor-associated factor 2) and RIP1 (Receptor-interacting protein 1) [15]. Multiple ubiquitination events on RIP1 and NEMO bring the TAK1 (TGF-β activated kinase 1) complex close to the IKK complex. TAK1-mediated IKKβ phosphorylation and IKKβ auto-phosphorylation activate IKKβ, which then phosphorylates IκBα [16]. Engagement of the IL-1 receptor, on the other hand, results in recruitment of IRAK1 (IL-1 receptor–associated kinase) and activation of TRAF6 (TNF receptor-associated factor 6), which then ubiquitinates and activates TAK1 resulting in IKK activation [17,18].
Many viruses with dissimilar life styles are known to interfere with the NF-κB pathway. In particular, poxviruses have evolved multiple strategies to counteract NF-κB function, indicating that inhibition of NF-κB-mediated transcription is important for successful infection of the host. This is not surprising as poxvirus infections are sensed by NF-κB-activating PRRs such as endosomal TLRs, RIG-I-like receptors, and the inflammasome [19]. General features of poxviral NF-κB inhibitors include, 1- individual viruses encode for multiple inhibitors, with vaccinia virus (VACV) encoding at least twelve [20]. While orthologs of some NF-κB inhibitors are found in viruses belonging to multiple poxvirus genera (e.g. VACV A52R, VACV E3L), others are restricted to a particular genus (e.g. VACV A46R and VACV B14R in Orthopoxvirus) or even selected viruses within a genus (PPV ORFV002) [21–29]; 2- in contrast to other classes of poxviral immunomodulators, poxviral NF-κB inhibitors have no or little resemblance to host proteins; 3- although inhibitors target extracellular, membrane, cytosolic, or nuclear events of NF-κB regulation, most inhibitors directly target NF-κB subunits or the proximal IKKs; 4- despite the multiplicity of inhibitors, there seems to be low or no redundancy as judged by the effect of individual gene deletions on viral pathogenesis; 5- with a few exceptions (myxoma virus MYXV150, cowpoxvirus CPXV006) no single gene-deletion rendered complete virus attenuation [30–34]; 6- most inhibitors are expressed early after virus entry into cells.
Apart from VACV E3L (ORFV020), PPV lack homologues of poxviral NF-κB inhibitors present in other poxviral genera, suggesting that PPV have evolved novel mechanisms to counteract the NF-κB signaling pathway. Recently, we have described four NF-κB inhibitors encoded by ORFV, ORFV024, ORFV002, ORFV121 and ORFV073 [35–38]. ORFV024 was shown to inhibit phosphorylation of IKK kinases, thus preventing activation of IKK complex. ORFV121, a virulence determinant, was shown to bind to- and inhibit phosphorylation and nuclear translocation of NF-κB-p65 [37]. ORFV002 binds NF-κB-p65 and reduces its acetylation by co-activator p300 thus inhibiting transactivation [36]. Decreased NF-κB-p65 acetylation is a consequence of ORFV002 interfering with NF-κB-p65 phosphorylation by mitogen- and stress-activated protein kinase 1 (MSK1) [39]. ORFV073 inhibits NF-κB signaling by preventing activation of IKK complex through interaction with NEMO, the regulatory subunit of the IKK complex [38].
The retinoblastoma tumor suppressor, pRb, is a multifunctional, predominantly nuclear protein encoded by the RB1 gene. pRb affects cell cycle, differentiation and metabolism, genome stability and apoptosis, mostly but not exclusively through transcription regulation [40]. Central to pRb function is its ability to nucleate complexes containing multiple interacting partners, thus participating in various regulatory circuits. Viruses have evolved functions to modulate those pathways by targeting pRb to their advantage. For example, adenovirus (Ad) protein E1A interaction with pRb and other factors represses select host genes to promote productive virus infection [41]. The human cytomegalovirus pp71 tegument protein and the human papillomavirus E7 oncoprotein bind to pRb and induce its degradation thus driving cells to a cell cycle stage that potentially favors efficient replication of viral genomes [42,43]. pRb has been shown to affect the regulation of the NF-κB pathway following TNFα signaling or vesicular stomatitis virus infection, although mechanisms involved are not yet defined [44,45]. Many viral and cellular pRb-interacting proteins contain the motif LxCxE (where x means any amino acid) that binds to the LxCxE cleft of the pRb pocket domain [46]. Select pRb-binding proteins that interact through LxCxE with pRb also bind p107 and p130, the other two members of the retinoblastoma family of proteins.
Here we show that ORFV119, a LxCxE motif-containing virion protein unique to PPV, interferes with NF-κB signaling in a pRb-dependent manner early in infection by inhibiting IKK complex activation.
Primary ovine fetal turbinate (OFTu) cells were obtained from Howard D. Lehmkuhl (USDA) and were maintained in minimal essential medium (MEM) supplemented with 10% fetal bovine serum (FBS) (Atlanta Biologicals, Flowery Branch, GA), 2 mM L-glutamine, gentamicin (50 μg/ml), penicillin (100 IU/ml), and streptomycin (100 μg/ml). Human osteosarcoma cells (Saos-2, provided by Timothy M. Fan, UIUC), Human embryonic kidney (HEK 293T) and cervical cancer (HeLa) cells (obtained from American Type Culture Collection), were cultured in Dulbecco's modified essential medium (DMEM) supplemented as above. Cells were incubated at 37°C with 5% CO2. ORFV strain OV-IA82 [47] was used as a parental virus to construct ORFV119 gene deletion virus OV-IA82-Δ119 and for experiments involving wild type virus infection. OV-IA82-Δ119 was used as parental virus to construct revertant virus OV-IA82-RV119Flag, and OV-IA82-RV119LxGxE-Flag, a revertant virus carrying a CxG substitution in ORFV119 LxCxE motif.
To knockdown pRb expression in OFTu cells, siRNA directed against ovine RB1 was used. A pool of three RB1 sense (S) and anti-sense (AS) siRNA with following sequences 1-RB1-(S): CTACCTATAGCAGAAGTAT, RB1-(AS): ATACTTCTGCTATAGGTAG; 2-RB1-(S): GCTTATATATTTGACACAA, RB1-(AS): TTGTGTCAAATATATAAGC; 3-RB1-(S): CTCAGATTCACCTTTATTT, RB1-(AS): AAATAAAGGTGAATCTGAG (Custom Oligos: Sigma Aldrich) were used. Pooled RB1 siRNA (15 nM of each siRNA) were transfected in OFTu cells (30,000–70,000/well) using MISSION transfection reagent (Sigma Aldrich, Cat # S1452) following the manufactures’ protocol. At 48 h post transfection, pRb knockdown was examined by SDS-PAGE, using antibody against pRb (abcam, Cat # ab85607). One MISSION siRNA Universal Negative Control (SIC001, sigma aldrich) was included in all experiments.
To obtain OV-IA82Rb- or OV-IA82-RV119LxGxE-Flag-Rb- virus stocks (OV-IA82 and OV-IA82-RV119LxGxE-Flag viruses propagated in cells with reduced pRb levels, respectively), OFTu cells with reduced pRb levels (OFTuRb- cells) were infected with OV-IA82 or OV-IA82-RV119LxGxE-Flag virus and supernatants from infected cultures were collected at 24 h p.i and used for NF-κB-p65 nuclear translocation assays.
To construct expression plasmids pORFV119Flag and pORFV119LxGxE-Flag, the ORFV119 or ORFV119LxGxE coding sequences were PCR-amplified from the OV-IA82 genome and a plasmid containing synthetic ORFV119LxGxE sequence (Genscript), respectively with primers 119-3xFlag-FW (EcoRI): 5’-TAAGGCCTCTGAATTCAATGGACTCTCGTAGGCTC GCTCTT-3’; 119-3xFlag-RV (BamHI): 5’-CAGAATACGTGGATCCTCAATCGCTGTCG CTGTCGCCGAG-3’ and cloned into p3xFlag-CMV-10 vector (pFlag) (Clontech, Mountain View, CA). Similarly, to obtain a ORFV119-green fluorescence protein (ORFV119GFP) expression vector, ORFV119 sequence was PCR amplified from OV-IA82 genome with primers 119-GFP-FW (EcoRI): 5’- TAAGGCCTCTGAATTCATGGACTCTCGTAGGCTCGCTCTT-3’ and 119-GFP-RV (BamHI): 5’- CAGAATACGTGGATCCAGATCGCTGTCGCTGTCGCCGA GCG-3’, and cloned into the vector pEGFP-N1 (Clontech, Mountain View, CA). DNA sequencing of plasmids confirmed the identity and integrity of the constructs.
To generate gene deletion mutant virus OV-IA82-Δ119, a recombination cassette (pΔ119-RFP) containing Vaccinia virus 7.5 promoter (VV7.5)-driven Red Fluorescent Protein (RFP) gene flanked by 528 bp sequences representing ORFV119 left and right flanking regions was synthesized and cloned in pUC57 vector (Genscript, Piscataway, NJ). Similarly, to generate revertant viruses OV-IA82-RV119Flag and OV-IA82-RV119LxGxE-Flag, recombination cassettes pRV119Flag-GFP and pRV119LxGxE-Flag-GFP were synthesized containing N-terminally tagged ORFV119 or ORFV119LxGxE sequences, respectively, and a GFP reporter gene, all flanked by approximately 600 bp of homologous sequence on either side to mediate recombination (Genscript, Piscataway, NJ).
To obtain OV-IA82-Δ119, OFTu cells were infected with OV-IA82 (MOI, 1) and transfected with recombination plasmid pΔ119-RFP. To obtain OV-IA82-RV119Flag and OV-IA82-RV119LxGxE-Flag, OFTu cells were infected with OV-IA82-Δ119 (MOI, 1) and transfected with recombination plasmids pRV119Flag-GFP or pRV119LxGxE-Flag-GFP. Recombinant viruses were isolated by limiting dilution and plaque assay using fluorescence microscopy as previously described [36]. Identity and integrity of DNA sequences in purified viruses was confirmed by PCR and DNA sequencing.
Semi-purified viruses were used in infection experiments. OFTu cells infected with OV-IA82, OV-IA82-RV119Flag, OV-IA82-Δ119, or OV-IA82-RV119LxGxE-Flag (MOI, 0.1) were disrupted by three cycles of freeze and thaw at three days p.i. Cellular debris were removed by centrifugation at 1500 rpm for 5 min, and supernatants pelleted by ultracentrifugation at 25,000 rpm for 1 h. Pellets were resuspended in MEM and aliquots frozen at -80°C. Viral titers were obtained by the Spearman-Karber’s 50% tissue culture infectious dose (TCID50) method.
Extracellular enveloped virus (EEV) and intracellular mature virus (IMV) were purified using double sucrose gradient purification protocol with modifications [48]. OFTu cells infected with OV-IA82, OV-IA82-Δ119, OV-IA82-RV119LxGxE-Flag or OV-IA82-RV119Flag (MOI, 0.1) were harvested at three days p.i. and supernatant (EEV) and cellular (IMV) fractions collected following centrifugation at 1,500 rpm for 5 min. Virus in supernatants was pelleted by ultra-centrifugation (25,000 rpm for 1 h), while cellular fractions were disrupted by three cycles of freeze and thaw and centrifuged to remove cell debris. Further purification steps were identical for EEV and IMV. Both preparations were sonicated, pelleted through a 36% sucrose cushion, and purified using double sucrose gradient centrifugation. Virus-containing bands were collected, and virus was recovered by centrifugation and resuspended in 10 mM TrisHcl. Whole cell extracts (60 μg) from OV-IA82-RV119Flag infected OFTu cells, and purified EEV and IMV virion proteins (10 μg) were resolved by SDS-PAGE, blotted to nitrocellulose membranes and probed with primary antibodies against Flag (Genscript, Cat # A00187), ORFV structural protein ORFV086 or Glyceraldehyde-3-Phosphate Dehydrogenase (GAPDH) (sc-25778; Santa Cruz) overnight at 4°C [49]. Blots were incubated with appropriate HRP-labeled secondary antibodies (1:15,000) (anti-mouse, sc-2031 and anti-rabbit, sc-2004; Santa Cruz) for 1 h at room temperature (RT) and membranes were developed using chemiluminescent reagents (SuperSignal West Femto, Thermo Fischer Scientific).
To examine the effect of protein synthesis inhibitor cycloheximide (CHX) on expression of ORFV119 and control cellular protein p53. OFTu cells mock treated or treated with CHX (50 μg/ml) (Sigma-Aldrich, St. Louis, MO) for 30 min were infected with OV-IA82RV119Flag in presence or absence of CHX and harvested at 30 min, 1 h, 1.5 h and 2 h p.i. Total protein extracts (50 μg) were resolved by SDS-PAGE, blotted and transferred to nitrocellulose membranes, probed with antibody against Flag, p53 (sc-6243; Santa Cruz) and GAPDH (sc-25778; Santa Cruz) as described above.
To evaluate ORFV119 expression, OFTu cells were infected with OV-IA82 or OV-IA82-RV119Flag (MOI, 10) for 2 h, 4 h, 6 h, 8 h, 12 h or 24 h p.i. Total protein extracts (50 μg) were resolved by SDS-PAGE, blotted and transferred to nitrocellulose membranes, and probed with primary antibody against Flag and GAPDH. Blots were then incubated with appropriate HRP-labeled secondary antibodies and developed using chemiluminescent reagents.
To assess the subcellular localization of ORFV119, OFTu cells (1.5 x 105 cells/well) were cultured in four-well chamber slides (ibidi, Martinsried, Germany) for 16 h and mock infected or infected with OV-IA82-Δ119 or OV-IA82-RV119Flag (MOI, 10). Cells were fixed at 3 h, 6 h, 12 h, 16 h or 24 h p.i with 4% formaldehyde for 20 min, permeabilized with 0.2% Triton X 100 for 10 min, blocked with 1% bovine serum albumin (Sigma-Aldrich, St. Louis, MO) for 1 h, and incubated with primary antibody against Flag (Cat # A00187-200; Genscript) overnight at 4°C. Cells were then stained with Alexa fluor 594 labelled secondary antibody (Thermo Fisher Scientific, Cat # A-11005), counterstained with DAPI (2 μg/ml) for 10 min and visualized by confocal microscopy (A1, Nikon).
To assess the effect of ORFV119 in virus replication, one-step growth curves were determined in OFTu cells infected with OV-IA82, OV-IA82-Δ119 or OV-IA82-RV119Flag (MOI, 10). Virus yields were quantitated at 6 h, 12 h, 24 h, 36 h, and 48 h p.i. as described above.
To investigate the effect of OV-IA82-Δ119 infection on expression of NF-κB regulated genes, OFTu cells were mock infected or infected with OV-IA82, OV-IA82-Δ119 or OV-IA82-RV119Flag (MOI, 10) and harvested at 2 h p.i. RNA was extracted using RNeasy Mini Kit (QIAGEN, Cat # 74104) and reverse transcribed as previously described [35]. mRNA for select NF-κB regulated genes was quantified using ABI Real time PCR system (Applied Biosystems, Foster city, CA), Power SYBR Green PCR Master Mix (Cat # 4368708, Applied Bio) and the following primers, TNFα (Fwd: 5’-CCTTCAACAGGCCTCTGGTT-3’; Rev: 5’-GTGGGCTACCGGCTTGTTAT-3’) IL1β (Fwd 5’-AAATCCCTGGTGCTGG ATAG-3’; Rev: 5’-GTTGTCTCTTTCCTCTCCTTGT-3’) NF-κB1 (Fwd: 5’- CAGAGAGGA TTTCGTTTCCGT-3’; Rev: 5’-TGCAGATTTTGACCTGAGGGT-3’) IL36α (Fwd: 5’-ATGTCTTCACACCTTGGCAGT-3’; Rev: 5’-ATCGGGTGTACCCTGGATAA-3’) TLR2 (Fwd: 5’-TTGCTCCTGTGACTTCCTGTC-3’; Rev: 5’- GAGCGTCACAGCGGTAGC-3’). Data analysis was performed as previously described [35]. Experiments were conducted with biological triplicates and at least three technical replicates. Statistical analysis was performed by using Student’s t test.
The effect of ORFV119 in TNFα-induced activation of NF-κB was assessed by NF-κB-p65 nuclear translocation assay. HeLa cells transfected with control plasmid (pEGFP-N1) or a plasmid encoding ORFV119-GFP fusion protein (pORFV119GFP) were treated with TNFα (20ng/ml) at 12 h post transfection for 30 min, fixed, permeabilized and blocked as described above, incubated with primary antibody against NF-κB-p65 (Cell Signaling, Cat # 8242s) for 2 h, stained with Alexa fluor 594-labeled secondary antibody (Thermo Fisher Scientific, Cat # A-11005) for 1 h, counterstained with DAPI, and examined by confocal microscopy. Numbers of GFP expressing cells (approximately 300 cells/sample) exhibiting nuclear NF-κB-p65 staining were determined in randomly selected fields and results were shown as mean percentage of GFP/ORFV119GFP expressing cells containing nuclear NF-κB-p65 over three independent experiments. Statistical analysis of data was performed by using the Student’s t test.
To examine the role of ORFV119 on NF-κB-p65 nuclear translocation during ORFV infection, OFTu cells were mock infected or infected with OV-IA82, OV-IA82-RV119Flag, OV-IA82-Δ119 or OV-IA82-RV119LxGxE-Flag (MOI, 10). Cells were fixed at 30 min, 1 h, 2 h, 4 h, and 6 h p.i. and processed for NF-κB-p65 staining as described above. Cells (approximately 300/sample) were randomly selected and scored as mean percentage of cells containing nuclear NF-κB-p65 over three independent experiments. Statistical analysis of data was performed by using Student’s t test.
To investigate the effects of CHX on NF-κB-p65 nuclear translocation during ORFV infection, OFTu cells were pre-treated with CHX for 30 min, infected with OV-IA82 or OV-IA82-Δ119 in presence or absence of CHX, and fixed at 30 min or 1 h p.i. NF-κB-p65 nuclear translocation assay, scoring, quantification and analysis were performed as described above.
To investigate the effect of OV-IA82, OV-IA82Rb-, OV-IA82-RV119LxGxE-Flag and OV-IA82-RV119LxGxE-Flag-Rb- virus infection on NF-κB-p65 nuclear translocation, OFTu or OFTuRb- cells were infected with OV-IA82, OV-IA82Rb-, OV-IA82-RV119LxGxE-Flag or OV-IA82-RV119LxGxE-Flag-Rb- virus (MOI, 10). Cells were fixed at 1 h p.i. and NF-κB-p65 nuclear translocation assay was performed as described above. Cells (approximately 300/sample) were randomly selected and scored for mean percentage of cells containing nuclear NF-κB-p65 over two independent experiments. Statistical analysis of data was performed by using Student’s t test.
Additionally, to assess the effect of pRb on NF-κB-p65 nuclear translocation, Saos-2 cells, a pRb negative cell line, were mock infected or infected with OV-IA82 or OV-IA82-Δ119 (MOI, 50) and fixed at 1 h, 1.5 h, 2 h, 4 h, and 6 h p.i. NF-κB-p65 nuclear translocation assay, quantification and analysis were performed as described above.
HeLa cells transfected with control plasmid (pFlag) or a plasmid encoding ORFV119-3xFlag fusion protein (pORFV119Flag) were treated with TNFα (20ng/ml), and harvested 10 and 20 min post treatment. OFTu cells mock infected or infected with OV-IA82, OV-IA82-RV119Flag, OV-IA82-Δ119 or OV-IA82-RV119LxGxE-Flag (MOI, 10) were harvested at 30 min and 1 h p.i. Total protein extracts (50 μg) were resolved by SDS-PAGE, blotted and transferred to nitrocellulose membranes and probed with specific antibody against phospho-IKKα/β (Ser176/180) (Cat # 2697; Cell Signaling), phospho-IκBα (Ser32/36) (Cat # 9246; Cell Signaling), phospho-NF-κB-p65 (Ser536) (Cat # 3033; Cell Signaling), IKKα/β (sc-7607; Santa Cruz), IκBα (sc-371; Santa Cruz), NF-κB-p65 (sc-7151; Santa Cruz) and GAPDH (sc-25778; Santa Cruz) or Flag. Blots were processed as described above. Protein bands were quantified for densitometric analysis using ImageJ software, Version 1.6.0 (National Institute of Health, Bethesda, MD) and fold changes calculated. Statistical analysis was performed by using Student’s t test.
To investigate the effect of ORFV infection on NF-κB-p65 activation, OFTu cells were infected with OV-IA82 or OV-IA82-Δ119 (MOI, 10) and harvested at 30 min, 1 h, 2 h, 4 h, and 6 h p.i. Total protein extract (50 μg) were resolved by SDS-PAGE, blotted and transferred to nitrocellulose membranes, probed with phospho-NF-κB-p65 and NF-κB-p65 antibodies, and developed as described above.
The effect of ORFV119 on poly(I:C), poly(A:T) or ORFV DNA induced NF-κB-mediated transcriptional activity was investigated using a NF-κB promoter luciferase assay. Firefly luciferase gene under the control of NF-κB promoter (pNF-κB-Luc) and with a plasmid encoding sea pansy (Renilla reniformis) luciferase under the control of herpesvirus TK promoter (pRL-TK) were used in studies. HeLa cells cultured in 12-well plates were co-transfected with the vectors pNF-κB-Luc (450 ng; Clontech, Mountain View, CA), and pRL-TK (50 ng; Promega, Madison, WI) and pFlag or pORFV119Flag. At 24 h after transfection, cells were induced with poly(I:C) (Amersham, Pittsburgh, PA) (500ng), poly(A:T) (Inviogen, San Diego, CA) (750ng) or ORFV DNA (1μg). ORFV DNA was extracted from OV-IA82 virus stock using QIAamp DNA Blood Mini Kit (Qiagen, Germantown, MD). Luciferase activities were determined at 20 h post- induction using the Dual Luciferase Reporter Assay (Promega) and a luminometer. Data were analyzed as previously described [35]. Statistical analysis of the data was performed by using Student's t test.
To investigate the effect of ORFV119 on E2F transcriptional activity in ORFV infected cells, OFTu cells were co-transfected with a pE2F-Luc (450 ng; Signosis, Santa Clara, CA) and pRL-TK (50 ng) plasmids. At 24 h post transfection, cells were mock infected or infected with OV-IA82, OV-IA82Δ119, or OV-IA82-RV119LxGxE-Flag (MOI = 10). Firefly and sea pansy luciferase activities were measured at 1, 2, 4 and 6 h p.i. and expressed as relative fold changes in luciferase activity as described above.
The interaction of ORFV119 or ORFV119LxGxE with pRb and ORFV119 with TRAF2 was assessed in virus-infected cells by co-immunoprecipitation. OFTu cells infected with OV-IA82-RV119Flag or OV-IA82-RV119LxGxE-Flag (MOI, 10) or mock infected were harvested at 12 h p.i. Total protein extraction and co-immunoprecipitation was performed using nuclear complex Co-IP Kit (Active Motif, Carlsbad, CA) following the manufacturer’s instructions. For co-immunoprecipitation, total protein extracts were incubated with antibodies against Flag, TRAF2 and pRb in the high stringency buffer (IP High buffer [Cat # 101676], 300 mM NaCl [Cat # 101684], detergent [Cat # 101683], 1M DTT [Cat # 3483-12-3; sigma], Protease inhibitor cocktail [Cat # P8340; sigma]) overnight at 4°C, and then incubated with pre-washed 50 μl slurry of protein G agarose beads (Cat # 16–266; Millipore) at 4°C for 2 h. Beads were washed four times with high stringency buffer (described above) and bound proteins were eluted in Laemmli buffer. For immunoblot analysis, eluted proteins and control total protein cell lysates were resolved by SDS-PAGE, blotted and transferred to nitrocellulose membranes, probed with antibodies against Flag, TRAF2 and pRb and developed as described above. Light chain specific antibody against Rabbit IgG (Cat # ab99697; Abcam) was used for TRAF2 blots. A flag-expressing protein ORFV113Flag was used as a control for specificity of ORFV119 and TRAF2 interaction.
The interaction of pRb with TRAF2 in the presence of pORFV119Flag also was assessed by co-immunoprecipitation. Antibodies for co-immunoprecipitation included anti-pRb (Cat # 9309s, Cell Signaling), anti-TRAF2 (Cat # sc-876, Santa cruz) and anti-Flag. Co-immunoprecipitations of total proteins extracts were performed as described above. The interaction of ORFV119 or ORFV119LxGxE with putative cellular binding partners was assessed by co-immunoprecipitation in pORFV119Flag or pORFV119LxGxE-Flag transfected HEK 293T or HeLa cells at 12 h post transfection. Antibodies for co-immunoprecipitation included pRb (Fig 3A, B-Cat # ab85607, abcam) (Fig 3E, F-Cat # 9309s, Cell Signaling), TRAF2 (Cat # sc-876, Santa cruz), TAK1 (Cat # sc-7126, Santa cruz), RIP1 (Cat # 3493, cell signaling), TRAF6 (sc-7221) or NEMO (sc-8330). Co-immunoprecipitations of total proteins extracts were performed as described above.
The interaction of ORFV119 with TRAF2 also was assessed in Saos-2 cells, which do not express pRb. Saos-2 cells transfected with control plasmid (pFlag) and pORFV119Flag or co-transfected with pRb (Origene, Rockville, MD) were harvested at 12h post transfection. Co-immunoprecipitations of total proteins extracts were performed as described above. Antibodies for co-immunoprecipitation included anti-TRAF2, anti-Flag and anti-pRb.
Co-immunoprecipitation efficiency was calculated by normalizing the band intensities of co-immunoprecipitated proteins to those corresponding immunoprecipitated proteins and to the expression of corresponding input lysates as previously described [50].
Five-month-old lambs randomly allocated to three experimental groups were inoculated with either OV-IA82-Δ119 (n = 4) or OV-IA82-RV119Flag (n = 4), or mock-infected (n = 3). Following anesthesia, the mucocutaneous junction of the right lower lip near the labial commissure was scarified along a two-centimeter line, and virus inoculum (0.5 ml) containing 107.5 TCID50/ml was applied topically using cotton swabs. In addition, the inner sides of hind limbs were scarified in a five-centimeter line and inoculated as above. Animals were monitored for 21 days for the presence of characteristic orf lesions. Pictures were taken of the labial inoculation sites at days 3, 5, 9, 12, 16 and 21 p.i. and the lesion sizes measured with a ruler. Skin biopsy specimens from hind limb inoculation sites were collected at days 2, 5, 8, 12 and 21 p.i. fixed in 10% buffered formalin, embedded in paraffin, sectioned, and stained with hematoxylin and eosin using standard methods.
All animal procedures were approved by University of Nebraska-Lincoln Institutional Animal Care and Use Committee (IACUC; protocol 1318) and were performed in accordance with the Guide for the Care and Use of Agricultural Animals in Agricultural Research and Teaching.
PPV ORFV119 amino acid sequences were aligned using Clustal Omega (EMBL-EBI). ORFV GeneBank accession numbers are (virus strains in parentheses) AY386263 (OV-IA82), AAP89015 (Orf11), ABA00637 (NZ2), 9AHH34303 (B029), ADY76823 (D1701), NP957896 (OV-SA00), AKU76741 (OV-GO), AKU76609 (OV-YX), AHZ33817 (NA1/11) and KP010356 (OV-SJ1). PCPV and BPSV GeneBank accession numbers are AEL20654 (PCPV F00-120R), AEO18268 (PCPV It1303/05), NP958027 (BPSV BV-AR02) and KM875471 (BPSV BV-TX09c5).
ORFV119 encodes for proteins of 170 to 206 amino acids, with predicted molecular weights of 18.6 to 22.2 kDa and percentage amino acid identities ranging from 77% to 100% (Fig 1). Homologs in Pseudocowpox virus (PCPV) and Bovine papular stomatitis virus (BPSV) are 89% and 54–56% identical to OV-IA82 ORFV119, respectively, whereas no ORFV119 homologue was found in the HL953 strain of parapoxvirus of red deer in New Zealand [51]. ORFV119 lacks homology to known proteins outside the PPV genus, and domains suggestive of protein function. However, a LxCxE motif (OV-IA82 ORFV119 positions 192–196) and a downstream stretch of acidic amino acids usually found in LxCxE motif-containing proteins are located at the C-terminus of the protein and highly conserved in PPV119 proteins (Fig 1). The LxCxE motif is required by several cellular and viral proteins to interact with members of the retinoblastoma family of proteins, which controls important aspects of cell physiology, including cell cycle progression, differentiation, and apoptosis [52]. The ORFV119 LxCxE motif and surrounding sequences follow the pattern XLXCXEXXX, where X should not be a positively charged amino acid (e.g. Lys or Arg) and X should preferably be a hydrophobic residue, which is predicted to bind pRb with high affinity [46]. A predicted mitochondrial localization sequence is found in the N-terminus of most ORFV strains and in PCPV (underlined in Fig 1), but not in BPSV119.
The kinetics of ORFV119 was assessed during ORFV replication in OFTu cells by Western blot using a recombinant virus expressing N-terminally Flag-tagged ORFV119 (OV-IA82-RV119Flag). A protein of approximately 32 kDa was detected at 2 h p.i. with increasing protein levels observed at later time points. At 24 h p.i., the last time point investigated, an additional slightly higher molecular weight species of ORFV119 (approximately 35 kDa) was detected (Fig 2A). The observed protein molecular weight was approximately 10 kDa higher than predicted, suggesting that the protein is post-translationally modified in some manner. Previous reports on ORFV replication have shown that early genes were expressed as early as 1 h p.i., with late genes expressed between 6 to 12 h p.i [35–38]. Newly replicated viral DNA was first detectable at 4 to 6 h p.i. accumulating rapidly between 8 and 16 h p.i [53]. Infectious virus was detected between 16 and 18 h p.i., with continuous virus production until 40 h p.i. [53]. Thus, ORFV119 is a early viral protein.
OV-IA82-Δ119, a virus lacking the ORFV119 gene, exhibited growth kinetics and virus yields comparable to those of wild type OV-IA82 and revertant OV-IA82-RV119Flag viruses in single step growth curves in OFTu cells, indicating that the gene is non-essential for growth in these cells (S1 Fig).
To examine the intracellular localization of ORFV119, OFTu cells were infected with OV-IA82-RV119Flag and examined by confocal microscopy at various times post-infection. ORFV119 staining was not evident at 3 and 6 h time points. At 12 h p.i., weak punctate ORFV119 staining was observed in the cytoplasm and adjacent to the plasma membrane. By 16 h p.i., enhanced ORFV119 staining was evident, and fluorescent circular to ovoid structures (389±30nm) were observed in the cytoplasm and especially in close proximity to the cell membrane. Remarkably, at late times p.i. (24 h), OV-IA82-RV119Flag infected cells exhibited abundant ORFV119 nuclear staining (Fig 2B). Fluorescence was specific for Flag-tagged ORFV119 as no signal was observed in cells infected with OV-IA82-Δ119 at all examined times p.i. (Fig 2B).
To investigate whether ORFV119 interacts with retinoblastoma protein pRb, 293T and HeLa cells were transfected with plasmids pORFV119Flag (ORFV119Flag), pORFV119LxGxE-Flag (ORFV119LxGxE) or control plasmid (pFlag), and protein extracts were prepared at 12 h post-transfection. Reciprocal co-immunoprecipitation assays with either anti-Flag or anti-pRb antibodies demonstrated that ORFV119 co-immunoprecipitates with pRb in both cell types. Co-immunoprecipitation of ORFV119 and pRb, however, was not observed with pORFV119LxGxE-Flag, a plasmid encoding ORFV119 in which C in the LxCxE motif was replaced by G, a change shown to abrogate interaction with pRb (Figs 3A and 3B, S2) [54]. To confirm the interaction in the context of the virus infection, OFTu cells were mock-infected or infected with OV-IA82-RV119Flag or OV-IA82-RV119LxGxE-Flag (MOI, 10), and cell lysates prepared at 12 h p.i. Reciprocal co-immunoprecipitation with either anti-Flag or anti pRb antibodies showed that ORFV119 but not ORFV119LxGxE-Flag co-immunoprecipitates with pRb (Fig 3C–3F). Together, these results indicate that ORFV119 directly or indirectly interacts with pRb. The observation that the integrity of the LxCxE motif is required for the interaction further suggests that ORFV119 might directly bind pRb.
Preliminary RNA-Seq experiments indicated increased transcription of multiple NF-κB regulated genes in cells infected with OV-IA82-Δ119 compared to cells infected with OV-IA82 virus, suggesting that ORFV119 inhibits NF-κB signaling. To rule out any confounding effect from cytokines that potentially might be present in the virus inocula, viruses used in these studies were semi-purified as described in Materials and Methods. Real-time PCR analysis of gene expression showed increased levels of NF-κB-regulated genes TNFα (6.68-fold), TLR2 (6.48-fold), NF-κB1 (3.26-fold) and IL36α (3.7-fold) in cells infected with OV-IA82-Δ119 compared to OV-IA82 at 2 h p.i (Fig 4A).
To assess the effect of ORFV119 on NF-κB-p65 nuclear translocation, OFTu cells were mock infected or infected with OV-IA82, OV-IA82-Δ119, OV-IA82-RV119LxGxE-Flag or OV-IA82-RV119Flag and NF-κB-p65 localization was examined by immunofluorescence. Infection with OV-IA82-Δ119 and OV-IA82-RV119LxGxE-Flag but not OV-IA82 or OV-IA82-RV119Flag led to rapid nuclear translocation of NF-κB-p65 as early as 30 min p.i. (Fig 4B and 4C). The effect was transient as the percentage of cells expressing nuclear NF-κB-p65 returned to those in wild type and revertant virus-infected cells between 1 and 2 h p.i. (Fig 4C, P<0.05). Notably, levels of NF-κB-p65 nuclear translocation observed for OV-IA82-RV119LxGxE-Flag were significantly reduced compared to those observed with OV-IA82-Δ119 (Fig 4C). Consistent with the nuclear translocation kinetics, levels of phosphorylated NF-κB-p65 (Ser536), which accumulates in the cytoplasm prior to nuclear translocation, were increased at early times p.i. (30 min and 1 h) with OV-IA82-Δ119 (S3 Fig). Together, these data show that ORFV119 is a poxviral NF-κB inhibitor acting transiently very early in infection and that ORFV119 LxCxE motif is important for the full inhibitory activity of the protein.
To explore the possibility that pRb transcriptional activity is involved in ORFV119 inhibition of NF-κB signaling in virus infected cells, we examined E2F-mediated gene transcription early in infection. OFTu cells were transfected with a plasmid encoding for a firefly luciferase reporter gene under the control of a E2F-responsive promoter and at 24 h post transfection cells were mock infected or infected with OV-IA82, OV-IA82Δ119 or OV-IA82-RV119LxGxE-Flag. Luciferase activities were measured at 1, 2, 4 and 6 h p.i. Similar low luciferase activity was observed in mock and virus-infected cells at 1, 2 and 4 h p.i. Significantly higher luciferase activity was observed in virus infected cells at 6 h p.i.; however, no significant difference was observed among the three viruses (S4 Fig). Thus, data suggest that ORFV119 mediated NF-κB inhibition does not involve E2F-mediated gene transcription early in infection.
To further investigate the mechanism of ORFV119 in NF-κB inhibition, OFTu cells were mock-infected or infected with OV-IA82, OV-IA82-Δ119, OV-IA82-RV119LxGxE-Flag or OV-IA82-RV119Flag for 30 min or 1 h p.i. and phosphorylation of IKKα/β, IκBα and NF-κB-p65 was assessed by Western blot. Infection by OV-IA82Δ119 and OV-IA82-RV119LxGxE-Flag led to marked and early phosphorylation of IKKα/β (Ser176/180), IκBα (Ser32/36) and NF-κB-p65 (Ser536) compared to OV-IA82 infected cells (Fig 5A). In OV-IA82Δ119-infected cells, relative fold increases of phosphorylated IKKα/β (14.5 and 19.7 fold), IκBα (21.5 and 11.23 fold) and NF-κB-p65 (9.24 and 15.35 fold) were observed at 30 min and 1 h p.i., respectively (Fig 5B and 5C). Similarly, in OV-IA82-RV119LxGxE-Flag-infected cells, relative fold increases of phosphorylated IKKα/β (27.35 and 21.42 fold), IκBα (19.35 and 21.24 fold) and NF-κB-p65 (13 and 13.24 fold) were observed at 30 min and 1 h p.i., respectively (Fig 5B and 5C). These results indicate that ORFV119 inhibits phosphorylation of the IKK complex, a NF-κB activating event.
The observation of ORFV119 staining structures of approximate virion size in infected cells at 16 h p.i. (Fig 2B) together with the early inhibitory effect of ORFV119 on NF-κB signaling suggested ORFV119 may be a virion component available during and/or immediately after virus entry. To examine this possibility, extracellular enveloped virus (EEV) and intracellular mature virus (IMV) were purified from OFTu cells infected with an OV-IA82, OV-IA82-Δ119, OV-IA82-RV119LxGxE-Flag and OV-IA82-RV119Flag virus. As in infected cell extracts at 24 h p.i. (Fig 2A), western blot analysis showed a protein doublet (~32 and 35 kDa) corresponding to ORFV119Flag or ORFV119LxGxE-Flag in both virion fractions. As expected, the ORFV119 protein was not detected in OV-IA82 (ORFV119 lacks Flag tag) and OV-IA82-Δ119 virions (Fig 6A). A control virion core protein ORFV086 was detected as a predominant 21 kDa band in OV-IA82, OV-IA82-Δ119, OV-IA82-RV119LxGxE-Flag and OV-IA82-RV119Flag EEV and IMV virions [49] (Fig 6A). As a control for potential contamination of purified virions with cellular proteins, a GAPDH control was used. GAPDH protein was not detected in OV-IA82, OV-IA82-Δ119, OV-IA82-RV119LxGxE-Flag and OV-IA82-RV119Flag purified EEV and IMV virions (Fig 6A).
To assess whether early inhibition of NF-κB-p65 nuclear translocation by ORFV119 involves de novo viral protein synthesis in the infected cells, OFTu cells were pre-treated with the protein synthesis inhibitor cycloheximide (CHX) for 30 min followed by infection with OV-IA82 or OV-IA82-Δ119 for 30 min and 1 h in presence of the drug. If de novo synthesis of ORFV119 is required for inhibiting NF-κB signaling, then increased levels of NF-κB activation should be observed in OV-IA82 infected CHX treated cells. Low levels of ORFV119 proteins were detected at 30 min and 1 h p.i. likely representing input virion-associated protein; however, ORFV119 protein levels declined in CHX-treated OV-IA82-RV119Flag infected OFTu cells at subsequent times (Fig 6B). Under these treatment conditions, expression of control host protein p53 also was inhibited (Fig 6C). Immunofluorescence analysis showed that inhibition of NF-κB-p65 nuclear translocation was unaltered in OV-IA82 infected cells regardless of CHX treatment (Fig 6D and 6E). Together, these results indicate that ORFV119 is a virion component, and suggest that virion-associated ORFV119 alone is responsible for early inhibition of NF-κB signaling.
To determine if ORFV119 alone is sufficient for inhibiting TNFα-induced nuclear translocation of NF-κB-p65, immunofluorescence assays were performed in HeLa cells transiently expressing GFP or ORFV119-GFP fusion protein (119GFP). Following TNFα induction (30 min) ORFV119-GFP-expressing cells exhibited significantly reduced nuclear translocation of NF-κB-p65 (12.7%) compared to control cells expressing GFP alone (68%) (Fig 7A and 7B, P<0.05).
ORFV119 effect on TNFα induced activation of NF-κB-p65 was further investigated by examining phosphorylation of IKKα/β (Ser176/180), IκBα (Ser32/36) and NF-κB-p65 (Ser536) in HeLa cells transfected with pFlag or pORFV119Flag plasmids. ORFV119 expression markedly reduced the TNFα induced phosphorylation of IKKα/β (20 and 50%), IκBα (25 and 68%) and NF-κB-p65 (40 and 60%) in cells expressing ORFV119Flag compared to control pFlag expressing cells at 10 and 20 min after TNFα induction. (Fig 8A–8D, P<0.05 and P<0.01). Together, results indicate that ORFV119 inhibits TNFα-induced NF-κB signaling by preventing activation of the IKK complex in the absence of any other viral protein.
To examine if ORFV119 also affected poly(I:C), poly(A:T) or ORFV DNA-induced NF-κB transcriptional activity, HeLa cells were co-transfected with pFlag or pORFV119Flag together with a plasmid encoding for a firefly luciferase reporter gene under the control of a NF-κB-responsive promoter. Cells were induced with poly(I:C), poly(A:T) or ORFV DNA at 24 h post-transfection and luciferase activities were determined at 20 h post-induction. No significant effect of ORFV119 expression on poly(I:C), poly(A:T) or ORFV DNA-induced NF-κB-mediated transcription was observed (S5 Fig). Thus, data suggest that ORFV119 functions primarily through TNFα-induced NF-κB signaling.
Given that: 1) ORFV119 interacts with pRb (Fig 3), 2) the ORFV119 LxCxE motif is required for that interaction (Fig 3) and 3) OV-IA82-RV119LxGxE-Flag—a virus containing a mutation in the ORFV119 LxCxE motif that abrogates pRb binding—was unable to efficiently inhibit NF-κB signaling (Fig 5), we examined the involvement of pRb in ORFV119-mediated inhibition of NF-κB signaling and NF-κB-p65 nuclear translocation in cells either lacking or with reduced levels of pRb.
Cells with reduced levels of pRb (OFTuRb-) were prepared from OFTu cells using siRNAs targeting ovine RB1 (see Materials and Methods). pRb protein knockdown of approximately 60% was routinely obtained for RB1 siRNA-transfected cells at 48 h post-transfection (Fig 9A, lanes 2 and 3). Given the possibility that pRb may be associated with the virion due to its interaction with ORFV119, virus stocks were prepared in either OFTu or OFTuRb- cells (see Materials and Methods). To evaluate the effect of reduced pRb levels on ORFV119 ability to inhibit NF-κB signaling, NF-κB-p65 nuclear translocation assays were performed in OFTu or OFTuRb- cells using OV-IA82, OV-IA82Rb- (OV-IA82 virus propagated in cells with reduced pRb levels), OV-IA82-RV119LxGxE-Flag or OV-IA82-RV119LxGxE-Flag-Rb- virus (OV-IA82-RV119LxGxE-Flag virus propagated in cells with reduced pRb levels) virus stocks. As expected from data described above for OV-IA82 (Figs 4 and 5), levels of NF-κB-p65 nuclear translocation at 1 h p.i. following infection of OFTu cells using OV-IA82 virus were low (1.5% positive nuclei). However, significantly increased NF-κB-p65 nuclear translocation was observed for treatment conditions where either OFTuRb- cells (OFTuRb- cells/ OV-IA82 virus) or OV-IA82Rb- virus (OFTu cells/ OV-IA82Rb- virus) were used (6.1 and 6.5% NF-κB-p65 positive nuclei, respectively). Notably, using both OFTuRb- cells and OV-IA82Rb- virus resulted in significantly increased levels of NF-κB-p65 nuclear translocation in infected cells (17.2% NF-κB-p65 positive nuclei) (Fig 9B and 9C). As expected for a ORFV119 protein lacking the pRb binding motif (LxCxE), no significant difference was observed in NF-κB-p65 nuclear translocation for treatment conditions where OV-IA82-RV119LxGxE-Flag or OV-IA82-RV119LxGxE-Flag-Rb- viruses were used to infect either OFTu or OFTuRb- cells (44.25% and 49.2% NF-κB-p65 positive nuclei, respectively) (S6 Fig). Thus, pRb contributes to the ORFV119-mediated inhibition of NF-κB signaling in infected OFTu cells.
Additional experiments examining involvement of pRb in ORFV119-mediated inhibition of NF-κB signaling were conducted in human osteosarcoma Saos-2 cells, a pRb-deficient cell line [55]. pRb was not detected in Saos-2 cell extracts by western blot. (Fig 9A, Lane1). If pRb is mediating ORFV119 inhibition of NF-κB-p65 nuclear translocation, increased NF-κB-p65 nuclear translocation would be expected in OV-IA82 virus infected Saos-2 cells. Saos-2 cells were mock infected or infected with OV-IA82 or OV-IA82-Δ119 and NF-κB-p65 localization was examined by immunofluorescence at indicated times p.i. No significant differences in NF-κB-p65 positive nuclei in OV-IA82-or OV-IA82-Δ119-infected cells were observed at any time post-infection (Fig 10A and 10B). Thus, in the absence of pRb, the early NF-κB inhibitory phenotype observed for OV-IA82 is lost. Together, data using cells with reduced pRb levels or lacking pRb altogether indicate that ORFV119-mediated inhibition of NF-κB-signaling is largely pRb-dependent.
To assess potential interactions of ORFV119 with components of the TNFα-induced NF-κB signaling pathway, co-immunoprecipitation assays were conducted with antibodies against TRAF2, TAK1, RIP1, TRAF6 and NEMO. OFTu cells were mock-infected or infected with OV-IA82-RV119Flag (MOI, 10) and harvested at 12 h p.i. Total cell protein extracts were immunoprecipitated with anti-Flag or anti-target cellular protein antibodies. Reciprocal co-immunoprecipitation demonstrated that ORFV119 co-immunoprecipitates with TRAF2 (Fig 11A and 11B). Similar results were obtained in 293T cells transiently expressing pORFV119Flag (Fig 11E and 11F); however, no interaction was observed between TRAF2 and transiently expressed ORFV119LxGxE (S7 Fig). As a control for specificity of ORFV119 and TRAF2 interaction, a Flag tagged ORFV113Flag was used. No interaction was observed between ORFV113Flag and TRAF2 in OFTu cells infected with OV-IA82-RV113Flag (Fig 11C and 11D) nor in 293T cells transfected with pORFV113Flag (Fig 11G and 11H). In the context of viral infection, reciprocal co-immunoprecipitation of ORFV119 with TAK1, RIP1, TRAF6 and NEMO was not observed. Together, these results indicate that ORFV119 directly or indirectly interacts with the scaffold protein TRAF2. Further, dependence on a LxCxE motif for interaction suggested that pRb might play a role in ORFV119-TRAF2 complex formation.
As ORFV119-mediated inhibition of the NF-κB signaling is pRb dependent (Figs 4, 5, 9 and 10) and ORFV119 interacted with TRAF2 in a LxCxE motif-dependent manner (S7 Fig), we hypothesized that a ORFV119-pRb complex may be required to efficiently interact with TRAF2. To examine this possibility, 293T cells were transfected with pFlag or pORFV119Flag and harvested at 12h p.i. Reciprocal co-immunoprecipitation assays using total cellular protein extracts with either anti-TRAF2 or anti-pRb antibodies demonstrate that TRAF2 and pRb co-immunoprecipitation, while weak or absent in pFlag transfected cells, is enhanced by the presence of ORFV119 in pORFV119Flag transfected cells (Fig 12A and 12B).
To further evaluate a pRb requirement for efficient ORFV119-TRAF2 interaction, co-immunoprecipitation experiments were performed in Saos-2, a pRb-deficient cell line. In contrast to results described above (Fig 11), where ORFV119-TRAF2 interaction was observed in pRb expressing OFTu and 293T cells, no interaction was detected in Saos-2 cells (Fig 12D and 12E). Notably, transfection of Saos-2 cells with both pORFV119 and pRb plasmids restored the interaction as reciprocal co-immunoprecipitation of ORFV119 and TRAF2 were observed (Fig 12F and 12G). Taken together, these data indicate that pRb is important for ORFV119-TRAF2 interaction and, further, they suggest that a ORFV119-pRb complex may be required for efficient interaction with TRAF2.
The contribution of ORFV119 to virus virulence was investigated in sheep, a natural ORFV host. Animals were inoculated with OV-IA82-Δ119 (n = 4), OV-IA82-RV119Flag (n = 4) or PBS (control group, n = 3) in the right labial commissure and the inner side of the thighs, and disease course was monitored for 21 days. All virus-inoculated animals developed clinical orf as evidenced by erythema, papules, pustules, and scabby tissue deposition on the lips (S8A Fig). However, beginning on day 3 p.i., lesions in sheep inoculated with OV-IA82-RV119Flag were 25% to 90% larger and exhibited more extensive pustules and scabby tissue deposition than those inoculated with deletion mutant virus (S8A and S8B Fig). By day 16 p.i., lesions in all sheep inoculated with OV-IA82-Δ119 were resolved. In contrast, three of four sheep inoculated with OV-IA82-RV119Flag (sheep # 79, #101, and #639) still exhibited well-defined lesions at day 16 p.i. and regressing lesions still were present in two of the animals (#79 and #101) at day 21 p.i. (end point of the experiment) (S8A Fig). No lesions were observed in PBS-inoculated controls. Histological examination of punch biopsies collected from the thighs at various times post-infection showed no significant differences in time to onset and type of skin changes between groups of virus-inoculated animals. Overall, results indicate that lesion development is more restricted and time to resolution is reduced in sheep inoculated with OV-IA82-Δ119 compared to revertant virus. Thus, ORFV119 contributes to ORFV virulence in the natural host.
The NF-κB signaling pathway plays key roles in the skin by regulating innate immune responses, inflammation, cell survival and cell proliferation [56–58]. Crucial cytoplasmic and nuclear events in the signaling pathway are targeted by various proteins encoded by the highly epitheliotropic ORFV [35–37]. Here, we describe an ORFV protein, ORFV119, that interacts with pRb and prevents activation of NF-κB signaling very early during infection. Inhibition of NF-κB-signaling by ORFV119 is largely dependent on its ability to interact with pRb, which prevents activation of the IKK complex and downstream NF-κB signaling.
ORFV119 was shown to interact with pRb in infected and uninfected cells, indicating that other ORFV proteins are not required for the interaction. Similar to oncoproteins of small DNA viruses, binding to pRb was dependent on the ORFV119 LxCxE motif since a CxG substitution in the motif abrogated the interaction completely (Fig 3). This suggests that ORFV119, like small DNA virus oncoproteins, might directly bind pRb. Interestingly, another highly epitheliotropic poxvirus, molluscum contagiosum virus (MCV), encodes a protein (MC007L) unrelated to ORFV119 that localizes to mitochondria and interacts with pRb through an LxCxE motif [59]. The function of MC007L in MCV infection is unknown.
pRb has not been previously implicated in antiviral responses against poxviruses; however, it has been associated with NF-κB modulation by other viruses. For example, transiently expressed adenovirus E1A, a pRb-binding oncoprotein, was shown to inhibit NF-κB-dependent transcription induced by TNFα, the effect being dependent on E1A/pRb interaction [44]. And pRb was shown to be required for the activation of the NF-κB pathway in response to vesicular stomatitis virus infection, although viral mechanisms involved were not described [45].
Through association with pRb, ORFV119 inhibits NF-κB signaling by preventing IKK complex activation (Fig 5), with ORFV119-TRAF2 interaction likely underlying the inhibition (Fig 11). Results suggest that a ORFV119-pRb complex may be required for efficient interaction of ORFV119 with TRAF2, leading to inhibition of NF-κB signaling (Fig 12 and S7 Fig). TRAF2 is a RING finger protein recruited to TNF receptors to regulate NF-κB signaling both positively and negatively [60]. Viral proteins that interact with TRAF2 activate or inhibit NF-κB signaling. For example, the MCV protein MC159 was shown to interact with both TRAF2 and NEMO and to inhibit NF-κB activation [61], while Kaposi’s sarcoma associated-herpes virus (KSHV) vFLIP and rotavirus VP4 proteins interact with TRAF2 activating NF-κB signaling [62,63]. The effect of these interactions on TRAF2 function remains unknown. Conceivably, viral proteins might interfere with NF-κB signaling by modulating TRAF2 E3 ubiquitin ligase activity and/or by affecting TRAF2 scaffold functions.
ORFV119 is a virion protein (Fig 6) functioning very early in infection (≤ 30 min) to inhibit NF-κB signaling. Observed early IKK complex activation on infection of cells with virus lacking ORFV119 (OV-IA82Δ119) indicates ORFV119 inhibits NF-κB signaling induced by an early infection event. Early inhibition of IKK complex activation during infection, ORFV119-TRAF2 interaction in infected cells, and a proposed role for TRAF2 in poxvirus entry [64] suggest that ORFV119 may be inhibiting ORFV entry mediated activation of NF-κB signaling. The importance of viral inhibition of NF-κB activation very early in ORFV infection is further supported by the presence of a second virion-associated NF-κB inhibitor ORFV073 (Fig 13). ORFV073 inhibits NF-κB signaling by preventing activation of IKK complex through interaction with NEMO, the regulatory subunit of the IKK complex [38]. As early infection events, including those related to cell entry, are likely conserved among poxviruses [65], early inhibition of NF-κB signaling in poxvirus infected cells may be of greater biological significance than currently appreciated.
Notably, at late times post-infection (≥24 h p.i) ORFV119 also is observed in the nucleus of infected cells (Fig 2), suggesting that in addition to the early virion-associated NF-κB inhibitory function described here, the protein may perform additional functions in the infected cell. The fact that most of pRb localizes to the cell nucleus raises the question as to whether ORFV119 interacts with nuclear pRb at late times post-infection affecting pRb functions such as transcriptional control and cell cycle regulation.
Other ORFV NF-κB inhibitors have been detected in the nucleus. ORFV002 localized to the nucleus and inhibited nuclear phosphorylation of NF-κB-p65 by interacting with mitogen stimulated stress kinase 1 (MSK1) [36,39]. And, ORFV073, a virion-associated NF-κB inhibitor, also was shown to localize to the nucleus yet a nuclear function for the protein has not been described [38]. Among other poxviruses, vaccinia virus protein K1 localized to the nucleus where it prevented acetylation of NF-κB-p65 [66].
Examination of published parapoxviral genomes shows that a genomic region encompassing ORFV119 is affected sporadically by gaps or deletions of as much as 5.7 kbp [67–69]. Notably, while the extent of the deletion varied considerably, ORFV119 was always affected. The genomic sequence of the PCPV F00-120R strain, for example, was found to contain a 5.5 kbp DNA deletion that removed genes 116–121 [67]. Retrospective PCR analysis, however, strongly suggested that loss of the 5.5 kbp region occurred during virus passage in tissue culture [68]. Deletions affecting this region were also found in ORFV isolates OV-NP (5.7 kbp deletion) and OV-SJ1 (1.5 kbp deletion), which were isolated from goat lesion material [69]. Although retrospective PCR analysis was not conducted, these virus isolates (cell culture-passage 6) were fully attenuated following inoculation of goats, further supporting the notion that the genomic loss occurred during early passage in cells, and indicating that important virulence functions are encoded within this genomic region. Genomic deletions resulting in complex fragmentation of the BPSV ORFV119 homolog have also been reported [70]. An explanation for instability of this genomic region upon virus passage in cell culture is lacking; however, it is possible that loss of these genes may be advantageous for virus growth in cell culture.
Here, infection with a deletion mutant virus lacking ORFV119 resulted in restricted lesion development and reduced time to resolution compared with revertant virus infection in the natural host. The attenuated phenotype of infection likely indicates improved host control of ORFV infection in the absence of ORFV119 (S8 Fig). Our results are at variance with a previous report where no differences in virulence and pathogenesis were observed between wild type and ORFV119 deletion viruses [71]. The discrepancy likely reflects differences in viral strains (IA82 vs SHZ1), extent of genomic deletion, and/or in vitro conditions for virus propagation.
Overall, ORFV pathogenesis studies in the natural host using viruses lacking single NF-κB inhibitor genes have shown a remarkable spectrum of phenotypes, ranging from wild type disease (ORFV002, ORFV024) to moderate (ORFV073, ORFV119), or marked attenuation (ORFV121) [35–38]. The multiple NF-κB inhibitors encoded by a poxvirus together with the possibility of overlapping or complementing functions may explain these observations. Alternatively, specific poxviral NF-κB inhibitors may exert only subtle and perhaps transient host range effects on specific infected cells or the infected tissue microenvironment. Regardless, the impact of these subtle changes on viral fitness in nature may be difficult to fully ascertain under experimental conditions.
|
10.1371/journal.pgen.1001314 | The Evolution of Host Specialization in the Vertebrate Gut Symbiont Lactobacillus reuteri | Recent research has provided mechanistic insight into the important contributions of the gut microbiota to vertebrate biology, but questions remain about the evolutionary processes that have shaped this symbiosis. In the present study, we showed in experiments with gnotobiotic mice that the evolution of Lactobacillus reuteri with rodents resulted in the emergence of host specialization. To identify genomic events marking adaptations to the murine host, we compared the genome of the rodent isolate L. reuteri 100-23 with that of the human isolate L. reuteri F275, and we identified hundreds of genes that were specific to each strain. In order to differentiate true host-specific genome content from strain-level differences, comparative genome hybridizations were performed to query 57 L. reuteri strains originating from six different vertebrate hosts in combination with genome sequence comparisons of nine strains encompassing five phylogenetic lineages of the species. This approach revealed that rodent strains, although showing a high degree of genomic plasticity, possessed a specific genome inventory that was rare or absent in strains from other vertebrate hosts. The distinct genome content of L. reuteri lineages reflected the niche characteristics in the gastrointestinal tracts of their respective hosts, and inactivation of seven out of eight representative rodent-specific genes in L. reuteri 100-23 resulted in impaired ecological performance in the gut of mice. The comparative genomic analyses suggested fundamentally different trends of genome evolution in rodent and human L. reuteri populations, with the former possessing a large and adaptable pan-genome while the latter being subjected to a process of reductive evolution. In conclusion, this study provided experimental evidence and a molecular basis for the evolution of host specificity in a vertebrate gut symbiont, and it identified genomic events that have shaped this process.
| The gastrointestinal microbiota of vertebrates is important for nutrient utilization, resistance against pathogens, and immune maturation of its host, but little is known about the evolutionary relationships between vertebrates and individual bacterial members of these communities. Here we provide robust evidence that the evolution of the gut symbiont Lactobacillus reuteri with vertebrates resulted in the emergence of host specialization. Genomic approaches using a combination of genome sequence comparisons and microarray analysis were used to identify the host-specific genome content in rodent and human strains and the evolutionary events that resulted in host adaptation. The study revealed divergent patterns of genome evolution in rodent and human lineages and a distinct genome inventory in host-restricted sub-populations of L. reuteri that reflected the niche characteristics in the gut of their particular vertebrate hosts. The ecological significance of representative rodent-specific genes was demonstrated in gnotobiotic mice. In conclusion, this work provided evidence that the vertebrate gut symbiont Lactobacillus reuteri, despite the likelihood of horizontal transmission, has remained stably associated with related groups of vertebrate hosts over evolutionary time and has evolved a lifestyle specialized to these host animals.
| Vertebrates are associated with trillions of microbes, the majority of which inhabit the digestive tract [1]. Research has led to an appreciation of the importance of these microbial communities, revealing substantial roles in development and performance of the host [2], [3]. As vertebrates evolved, they did so in association with microbes, and these reciprocal interactions have shaped both the attributes of the microbiomes and the phenotypic complexity of the host species [4]. It is conceivable that the beneficial functions of the gut microbiota conferred important selective traits during vertebrate evolution [3], [5]. A joint evolutionary trajectory between host and microbes is evident in anatomical features of vertebrates (rumen, cecum) which allow bacterial fermentations that provide energy to the host and an intensive gut associated immune system that is in place to maintain beneficial microbial communities [6], [7]. These features serve as clear testimony that we cannot attempt to understand the evolution of vertebrates without considering their microbial partners [1], [8].
Comparative analysis of genomes of bacteria originating from human hosts, greatly facilitated through the Human Microbiome Project, provided important insight into the adaptations and ecological roles of different microbial species in the human gut [9]–[11]. Despite these advances, very little is known about the evolutionary strategies of vertebrate gut symbionts. It is often postulated that the evolution of gut microbes involved coevolution of individual lineages with their host species, which is supported by the presence of phylotypes that are specific to particular vertebrate species [3]. However, conclusive evidence for stable associations of specific lineages with vertebrate hosts over evolutionary time-scales has not been provided by 16S rRNA data. Patterns of community similarity provide evidence for codiversification of entire gut communities with their hosts, which suggests that there are host-specific evolutionary interactions between mammals and their microbiomes [4]. In addition, some gut microbes are highly host specific, such as Helicobacter pylori, which has been used to track human migrations over long-time spans [12]. However, many microbial lineages in the mammalian gut are shared across host species [4], implying that some members of the gut microbiota are generalists that pursue promiscuous lifestyles. Such an evolutionary strategy is exemplified by commensal Escherichia coli, which have a broad host range and alternate between niches within the environment and their vertebrate hosts [13], [14]. To date, there are very few vertebrate gut symbionts for which host specificity has been clearly established. Furthermore, little is known about the mechanisms by which gut microbes, for whom symbiotic life is facultative and which have ample opportunities for horizontal transmission, can evolve stable associations with their host species that would allow for reciprocal evolutionary interactions between bacterial lineages and host genotypes.
The Gram-positive bacterium Lactobacillus reuteri is an excellent model organism to study the evolutionary strategy of a vertebrate gut symbiont as this species inhabits the gastrointestinal tract (GIT) of mammals as diverse as humans, pigs, mice, and rats as well as different species of birds. In rodents, pigs, and chickens, it is one of the dominant species in the GIT and forms biofilm-like associations with the stratified squamous epithelial lining of the proximal regions of the digestive tract [15]–[19]. We recently observed that strains of L. reuteri from global sources comprised distinct phylogenetic clusters that can be detected with Multilocus Sequence Analysis (MLSA) and Amplified Fragment Length Polymorphism (AFLP), and these clades show significant association with host origin [20]. The population structure suggests a stable association of L. reuteri with particular vertebrates over evolutionary time and the emergence of host adapted subpopulations. In addition to the genotypic patterns, an adaptive evolutionary process is also reflected by the phenotypic characteristics of L. reuteri strains in terms of ecological performance in the gut and adhesion to epithelial cells [20]–[27]. However, the molecular basis for these host adaptations is still unknown, and it is unclear to what degree the lifestyle and evolution of L. reuteri have remained restricted to particular hosts.
Genomic approaches in combination with experiments in animal models offer mechanistic insight into the evolution and ecology of microbial symbionts of vertebrates. In this study, we used such an approach and showed that only rodent isolates of L. reuteri colonize the gut of reconstituted Lactobacillus-free (LF) mice in high numbers, while isolates from humans, swine, and chicken form either lower populations or fail to colonize. We determined the genome sequence of the rodent isolate L. reuteri 100-23 and performed a comparative genomic analysis with the genome of the human isolate F275. A microarray analysis using genes representative of both strains was used to probe 57 L. reuteri strains, revealing specific gene combinations in host-adapted lineages of L. reuteri. Further genomic comparisons of nine isolates across five MLSA lineages confirmed the microarray data and further allowed the identification of the evolutionary processes that resulted in host-specific genomic features. The ecological significance of rodent-specific genes was demonstrated in gnotobiotic mice, where perturbations in 7 out of 8 genes unique to the rodent lineage resulted in impaired ability to propagate in the murine host.
We tested the ability of thirteen L. reuteri isolates originating from different vertebrate hosts (mouse, rat, human, chicken, and pig) to colonize the digestive tract of LF mice (Table 1). LF mice were previously derived from conventional mice by treatment with penicillin followed by reconstitution with cultures of microbes, non-cultivable microbes attached to epithelial cells, and cecal contents from mice treated with chloramphenicol [28]. These mice are maintained under gnotobiotic conditions and contain a gut microbiota functionally equivalent to conventional animals, but without any lactobacilli. Therefore, these mice are different from germ-free animals in that they allow the investigation of adaptations of Lactobacillus strains under more ecologically relevant conditions. The experiments revealed that only strains originating from mice and rats reached colonization levels after two weeks that were equivalent to those of Lactobacillus populations in conventional mice (Table 1). Isolates from other hosts formed either smaller populations (≤106 cells per gram in the forestomach, <105 cells per gram in the cecum) or were not detectable. These findings provide experimental evidence that the evolution of L. reuteri with rodents has resulted in host specialization. This reinforces previous findings which identified a population structure composed of subpopulations that were to a large degree host confined [20]. The colonization phenotypes in LF mice and the population structure further imply that host restriction does not occur at the host genus level, as L. reuteri isolates from rats showed excellent colonization of mice and group with isolates from mice in MLSA lineages [20].
The human isolate L. reuteri strain CF48-3A1 was detectable in the forestomach and cecum of female mice, but not of male mice, 14 days after gavage. The numbers of CF48-3A1 detected in the gut of female mice were considerably lower than those attained by strains of rodent origin, supporting host specificity of L. reuteri, but the apparent influence of gender was nevertheless noteworthy. The molecular mechanism that supports this gender-specific persistence of lactobacilli in the gut is unknown. Speculatively, gender-specific effects might involve the relative frequencies of receptors for Lactobacillus adhesins on the forestomach epithelium of female compared to male mice. Other factors may include relative retention times of digesta in the large bowel coupled with coprophagy, along with differences in coprophagous habits. These possibilities, however, remain unexplored for the present.
The species L. reuteri shows a significant degree of genetic variation, especially between strains from different MLSA lineages [20]. Sequence divergence can confound the CGH data as it impairs hybridizations. This was apparent because even though hybridizations were very reliable for the genomes of the reference strains 100-23 and F275 (>96% accuracy), the error rate was approximately 18.5% for strain CF48-3A of lineage VI. Therefore, to confirm the findings obtained with the CGH analysis and to gain further insight into the distribution of host-specific gene content throughout the entire L. reuteri population, we performed additional genomic comparisons in combination with PCR. First, we generated draft genome sequences (>15× coverage) of two additional rodent strains (lpuph1 and MLC3) and one pig strain (ATCC 53608). We then determined the presence of the host-specific genes identified by CGH and the pdu-cbi-cob-hem cluster in all available L. reuteri genomes (100-23, lpuph1, MLC3, ATCC 53608, F275, MM4-1a, MM2-3, ATCC55730, and CF48-3A). These genomes represent five MLSA lineages, lineages I and III (rodent), lineage II (human), lineage IV (pig), and lineage VI (poultry/human), and the genome characteristics are shown in Table S6. The average nucleotide identity (ANI) of a core set of genes within these L. reuteri genomes and L. vaginalis is shown in Table S7. An ANI of >95% was determined in all the L. reuteri genome comparisons, providing additional evidence that these strains, despite their considerable genomic differences, fall within what is currently considered to be one prokaryotic species [46].
As shown in Figure 4A, the genomic comparisons confirmed the findings obtained with the CGH analysis. The pdu-cbi-cob-hem cluster was detected in all human isolates (MLSA lineage II and VI) and the pig isolate ATCC 53608 (MLSA lineage IV), but it was only present in one of the three rodent strains. The urease cluster was strictly conserved among the three rodent strains and absent in all other genomes, while the surface proteins and the TCS2 cluster were to a large degree specific to rodents but more variable. The SecA2 and xylose clusters were detectable in rodent and porcine strains but completely absent in strains from lineage II and VI. The SPS and TCS1 clusters showed a much higher variability among rodent strains and several of the genes were datable in the lineage VI and IV strains, while most of the genes were absent in human lineage II strains. Consistent with CGH, the ABC transporter was specific to strain 100-23. To study the distribution of host-specific genomic features throughout the L. reuteri population, PCR was used to determine the presence of genes encoding SecA2, several surface proteins (Lr_70131, Lr_70581, Lr_70697, Lr_69916), UreC (the urease alpha subunit), and PduC (diol/glycerol dehydratase encoded by the pdu-cbi-cob-hem cluster) in 88 L. reuteri strains (Table S3). The results are shown in Figure 4B in a phylogenetic context. This analysis confirmed that several of the key genetic determinants identified by CGH are to a large degree associated with specific MLSA lineages and vertebrate hosts.
The functions of the genetic features associated with L. reuteri ecotypes are reflective of their lifestyle in respective hosts. In rodents, L. reuteri adheres directly to the stratified squamous epithelium present in the murine forestomach and forms thick cell layers that show characteristics of biofilms [25], [26], [36], [39]. Accordingly, several of the rodent-specific surface proteins are predicted to function as adhesins or mediators of biofilm formation, and the SecA2 system is likely involved in the secretion of some of these proteins (e.g. Lr_70902). Other factors, such as the TCS2, fructosyltransferase (Ftf), IgA specific metallopeptidase, and the urease cluster are likely to play roles in biofilm formation, cell aggregation, and the mitigation of low pH and exposure to IgA, respectively. It is striking that several of the genes identified as rodent-specific by CGH were also detectable in at least some strains that originate from pigs and poultry (Figure 2 and Figure 3), reflecting the similar lifestyle of rodent, porcine, and poultry lactobacilli which all form biofilm-like associations with epithelial surfaces in the proximal GIT [18], [57], [58].
The genome content of strains within the human MLSA lineage II is strikingly different when compared to other L. reuteri lineages. The absence of many genetic features involved in biofilm formation and adhesion reflects the lifestyle of L. reuteri in the human gut. Squamous stratified epithelia are absent, and epithelial cell layers rich in lactobacilli equivalent to those found in animals have not been described in the human GIT [19]. The genome content of strain F275 suggests a planktonic lifestyle in more distal regions of the human gut and limited, if any, interactions with the gut epithelium. This lifestyle would require fast multiplication rates, which could explain the absence of the large surface proteins in lineage II strains, which are likely to be a significant energetic burden. In addition, easily accessible nutrients are in low supply in the human colon having been absorbed in the small intestine, and the ability of L. reuteri to use 1,2-propanediol as an energy source through the pdu-cbi-cob-hem cluster might therefore constitute an important colonization factor in the human gut. The production of reuterin, which is also conferred by this cluster, might contribute to the fitness of L. reuteri in the human gut through inhibition of competitors in the same niche (as reviewed in [5]). Enzymes involved in 1,2-propanediol utilization and reuterin formation require Vitamin B12 as a co-factor [43], [55]. The synthesis of Vitamin B12 is also encoded by the pdu-cbi-cob-hem cluster, and it appears to be an important colonization factor for colonic bacteria, as demonstrated for Bacteroides thetaiotaomicron [59].
Although it is striking how gene content of L. reuteri lineages reflects niche characteristics in particular hosts, differences in gene frequencies within populations can arise not only through natural selection but also random genetic drift. In order to test whether the rodent specific genes were of ecological significance in the GIT of mice, we investigated the fitness of isogenic mutants of strain 100-23C in the gut of LF mice in competition with the parental strain. Eight genes representing major groups of genetic functions among the lineage-specific genes were selected for these experiments: Lr_70902 (serine-rich surface protein), Lr_70770 (putative adhesin), Lr_70892 (SecA2 translocase), Lr_70890 (Asp2, involved in SecA2 transport system), Lr_70894 (SecY2, involved in SecA2 transport system), Lr_70430 (two-component system histidine kinase), Lr_70458 (ABC-type multidrug transport system), Lr_70532 (ABC-type transporter of TCS2). This selection included sets of genes with high (Lr_70902, Lr_70770, Lr_70892, Lr_70890, Lr_70894, Lr_70532) and low conservation (Lr_70430, Lr_70458) among rodent strains. Further, it included genes with a variety of functions, such as adherence, secretion of surface proteins, and environmental sensing. As shown in Figure 7, when the parental strains and their mutant derivatives were introduced into LF mice, seven out of the eight mutants had impaired ecological fitness. The most significant defect in competitive fitness was caused through the inactivation of Lr_70890, Lr_70894, and Lr_70902, which are all associated with the secA2 operon. The only gene that did not contribute to ecological performance was Lr_70770, which encoded a putative adhesin. Given the large number of putative adhesins in the genome of L. reuteri 100-23 (Table 3), it is possible that redundancy exists in mechanisms that confer adherence.
The genetic architecture reflected in the genomes of the rodent and human-adapted L. reuteri strains 100-23 and F275 provides insight into the evolutionary processes that underlie host specialization. First, it is clear that that LGT played an important role in the evolution of L. reuteri, as many of the host-specific functions were found to be encoded on putative genomic islands or on regions with lost synteny between the two related strains (Figure 1). In addition, the pdu-cbi-cob-hem cluster, which is absent in most rodent strains, has previously been identified to be a horizontal acquisition of L. reuteri [34], [60]. Therefore, the acquisition of novel genetic material could have led to phenotypic innovations in L. reuteri and might have allowed lineages to become associated with vertebrates, radiate among vertebrate hosts, or to switch hosts during evolution.
However, closer scrutiny of the gene organizations at the loci of genomic difference between L. reuteri strains 100-23 and F275 suggested an additional mechanism of genome evolution. As shown in Figure 6 and Figures S3, S4, and S5, the pdu-cbi-cob-hem, SecA2, urease, and SPS clusters as well as the xylose operon and most of the surface proteins (Lr_70770, Lr_70131–Lr_70137 cluster, Lr_69916, Lr_70580/Lr_70581 cluster, and Lr_71380) are all replaced or interrupted by mobile genetic elements (e.g. putative IS elements and phage related genes) in the genomes of strains 100-23 and F275, respectively. These findings indicate that most of the lineage-specific genes in rodent and human lineage II strains were ancestral and appeared to be jettisoned after divergence of the two lineages. This means that genome evolution of L. reuteri strains is, in many cases, a process associated with gene deletions, possibly caused by mobile genetic elements that mediated rearrangements through recombination. Functional gene loss is a common mechanism that underlies host specialization in both pathogenic and symbiotic bacteria from various phylogenetic groups [30]–[32], [61]. Our findings indicate that it also plays an important role for host specialization in L. reuteri, especially in the human lineage II.
Given the long time periods involved and the lack of intermediate steps, it is currently difficult to reconstruct the evolutionary processes that have shaped L. reuteri subpopulations. However, the genomic comparisons of strains spanning several MLSA lineages allowed us to pinpoint some specific key events in the evolution of the species. The pdu-cbi-cob-hem cluster appears to be an ancient acquisition of L. reuteri as it is distributed through the entire phylogenetic spectrum of the species (Figure 4). This is in accordance with conclusions based on codon adaptation index and GC content [60]. The cluster is absent in most rodent strains, and the analysis of the loci in strain 100-23 indicated that the cluster was deleted through the action of mobile elements (Figure S4A). It is one of only very few examples of gene loss exclusive to this lineage, making it interesting to speculate as to why its function may be obsolete for the success of L. reuteri in the rodent forestomach.
The SecA2 cluster, which is highly conserved in rodent and porcine strains (Figure 4), appears to be a later acquisition of L. reuteri, as all but one strain from the lineage VI lack this cluster. As shown in Figure 6, there is no evidence for deletion of the cluster in lineage VI strains, while strains of MLSA lineage II showed evidence for deletion through mobile genetic elements. This indicates that the cluster was acquired after diversification of more recent lineages from lineage VI. The acquisition of the SecA2 cluster might have been a pivotal innovation of L. reuteri strains to colonize the gut of mammals. The biological significance of the SecA2 cluster for life in the rodent gut was clearly demonstrated in our competition experiments in LF mice, in which inactivation of four different genes in strain 100-23C associated with this cluster (Lr_70890, Lr_70892, Lr_70894, and the surface protein Lr_70902) had the most detrimental effects when compared to the other mutants tested (Figure 7).
The comparison of the genomes of L. reuteri 100-23 and F275 revealed evidence for only one event of LGT since the split of the two lineages. The surface protein Lr_70697 is arranged in an island with two transposases and two phage integrases next to a transfer RNA gene (tRNA-Val) in the genome of 100-23. This locus is intact in the genomes of F275, CF48-3a, and ATCC55730. Therefore, this gene cluster was likely acquired by a recent ancestor of 100-23 and inserted into a tRNA-Val gene, as described for islands in meseorhizobia and several pathogenic bacteria [62]. As with mesorhizobia, insertion of the cluster in L. reuteri left the entire tRNA gene (a Thr-tRNA) intact upon integration, whereas a small part (22 nucleotides in L. reuteri) became duplicated as a direct repeat (see Figure S5B). Both CGH (Figure 3) and PCR (Figure 4B) analyses showed that Lr_70697 was to a large degree specific to strain 100-23, supporting the hypothesis that this cluster was a recent genomic acquisition.
A recent study on the genomes of human L. reuteri strains revealed a closed pan-genome, with individual strains contributing to a very small number of new genes [35]. Our CGH analysis supported these observations, showing similar genome content and little genetic diversity among strains belonging to the human MLSA II lineage (Figure 2). However, strains from other hosts, and especially rodents, possessed a more variable gene content, and the majority of the rodent-specific genes detected by CGH were not conserved among rodent strains (Figure 2). Comparisons of the genomes of the three rodent L. reuteri strains 100-23, MLC3, and lpuph1 confirmed that rodent strains possess a larger pan-genome with a gene repertoire that extends beyond that of individual strains. Open pan-genomes have been described for many bacterial species, and they consist of a ‘core genome’ (genes present in all strains) and an ‘accessory’ genome (genes variable among strains) [14], [63], [64].
As shown in Figure 8, the three rodent strains shared around 1463 of the predicted protein coding genes. Of this core genome, only 25 genes were unique to rodent strains (Figure 8), confirming the CGH analysis in that only a small number of rodent specific genes are conserved among strains. Each strain possessed a significant proportion of genes that were absent in the other strains (528 proteins in 100-23; 235 in MLC3; and 309 in lpuph1), confirming the more variable gene pool among rodent strains. Of note, a large proportion of these genes were not found in the genomes of L. reuteri strains from non-rodent hosts (Figure 8). This rodent-specific accessory genome was comprised, apart from a large portion of mobile genetic elements, of the same functional groups as the genes identified by the CGH analysis to be rodent-specific (Figure 8 and Table S8). Thus, many of the rodent-specific surface proteins, glycosyltransferases involved in SPS synthesis, transport proteins, and regulatory proteins that are present in the genome of strain 100-23 are substituted by genes that are predicted to perform similar functions in strains MLC3 and lpuph1. The genomic comparisons revealed only one group of rodent-specific genes that were absent in the genome of 100-23 and were therefore not detected by CGH. These genes were all CRISPR-related and are likely to be involved in phage resistance.
It is important to point out that the seven rodent-specific genes that contributed to ecological fitness in colonization experiments in LF mice (Figure 7) were not conserved among rodent strains. A key conclusion of this study is therefore that adaptive traits that allow life in the murine gut are encoded by a rodent-specific accessory genome and that different combinations of these genes promote successful colonization. This of course begs the question of why plasticity is favored in the rodent L. reuteri population but not the human lineage II. It has been suggested that bacterial accessory genomes encode special ecological adaptations in genes that remain unbounded and can be more rapidly incorporated where and when they become advantageous [65], [66]. Thus, the larger gene pool within in the rodent L. reuteri population might be sampled by individual cells through LGT to form the basis for adaption to environmental fluctuations. The population genetic structure of L. reuteri {Oh, 2010 #380} and the colonization phenotypes in LF mice imply that lineages maintained a broader host range and evolved with at least two diverse host genera (Mus and Rattus), and probably many species (around 40% of the world's mammalian species are rodents). Such an evolutionary strategy would require individual cells to adapt not only to physiological and immunological differences of individual animals but different host genera, and the larger accessory genome of the rodent L. reuteri population might reflect a higher diversity among the host population.
The ecological forces that have shaped the autochthonous L. reuteri population in the human GIT appear fundamentally different than those in other hosts. Strains within the human-specific MLST lineage II, although obtained from world-wide locations, are highly conserved genetically and are clonally related [5], [20], suggesting a recent population bottleneck, founder effect, or clonal expansion. The genomic comparison of strain 100-23 and F275 further revealed that human strains underwent a process of reductive genome evolution. These evolutionary patterns resemble to some degree those found for genetically monomorphic pathogens, such as Yersinia pestis and Mycobacterium leprae [67]–[69], which show high clonality and genome evolution characterized by functional gene loss.
We can only speculate on what caused the specific genetic features of the human L. reuteri population. It has been suggested that the evolution of monomorphic pathogens was influenced by an expansion of the human population within the last 10,000–20,000 years, which possibly led to a significant increase of the available niche and a restriction to the human host [67]. The population bottleneck might also have been caused through altered transmission dynamics and changes in the human environment, which could have reduced the effective population size [5]. Low population sizes favor genetic drift and can lead to both decreased genetic variability [70] and the loss of genes (even if slightly beneficial) [71]. Alternatively, L. reuteri might have been acquired by humans more recently. Restriction to particular hosts or host changes have both been accompanied with a clonal population structure and functional gene loss, especially those associated with the cell envelope [72]–[75]. As described above, the genome of F275 shows clear evidence for pseudogene formation, gene deletions, and genome reduction, and although we do not yet know the causes of these patterns, the dramatic removal of surface proteins L. reuteri F275 suggests a process by which to bypass deleterious responses from the human immune system.
The gut of vertebrates provides a multitude of nutrient rich habitats inhabited by complex microbial communities, whose composition is remarkably host specific and stable [4], [76]. These communities are important for normal development and growth of the host, but must be acquired during each generation as most vertebrates are essentially germ-free at birth. This process is poorly understood but relevant as benefits to the host are increased by the correct selection of true mutualists and their stable maintenance over evolutionary time [5], [29]. This study clearly established host specificity within the species L. reuteri through a combination of animal experiments and evolutionary genomics, and it revealed a first insight into the genomic changes that underlie host adaptation. Host specificity of L. reuteri in the mouse gut appears to be mediated to a large degree by specific adhesins. However, other factors are likely to contribute to host specificity and include adaptations to the environmental conditions (the urease cluster, Ig-A protease, factors for biofilm formation) and their regulation (possibly through TCS involved in quorum sensing).
In the last decade, our understanding of genome evolution in host-associated bacteria has advanced dramatically due to the availability of hundreds of sequenced genomes [77]. Common trends have been identified and range from those observed in obligate bacterial symbionts, who show extensive reductive genome evolution, to those of facultative symbionts with free-living stages, who have expanded genomes and high levels of LGT [77]–[79]. Genome evolution of L. reuteri shares some patterns that have been observed in other host associated bacteria, and the findings suggest an evolutionary intermediate transitioning from a facultative to an obligate, mutualisitic lifestyle, which concurs with the observed degree of host specialization. Accordingly, the high amount of mobile elements (e.g. IS elements) in L. reuteri genomes is a characteristic that is often associated with recent obligate host associations in bacteria [31]. Although mobile elements are common in all L. reuteri genomes, there are distinct trends of genome evolution in the rodent and human lineages, with the former possessing a large and adaptable pan-genome while the latter being subjected to a process of reductive evolution. These distinctions are likely related to differences in the microbe's host range and the ecology and genetic diversity of the host population.
Taken together, the results of this study revealed host adapted subpopulations among the species L. reuteri whose genome content reflected niche characteristics in their respective hosts. Although physiological and immunological differences of vertebrates were likely to constitute important selective forces that drove this specialization, the distinct patterns of genome evolution in rodent and human lineages suggest that the evolutionary trajectories of a vertebrate gut symbiont are not only determined by microbial competition but also by the ecology and evolutionary history of the host.
All animal experiments were approved by the Otago University Animal Ethics Committee (approval number 2/09).
Lactobacillus reuteri strains used in this study are listed in Table S3 and were grown anaerobically on MRS (Difco) plus 5g/L Fructose and 10g/L Maltose at 37°C or 45°C (where indicated). Escherichia coli EC1000, which was used for cloning vectors for gene inactivation in L. reuteri, was grown aerobically in LB media at 37°C. Erythromycin (200 µg/mL for E. coli, 5 µg/mL for Lactobacilli), kanamycin (40 µg/mL for E. coli), and chloramphenicol (7.5 µg/mL for lactobacilli) were used for the propagation of recombinant strains. L. reuteri 100-23C, which is a plasmid-free derivative of strain 100-23, was used to test the ecological relevance of selected genes (see below).
LF mice were raised under gnotobiotic conditions, and the absence of lactobacilli was regularly tested by anaerobic culture on Rogosa SL agar for 48 hours. Mice (around 6 weeks of age) were inoculated by gavage on a single occasion with ∼106 Lactobacillus cells that had been cultured anaerobically in MRS medium overnight. Cell numbers of lactobacilli in fecal samples, the forestomach, and the cecum were determined by quantitative culture on Rogosa SL agar as described previously [39].
Sequencing of L. reuteri 100-23 (rodent isolate) and DSM20016T (human isolate F275) genomes were accomplished through the Community Sequencing Program of the Joint Genome Institute (Walnut Creek, CA), using a combination of whole-genome shotgun sequencing of three libraries with 3-Kb, 8-Kb, and 40-Kb DNA inserts. The genomes were further sequenced using a Roche Genome Sequencer (FLX-GS) to reduce the amount of contigs, and gaps were closed manually by sequencing PCR products generated from the ends of contigs. This process resulted in a circular genome for DSM20016T and two scaffolds for 100-23 (729,351 bp and 1,576,206 bp). PCR reactions to amplify the DNA between these scaffolds failed on several attempts, probably due to the highly repetitive nature of the termini. Genomes were annotated using the JGI annotation pipeline, and the genome sequences have been deposited in GenBank under the accession numbers NC_009513 (strain DSM20016T) and NZ_AAPZ00000000 (strain 100-23).
The genomes of L. reuteri lpuph1 and MLC3 (rodent isolates) were sequenced to draft status at the Core for Applied Genomics and Ecology (CAGE, University of Nebraska, Lincoln, USA) with a standard shotgun library prep kit of the Roche GS FLX Titanium series. The genome of L. reuteri ATCC53608 (pig isolate) was sequenced at the Biotechnology and Biological Research Council's TGAC (The Genome Analysis Centre, Norwich Research Park, UK). Sequencing resulted in 185,905 (lpuph1), 115,542 (MLC3), and 617,241 (ATCC53608) reads that were assembled de novo using the gsAssembler (Newbler) module of the GS-FLX Off-Instrument Sofware Suite. This resulted in draft sequences of 127, 126, and 142 contigs, for lpuph1, MLC3, and ATCC53608 respectively. The draft sequencing resulted in a final coverage of around 30 fold (lpuph1), 20 fold (MLC3), and 100 fold (ATCC53608). The genome characteristics are listed in Table S6. Genome sequences for mlc3 and lpuph are available at DDBJ/EMBL/GenBank under the accession numbers AEAW00000000 and AEAX00000000, respectively. Genome sequences for ATCC 53608 are available at EMBL under the accession numbers CACS01000001 to CACS01000142.
L. reuteri F275 was isolated in the 1960s and later deposited in both the Japanese and German culture collections. The genome sequence of the strain deposited in the Japan Collection of Microorganisms (JCM1112T) was recently published [34]. In the present study, the strain deposited in the Deutsche Sammlung von Mikroorganismen und Zellkulturen (DSM20016T) was sequenced. As described by Morita and coworkers, strain DSM20016T has undergone genomic modifications of its genome during in vitro propagation [34], and as a consequence, JCM1112T contains two additional regions (a total of 40 kb) compared to the genome of DSM20016T. For the analysis and comparisons of gene content in 100-23 and F275, we used the genome annotations of strains 100-23 and DSM20016T, as they were both done with the JGI annotation pipeline. The genes that were encoded by the extra sequence identified in JCM112T were added to the gene set of DSM20016T and considered in all comparisons.
The Integrated Microbial Genomes (IMG) system of the JGI was used to analyze genome characteristics and compare genomes [80]. Unique and conserved genes between strains 100-23 and F275 were determined by the BLASTP algorithm implemented in the IMG Phylogenetic Profiler with a maximum E value of 1e−5 and a minimum amino acid identity of 70%. Whole genome comparisons were completed using the Artemis Comparison Tool (ACT) [81]. For this analysis, the two remaining scaffolds of the 100-23 genome were combined and the chromosome replication initiation site was identified. Visual genome comparisons of the genomes of strains 100-23 and JCM1112T were prepared by using ACT (BLASTN with a score cutoff of 1900). Alien_hunter was used to identify areas affected by LGT [82]. This program utilizes interpolated variable order motifs to identify regions of the genome with atypical sequence composition and thus integrates codon, and nucleotide compositional changes into its predictions.
BLASTP was used to identify homologous genes (>70% identity, >70% coverage) found in all L. reuteri strains and the closely-related L. vaginalis. Nucleotide sequences for these 169 orthologous genes were individually aligned in MUSCLE and concatenated and used to calculate the average nucleotide identity (ANI) as described by Konstantinidis and Tiedje [83]. The same BLASTP criteria were applied to determine the core and accessory genomes of L. reuteri strains (Figure 8).
Spotted microarrays were designed to contain probes representing all detected open reading frames (ORFs) of the rodent strain 100-23 and ORF unique to strain F275 when compared to 100-23. The phylogenetic profiler tool of the IMG platform was used to identify unique genes of F275 (using the sequence of strain DSM20016T) with a maximum E value of 1e−10 and an amino acid percentage of less than 90%. This analysis revealed 403 unique genes for F275. Probes (60 bp) were designed for all ORFs of sufficient size by using Oligo Array 2.1 [84]. Multiple probes were designed for genes of 100-23 larger than 4.5 kb (3 per gene). In total, the probe set comprised 2192 probes representing 2170 genes of strain 100-23 and 320 probes representing 320 genes of F275. Oligomers were synthesized by Invitrogen (Carlsbad, CA USA) and spotted in duplicate using an Omnigrid arrayer (Gene Machines, San Carlos, California).
L. reuteri strains used in microarray typing are listed in Table S3, which include 24 isolates from humans (including DSM20016T), 24 from rodents (including 100-23), 5 pig isolates, and 5 chicken isolates. Chromosomal DNA of bacteria was prepared as described by Oh and coworkers [20]. DNA of strains 100-23 and DSM20016T was mixed at a 1∶1 ratio, and 2 µg was amplified by random priming using Cy5 dye-labeled nucleotides and the BioPrime DNA labeling kit (Life Technologies, Rockville, Md.) to generate the reference DNA. Test DNA was generated by random priming PCR from all strains with Cy3 dye-labeled nucleotides. Concentrated labeled products from each reference test pair were hybridized in formamide-containing buffer (Array Hyb Low Temp; Sigma, St. Louis, Mo.) for 4 h at 47°C. Slides were washed once each in 1× SSC (0.15 M NaCl plus 0.015 M sodium citrate)−0.03% sodium dodecyl sulfate, 0.2× SSC, and finally 0.05× SSC. Fluorescence intensities of the array addresses were determined using a GenePix4000 multicolor microarray scanner and GenePix software (Axon Instruments, Union City, CA USA).
Genome content comparisons were performed using MARKFIND, as described by Zhang et al 2003 [45]. MARKFIND performs a cluster analysis based on genome polymorphisms implementing the unweighted pair-group method with arithmetic means (UPGMA). The program also uses an algorithm for sorting polymorphic characters in the binary strings relative to user-specified groups of taxa. For those genes being represented by three probes (i.e. large surface proteins), the gene was marked as present if at least two probes showed hybridization.
The accuracy of the microarray analysis was tested by comparing the results obtained by hybridizations with whole genome BLASTN comparisons. BLASTN was performed by comparing all gene sequences that are represented on the microarray slide with the genome sequences of L. reuteri 100-23 (rodent III cluster), lpuph1 (rodent I), F275 (human cluster II), and CF48-3A1 (human/chicken cluster IV). Genes were considered present if BLASTN resulted in alignments with more than 70% identity and at least 50% coverage to the query sequence. This analysis revealed that the microarray analysis had a very high accuracy for the two reference strains, showing >96.7% and 96.3% accuracy for 100-23 and DSM20016T, respectively. The accuracy dropped for strain lpuph1 to 92.5%, and it was lowest for strain CF48-3a (81.5%). So as expected, the accuracy of the microarray analysis decreased as gene divergence between the test and reference strains increased (see Table S7 for ANIs).
Eighty-eight L. reuteri strains from all known MLSA lineages of the species (Table S3) were tested by PCR for the presence of representative rodent-specific genes: surface proteins (Lr_70131, Lr_70581, Lr_70697, Lr_69916), secA2 (Lr_70892), pduC (encoding a subunit of diol/glycerol dehydratase, the first enzyme in the propanediol fermentation/reuterin formation pathway), and ureC (encoding the urease alpha subunit). Primers were constructed based on the sequences of all strains that possessed the gene to first amplify an internal region of the gene, and second, to target the flanking genes and amplify the loci in which the gene was located in strain 100-23. The PCRs were carried out in 25 µl volumes containing 20 pmol of each primer and 0.5 units of Taq polymerase (Takara). After an initial denaturation for 3 min at 94°C, the reaction mixtures were cycled 30 times at 94°C for 30 s, 30 s at appropriate annealing temp, and 72°C for 3 min, followed by a 7-min extension at 72°C. Primer sequences and annealing temperatures are listed in Table S9.
The contribution of genes for ecological performance was determined as described previously [37]. Briefly, genes were inactivated in strain 100-23C by insertional mutagenesis by inserting the plasmid pORI28 into the target sites, which renders the mutant erythromycin-resistant. 1∶1 mixtures of mutant and wild type were administered by intragastric gavage to anesthetized LF mice. The mice were killed 7 days after inoculation, and lactobacilli were cultured quantitatively from the forestomach and cecum. To determine the proportion of the mutant strain, lactobacilli were quantified on agar plates with and without erythromycin.
|
10.1371/journal.pbio.1000343 | Binding Site Turnover Produces Pervasive Quantitative Changes in Transcription Factor Binding between Closely Related Drosophila Species | Changes in gene expression play an important role in evolution, yet the molecular mechanisms underlying regulatory evolution are poorly understood. Here we compare genome-wide binding of the six transcription factors that initiate segmentation along the anterior-posterior axis in embryos of two closely related species: Drosophila melanogaster and Drosophila yakuba. Where we observe binding by a factor in one species, we almost always observe binding by that factor to the orthologous sequence in the other species. Levels of binding, however, vary considerably. The magnitude and direction of the interspecies differences in binding levels of all six factors are strongly correlated, suggesting a role for chromatin or other factor-independent forces in mediating the divergence of transcription factor binding. Nonetheless, factor-specific quantitative variation in binding is common, and we show that it is driven to a large extent by the gain and loss of cognate recognition sequences for the given factor. We find only a weak correlation between binding variation and regulatory function. These data provide the first genome-wide picture of how modest levels of sequence divergence between highly morphologically similar species affect a system of coordinately acting transcription factors during animal development, and highlight the dominant role of quantitative variation in transcription factor binding over short evolutionary distances.
| The differentiation of cells, tissues, and organs during animal development is established by a process in which genes that control cell identity and behavior are turned on and off at specific times and places. This process is choreographed, to a large extent, by a collection of proteins known as transcription factors that bind to specific sequences in DNA and thereby modulate the expression of neighboring genes. Because of the central role that transcription factors play in shaping organismal form and function, they have long been suggested to be major players in phenotypic evolution. However, we have a poor understanding of how changes to DNA affect transcription factor binding in living systems. Here, we use a combination of biochemical and genomic techniques to compare, between two closely related species of fruit flies in the genus Drosophila, the binding of six transcription factors that help establish the characteristic segments that form along the anterior-posterior (head to tail) axis in developing flies. We show that the patterns of transcription factor binding between these closely related species are broadly conserved, consistent with the nearly identical development and appearance of these species. However, we also show that, whereas the DNA changes that have accumulated between these species in the five million years since their divergence—roughly one difference per 10 basepairs—have not altered the locations where these factors bind, they have had a considerable effect on the amount of factor bound at each site across a population of embryos. We can trace these quantitative differences in binding to the gain and loss of the short sequences known to be preferentially recognized by these factors, giving us key insights into the effect that sequence changes have on the biochemical events that underlie animal development.
| Despite four decades of interest in the evolution of transcriptional regulation, we still have a poor understanding of the molecular bases for regulatory divergence and the constraints under which cis-regulatory sequences evolve. Most regulatory sequences appear to be under strong selection to maintain their transcriptional output, and as a result, binding sites for the sequence-specific transcription factors that regulate mRNA synthesis are preferentially conserved [1],[2]. However, even in regulatory sequences with highly conserved function, transcription factor binding sites can be gained and lost over time at a high rate, leading to considerable differences in the composition and arrangement of binding sites between even closely related species [2]–[10]. Whether and how this binding site turnover affects transcription factor binding, and what the consequences of changes in binding on transcription might be, remains unknown.
After years in which the study of regulatory evolution was primarily a computational exercise, a series of recent studies have compared genome-wide in vivo binding of transcription factors in the same conditions or tissues of related species [11]–[14]. Among yeasts of the genus Saccharomyces [11],[12] and between human and mouse [13],[14], a substantial fraction of experimentally observed interactions between transcription factors and DNA are species-specific. While these differences could, in principle, be due to divergence of transcription factors and other trans-acting factors, binding differences appear to be driven primarily in cis [13], suggesting that differences in the sequences, and not the factors binding to them, drive the divergence in binding. Species-specific binding is generally associated with the gain/loss of sequence motifs recognized by the relevant factor [11],[14], although the correlations are weak.
Here we examine how the binding of a group of six factors that direct temporal and spatial patterns of gene expression along the anterior-posterior (A-P) axis during early development differs between Drosophila melanogaster and its sister species D. yakuba. These two species, whose genomes have been fully sequenced [15],[16], diverged only five million years ago [17]. They are separated by a molecular distance less than half that between mouse and human [18], and D. yakuba orthologs of virtually all D. melanogaster genomic regions can be readily identified and aligned. Though there are some subtle changes in the levels of expression of key regulators between these species (our unpublished data), there is little difference in either their spatial expression patterns or those of their targets, a product at least in part of strong selection to maintain them [10].
In our earlier work on the binding of these factors in D. melanogaster, we showed that they bind to an overlapping set of thousands of genomic regions in vivo [19],[20], as has subsequently been observed for many other animal transcription factors [21]. A wealth of evidence suggests that, at least in D. melanogaster, and probably generally, only the several hundred most highly bound regions are directly involved in transcriptional regulation, with the remainder having a different, or more likely no, function [19],[20].
Thus these two fly species provide an ideal opportunity to study the effects of modest sequence divergence on transcription factor binding, its origins in changes in genomic sequence, and its functional consequences. We expected binding differences between D. melanogaster and D. yakuba to be more modest than those observed between mouse and human, or between Saccharomyces species. However, we hoped that the more modest differences in their genomes would improve our ability to associate sequence and binding divergence, and that our earlier work establishing the relationship for these factors between binding levels and regulatory function would provide an invaluable context for analyzing the functional consequences of the binding differences we observe.
We collected embryos spanning the hour immediately preceding gastrulation, during which the regulatory events that initiate patterning along the A-P axis occur, from large laboratory populations of D. melanogaster (Oregon R) and D. yakuba (Tai8E2), and immediately immersed them in formaldehyde to covalently stabilize protein-DNA interactions. We isolated chromatin from each species, and immunoprecipitated bound regions with affinity purified rabbit polyclonal antibodies raised against the D. melanogaster versions of the key A-P regulators: Bicoid (BCD), Hunchback (HB), Krüppel (KR), Giant (GT), Knirps (KNI), and Caudal (CAD). We sequenced recovered fragments on an Illumina Genome Analyzer II, mapped reads to the reference genomes of each species using Bowtie [22], and calculated fragment coverage based on the average fragment length in the immunoprecipitated material (Table 1 gives statistics on the numbers of sequenced and mapped tags for each experiment in both genomes). We normalized the signal between species so that the average binding across a diverse set of known targets of these factors was equal, and projected the normalized binding signals from each species onto the coordinates of a whole-genome pairwise alignment computed using Mercator [23] and FSA [24].
We began our analysis of binding divergence by examining previously identified targets of these six factors (Figure 1) [19]. Overall, binding to these loci is remarkably similar between species (Figure 1A), with both bound regions and their relative binding intensities similar for most factors across most loci (we note that the normalization did not consider the pattern of binding—just levels across the locus). Several types of binding divergence are evident, including the gain or loss of binding (Figure 1B), shifts in the precise location of binding (Figure 1C), and changes in the height, but not location, of binding peaks (Figure 1D). Note that with only two species it is impossible to determine whether features found in one species but not the other represent gains or losses relative to the common ancestor.
To get a comprehensive picture of this variation, we identified genomic regions significantly bound by each factor independently in both species using MACS [25] with total chromatin as controls (“Input” controls). While the signal-to-noise ratio was higher in D. melanogaster than in D. yakuba, yielding more detected peaks in D. melanogaster for all factors (Table 2), the relative numbers of peaks identified for each factor were similar in the two genomes. For each bound region in each species we quantified the number of sequence reads observed in the region in the source species and in the orthologous region of the other species.
Before analyzing species-specific differences in binding in detail, we sought to establish that observed differences between D. melanogaster and D. yakuba were due to bona fide interspecies differences in binding, and not experimental noise or bias. As in our earlier work [19], we performed chromatin immunoprecipitation (ChIP) with antibodies recognizing different domains of several of the targeted proteins. Antibodies recognizing the N and C terminal domains of HB and KR give nearly identical results in both species, with correlations of 0.99 and 0.97 over called peaks for these antibodies in D. melanogaster and correlations of 0.98 and 0.94 in D. yakuba (Figure S1). In contrast, correlations of the binding levels for the same antibody between species range from 0.57 to 0.75, demonstrating that the binding differences are not due to experimental noise. It is also highly unlikely that these difference arise from differential affinity of the antisera for transcription factors from the two species, as there are three or fewer amino acid changes between the species for five of the six factors (KR has more than 10).
We were also concerned that differences in sequence composition or chromatin state might interact with the sequencing protocol to produce apparent interspecies differences in binding. To evaluate this, we examined genome-wide variation in the total chromatin sequencing signal (“Input” control). There was no correlation between the Input signal and binding in the individual species (Figures S2 and S3) and only a weak correlation between interspecies differences in ChIP and Input signals (from 0.04 to 0.14; Figure S4). This latter correlation is likely due to interspecies differences in chromatin state and corresponding differences in fragmentation [26], but is too weak to explain the observed differences in factor binding.
Unlike in the yeast and mammalian studies described above, the gain or loss of bound regions between D. melanogaster and D. yakuba was rare, with fewer than 1% to 5% of peaks (depending on the factor) found in one species clearly absent or displaced in the other (Table 2). The rate of gain/loss near known targets of the A-P factors was similar to the genome-wide rate (Table 2).
The measured binding at orthologous regions bound in both species varied considerably (Figures 2, S5, and S6) both in the highly bound regions that our previous studies suggested are functional targets of these factors [19],[20] and in the poorly bound regions that likely are not. The more highly bound regions showed a greater total variation in binding (Figure S7), with the normalized divergence (difference in binding over average binding level) roughly constant across binding levels (Figures 3 and S8) and relative to annotations (Figure S9).
The divergence was marginally lower within the 44 characterized D. melanogaster cis-regulatory modules (CRMs) known to be targeted by one or more of these factors (correlation rA-P from 0.62 to 0.91 compared to 0.57 to 0.75) [27] and in peaks near genes (within 10 Kb of the 5′ end) known to be regulated by these A-P factors (correlation rA-P from 0.59 to 0.92, depending on the factor).
We sought to determine the extent to which sequence changes in the bound regions drove quantitative differences in binding. We first examined overall measures of sequence divergence. Levels of single-nucleotide divergence (sequence identity) and frequency of insertions and deletions in the 100 base pairs centered on the inferred peak of binding exhibited only low to moderate correlations with binding divergence (0.07 to 0.24; Figures S10 and S11), consistent with our expectation that changes to specific short sequences, rather than entire regions, would have a disproportionate effect on binding.
We next sought to identify short sequences (e.g., transcription factor binding sites) whose gain or loss was associated with changes in binding levels. We devised an unbiased statistical approach that assessed the impact on binding of changes to a short sequence (word) by comparing the distribution of binding intensities in all bound regions where the word was conserved to the distribution in all bound regions where the word was present in one species but not the other (defining bound regions as the 100 bp centered on peaks of maximal binding intensity). If alterations to a word affect binding, then these distributions should be different. We identified such words (which we call divergence-driving words, or DDWs) by comparing the conserved and non-conserved distributions for all 16,384 words of length 7 bp and picking those that showed a statistically significant difference. We found DDWs for four of the six factors, and in each case, virtually all of these DDWs matched the known sequence specificities of the corresponding factor (Figure 4).
To quantify the fraction of binding divergence that is explained by the DDWs, we developed a method that used the gain and loss of DDWs to predict binding divergence between the species. For each factor for which we had identified DDWs, we built a simple linear model relating the divergence of DDWs in a bound region to interspecies difference in binding at that bound region. In the model, each divergent DDW in a bound region contributed a fixed amount to the predicted binding difference, with the effect of multiple divergence DDWs adding independently. The contribution of each DDW was determined by a regression using the least angle regression method [28] with extensive cross-validation (see Methods).
The correlations between predicted and observed divergence in binding of single factors across all peaks with at least one DDW in the two genomes ranged from 0.3 for HB to 0.41 for BCD (Figures S12–S27). While far from perfect, these correlations demonstrate that changes in a highly restricted collection of sequences (for example, BCD has only a single 7 bp DDW) drive an appreciable fraction of binding divergence between species. We additionally performed the same predictions using words derived from the in vitro factor binding specificities described by [29]. The correlations between predictions and observations ranged from 0.18 for HB to 0.39 in BCD, similar to or lower than the correlations resulting from our DDWs (unpublished data).
We investigated whether the lack of a strong relationship between probable enhancer function and quantitative conservation of binding was associated with similar trends at the sequence level. For each factor for which we identified DDWs, we quantified motif enrichment and conservation as a function of the level of transcription factor occupancy in D. melanogaster. Motif enrichment and conservation were elevated within bound regions above background levels across the genome (Figure 5). The fraction of peaks with motifs showed a weak dependence on binding levels, with the most strongly bound regions exhibiting the greatest density of motifs. The level of conservation of these motifs was weakly correlated with overall binding levels, consistent with our observation that quantitative divergence in binding strength decreased slightly near genes regulated by these factors.
In our initial comparison of binding between species, we noticed that increases in binding of a single factor were often correlated with increases in binding of many other factors (Figures S28–S33). For example, changes in the binding of KR correlated with changes in the binding of other factors with r = 0.36 (KNI) to 0.62 (CAD), and such coordinated changes are recapitulated for all pairs of factors. This widespread correlated change suggests a factor-independent mode of binding divergence.
To obtain an unbiased assessment of the extent of these correlated changes in binding, we quantified binding divergence for all six factors in all regions significantly bound by any factor and performed principal component analysis (PCA), a method for analyzing variation between many factors simultaneously rather than only pairs of factors, on these data (Figure 6A). The first principal component, which represents the most significant axis of variation in the dataset, has the same direction and similar magnitude for all six factors, demonstrating that a pan-factor coordinated binding shift is the dominant driver of A-P factor binding divergence (this principle component explains 38% of the overall variation in binding between the species). A similar effect was observed when we performed PCA on the binding levels in each species independently (Figure 6B and 6C), suggesting that a common effect is responsible for much of the variation in binding both between species and within a single genome.
The single-genome PCA revealed several interesting factor-specific correlations: increases in binding of the repressor GT are associated with decreases in binding of the activator HB (PC2 in Figure 6B), increases in HB are associated with decreases in BCD (PC3 in Figure 6B), etc. As expected, given the overall similarity of binding between the species, the single-genome PCA analyses of D. melanogaster and D. yakuba yielded essentially identical results.
To investigate whether the features captured by these different principal components are related to specific sequences, we applied the same motif discovery method described above to projections of the binding data along each of the principal components shown in Figure 6A. We discovered substantially more motifs in this analysis (Figure 7) than in the single-factor analyses, likely because of the increased statistical power derived from considering all regions bound by any, as opposed to a single, factor.
Interestingly, one of the words whose divergence is associated with the first principal component is the “TAGteam” motif, CAGGTAG [30], the binding site for Zelda, an activator of the early zygotic genome [31]. Zelda's mechanism of action is unknown, but the strong correlation between gain and loss of its binding site with variation in changes in binding of all factors supports a direct or indirect role for Zelda in nucleosome positioning and chromatin remodeling.
We have provided the first genome-wide picture of how modest levels of sequence divergence between highly morphologically similar species affect a system of coordinately acting transcription factors during animal development. The pervasiveness of changes in binding levels highlights the importance of treating transcription factor binding as a quantitative trait. This is in contrast to previous interspecies studies of in vivo binding [11],[13],[14], which focused on the gain and loss of bound regions.
Although the gain/loss of bound regions is often associated with the gain/loss of cognate binding sites, we establish here that the primary biochemical effect of binding site turnover is to alter levels of binding to existing bound regions. What remains unclear is whether and how the small changes in the amount of bound factor affect transcription, and under what circumstances such changes have demonstrable phenotypic consequences. That there are no clear differences in binding divergence between functional and non-functional targets, and that the most strongly bound (and presumably functional) regions show more absolute, and equal relative, divergence suggests that much of the variation we observe between these two species does not significantly affect organismal fitness, consistent with the observation that binding site gain and loss in active CRMs often does not result in significant changes in regulatory function [6],[10]. This is, however, far from definitive proof, and there are many alternative explanations for this observation, such as compensatory or directional selection on binding to functional regulatory sequences. Exploring the molecular and developmental consequences of quantitative variation in transcription factor binding will be an exciting avenue for future research.
Although we and others have previously described a broad correlation between factor binding in the Drosophila blastoderm [19],[20],[32] and other systems, we were surprised at how strong this common effect was in driving interspecies binding differences. It is tempting to speculate that this effect arises from interspecies differences in chromatin structure, which could readily produce such a uniform effect on the binding of a large collection of factors. However, the only direct evidence that chromatin differences may cause binding differences is the association of the gain/loss of CAGGTAG with the increase/decrease of the common factor signal (PC1). CAGGTAG is the binding site for the factor Zelda, a general zygotic activator of transcription with a putative association with chromatin. CAGGTAG, however, explains only a small fraction of the common signal.
Indirect cooperativity between factors, in which binding of one factor alters chromatin state and thereby facilitates the binding of other factors, may also play a significant role in binding divergence. We have examined only six of the approximately 40 transcription factors active at this stage of embryogenesis. Given the extensive cross-binding of A-P and dorsal-ventral regulators [19],[33], it is likely that changes in the binding of some of these additional factors influences the A-P factor binding.
Although D. melanogaster and D. yakuba are closely related, we were not always able to accurately identify orthologous sequences, largely due to ambiguities in the draft D. yakuba assembly. Even where the orthology of regions was unambiguous, and despite this close evolutionary distance, base-level alignments were frequently uncertain. Our analysis of sequence-specific effects required a precise alignment, and inevitable alignment errors will make nucleotide-level analysis of regulatory changes challenging for more distantly related species (although the alignment accuracy estimates produced by FSA may help to identify reliably aligned loci).
Several aspects of this experiment should help direct future efforts to use comparative ChIP-Seq to study the relationship between sequence and binding divergence. The widespread quantitative binding divergence between D. melanogaster and D. yakuba demonstrates that even relatively similar species can be used to study binding changes. Indeed, given the magnitude of the binding divergence that we observe, we expect there to be quantitative differences between D. melanogaster and more closely related species, such as D. simulans, as well as among D. melanogaster individuals. While comparisons with more distantly related species will likely reveal greater binding divergence, and will help explain how such divergence affects expression and phenotype, the difficulties with aligning genomes at this distance, and comparing embryonic stages, may render sequence-based analyses less powerful.
Even though we were working with very similar organisms, with similar timing and structure of embryonic development, there were undoubtedly subtle differences in our sampling of developmental stages in the two species. Because transcription factor binding is dynamic, such sampling differences have the potential to manifest themselves as apparent interspecies differences in binding. We do not believe this effect was significant in our data, however, as it is unlikely that this type of false-positive binding divergence would be associated with the specific sequence changes that we repeatedly observed. Nonetheless, this will be a major difficulty in future studies, especially when developmentally and morphologically different organisms are compared, as precisely those changes that make such comparisons interesting also make them far more difficult.
Both D. melanogaster and D. yakuba embryos were collected from population cages for 1 h, and then allowed to develop to late stage 4 and early stage 5 before being harvested and fixed with formaldehyde. The embryos from the two species developed very similarly, and the aging times to reach the desired age were 2 h for D. melanogaster embryos and 1 h and 45 min for D. yakuba embryos. The staged embryos were harvested and cross-linked with formaldehyde, and the chromatin was isolated through CsCl gradient ultracentrifugation essentially as previously described [19].
The chromatin used for immunoprecipitation was fragmented through sonication using a Branson Sonifier 450 to an average fragment size of 225 to 250 bp, which is shorter than the average size of chromatin used in our previous ChIP-chip experiments [19]. ChIP was carried out using affinity purified rabbit polyclonal antibodies, and for two of the factors, HB and KR, two affinity purified antibodies that recognize non-overlapping parts of each factor were used. These antibodies and the ChIP procedure were identical to those described in [19].
The DNA libraries for sequencing were prepared from the ChIP reaction and from Input DNA following the Illumina protocol for preparing samples for ChIP sequencing of DNA using the reagents provided in the genomic-DNA or ChIP-DNA sample preparation kits, with some modifications. Briefly, the DNA fragments were converted to phosphorylated blunt ends using T4 DNA polymerase, Klenow DNA polymerase, and T4 polymerase kinase, a 3′ A base overhang was added using Klenow DNA polymerase exo- (3′ to 5′ exo minus), and Illumina adapters were ligated to the fragments. We carried out the PCR step for enrichment of adapter-modified DNA prior to the library size selection, and limited the amplification to 10–13 cycles to minimize the potential bias associated with PCR amplification. After the amplification step, we size-selected DNA fragments of 150–500 bp (including the adapter sequence) for BCD, HB, GT, and KNI samples, and 200–500 bp for KR and CAD. The DNA library was quantified by QPCR using ABI Power SYBR green PCR master mix and pair primers that match the adapter sequences. We used a Solexa DNA library, which we generated with known concentration as a standard. Due to the extreme sensitivity, the DNA used in the reactions ranged from 0.0001–0.01 ng. The sequencing of the library DNA was performed on the Solexa/Illumina platform according to the manufacturer's instruction. Each library was analyzed in two lanes on the flow cell.
We used the Apr. 2006 assembly (dm3, BDGP Release 5) of the D. melanogaster genome, downloaded from http://hgdownload.cse.ucsc.edu/goldenPath/dm3/bigZips/chromFa.tar.gz, and the Nov. 2005 assembly (droYak2) of the D. yakuba genome, downloaded from http://hgdownload.cse.ucsc.edu/goldenPath/droYak2/bigZips/chromFa.tar.gz.
We trimmed all sequenced tags to 20 bp and mapped the tags to the genomes using Bowtie v0.9.9.1 [22] with command-line options ‘-v 1 -m 1’, thereby keeping only tags that mapped uniquely to the genome with at most one mismatch. Table 1 gives statistics on the total numbers of sequenced and mapped tags for all experiments. Note that while we mapped tags to the entire genomes, we did not use the heterochromatic chromosomes or unassembled sequence for any analyses.
We used annotations from FlyBase r5.15 [34] for analyses using genes in D. melanogaster.
We called peaks for each experiment using MACS v1.3.5 [25] with the option ‘--pvalue 0.00001’. We used total chromatin as background controls, and set the ‘--mfold’ option to the maximum value for which MACS could find a sufficient number of paired peaks. In order to only consider peaks for which we could reliably assign orthology and to control for potential assembly errors in the draft D. yakuba genome, we used exonerate [35] to search for peaks whose associated sequence was duplicated in either genome. For each peak, we (1) searched for duplicated sequence in the genome where the peak was called and (2) used the whole-genome alignment to pull out the orthologous sequence in the other genome and searched for duplicates of that sequence in the other genome, which frequently indicated a potential assembly error due to the unfinished nature of the D. yakuba assembly. We discarded any peaks whose associated sequence was duplicated in either genome.
We used a large-scale orthology mapping created by Mercator [23] to identify syntenic regions of the genomes, which were each aligned with FSA v1.11.0 with the options ‘--exonerate --softmasked --refinement -1 --mercator cons seqs.fasta’. The resulting whole-genome alignment can be downloaded here: http://www.biostat.wisc.edu/~cdewey/data/fsa_mercator_alignments/drosophila_melanogaster-5.0-drosophila_yakuba-2.0-1.0.tar.gz.
We first normalized the total number of sequenced tags to a fixed number for each experiment, the standard method of controlling for the variable success of amplification and sequencing. This normalization, however, is insufficient for our purposes, since it does not take into account differences in genome size and background between the species. We therefore performed an additional comparative normalization step. Assuming that the total amount of binding near known regulatory targets of the six factors studied here (A-P and D-V genes, as identified in [19] and listed below) is constant, we scaled the total number of sequenced tags in D. yakuba for each factor such that the total difference in inferred binding strength across the 50 most highly bound peaks in each genome (for a total of 100) within 10 kb of A-P targets was minimized (using a least-squares linear regression).
This comparative normalization procedure assumes there are no differences in the total number of molecules bound to A-P targets in the two genomes. Although this may not always be the case, we do not expect to see such global differences between such closely related species. It is also possible that by using the 50 most highly bound peaks near known A-P target genes for normalization we would underestimate variation in these genes. However, the effect of any single peak on the normalization was minimal, and the inferred divergence for any of these peaks did not change significantly when they were not included in the normalization (unpublished data).
We assessed binding strength by estimating a fragment density by extending each sequenced tag to the average fragment length based on the selected size distribution. We modified the SynPlot program [36] to display quantitative data along an alignment in order to create the plot in Figure 1.
We compared binding between the two genomes as follows: Given a peak called in one genome, we used the whole-genome alignment to project the 100 bp containing the peak onto the other genome and computed the maximum binding strength within that homologous sequence in the other genome. Note that therefore our maximum spatial resolution when assessing binding divergence is 50 bp, implying that if, for example, a binding site is present in D. melanogaster, and lost in D. yakuba but replaced by another site 30 bp away, then we will not detect any binding divergence if the two sites are bound at similar levels.
We labeled peaks that were within 10 Kb of a gene in D. melanogaster known to be regulated by A-P factors as A-P target loci. We used the following list of genes: Brk, D, Doc1, Doc2, E(spl), Kr, Phm, SoxN, Vnd, bowl, btd, cad, croc, dpp, ems, eve, fkh, ftz, gt, h, hb, hkb, ind, kni, knil, noc, nub, oc, odd, opa, os, pdm2, pnr, prd, pxb, rho, run, salm, shn, sim, slp1, slp2, sna, sob, sog, ths, tld, tll, tsh, tup, twi, vn, wntD, zen.
We identified DDWs for each factor as follows. For each word of a fixed length k, we identified all (non-softmasked) instances of the word (on both strands) within a 100 bp window centered on the empirical maximum of peaks called in D. melanogaster for that factor. We then accumulated two distributions of binding strength divergence (D. melanogaster − D. yakuba) for the word, pcons and pdiv, with pcons consisting of instances where the word was exactly conserved in D. yakuba and pdiv consisting of instances where the word was diverged in D. yakuba. We used a non-parametric statistical test, Kolmogorov-Smirnov test, to test for equality of distribution pcons ∼ pdiv. If equality of distribution could be rejected with p<0.01, then we called the word a candidate DDW. We then performed the identical procedure in the opposite direction, wherein we examined peaks called in D. yakuba and assessed the conservation of words in D. melanogaster, and identified a second set of candidate DDWs. We took the intersection of these two sets to obtain final lists of DDWs. We performed this procedure to identify words of length k = 6 and 7.
We assessed whether sequence motifs matched the known DNA-binding specificities of A-P factors with position weight matrices (PWM) from [29]. When creating Figures 4 and 7, we said that a word matched the specificity for a factor if it matched a subsequence of the corresponding PWM with ln (p value) <−4 as reported by Patser [37].
We used the Least Angle Regression (LARS) algorithm [28], implemented in the package lars for R [38], to learn a linear model of binding divergence using DDWs of length k = 6. We performed 5-fold cross-validation to estimate the mean-squared prediction error (MSE) associated with each value of the lasso regularization parameter β and then chose the model given by the β that yielded the lowest MSE. This cross-validation procedure helps to prevent the over-fitting characteristic of standard least-squares linear regression, making the correlations that we estimated robust to generalization error.
In order to ensure that (1) the DDWs that we identified truly have predictive value and (2) that the correlations reported are not due solely to base-composition effects, we randomly shuffled the nucleotides of each DDW to create a set of shuffled words with unchanged base composition, and then built a predictive model using these shuffled words. Models constructed using these shuffled words had no predictive value, indicating that the correlations that we report for our DDWs are not statistical artifacts. Figures S12–S27 show lasso variable selection curves and cross-validation curves for all values of β, as well as scatterplots of predicted and observed binding divergences, for predictive models constructed using our DDWs as well as their shuffled counterparts. The cross-validation curves make clear that while the DDWs are correlated with binding strength, the shuffled words are not: MSE decreases as more DDWs are included into the model, indicating the gain and loss of these words correlates with changes in observed binding strength, whereas MSE increases as more shuffled words are included into the model, indicating that these words are uncorrelated with binding. This provides clear evidence that our cross-validation procedure correctly chooses the model with the minimum generalization error, for example, that the models are not over-fit to the data.
We performed an identical analysis using words derived from the in vitro binding specificity data described in [29]. We enumerated all k-mers that matched a subsequence of the corresponding PWM with ln (p value) <−8 as reported by Patser [37], identifying four 6-mers for BCD, HB, and GT and sixteen 6-mers for KR, and then used the learning procedure described above to learn models of binding divergence using these words.
We calculated binding strengths of the six factors across all called peaks, subtracted the empirical means for each factor, and scaled the data for each factor such that it had unit variance. We used the singular value decomposition routine in IT++, a linear algebra library for C++, to perform PCA, and created heatmaps of the PCA results using a modified version of the aspectHeatmap function in the ClassDiscovery package.
In order to confirm that the putative chromatin signal represented by the first principal component did reflect coherent increases and decreases in binding of all six factors in our data, we randomly interchanged the measured binding strengths for a single factor across called peaks while holding all others unchanged (Figure S34, panels A–F) and similarly randomly interchanged the binding strengths of all factors (Figure S34, panel G), thereby removing spatial correlations between the binding of single factors and the other five (Figure S34, panels A–F) and removing spatial correlations between the binding of any factors (Figure S34, panel G). As expected, the chromatin signal disappeared after performing any of these transformations on the data.
We identified sequence motifs associated with interspecies divergence of each principal component using the same procedure described above, but with the data projected along the principal component of interest. For each principal component, we accumulated the distributions pcons and pdiv across all peaks called for any of the six factors.
All sequence reads from the experiments described are available from the NCBI's GEO database with accession number GSE20369. Processed datasets, including mapped reads, called regions and peaks, D. melanogaster − D. yakuba alignments, and all software described here, are available at http://rana.lbl.gov/data/melyak.
|
10.1371/journal.pbio.1001299 | ROP GTPase-Dependent Actin Microfilaments Promote PIN1 Polarization by Localized Inhibition of Clathrin-Dependent Endocytosis | Cell polarization via asymmetrical distribution of structures or molecules is essential for diverse cellular functions and development of organisms, but how polarity is developmentally controlled has been poorly understood. In plants, the asymmetrical distribution of the PIN-FORMED (PIN) proteins involved in the cellular efflux of the quintessential phytohormone auxin plays a central role in developmental patterning, morphogenesis, and differential growth. Recently we showed that auxin promotes cell interdigitation by activating the Rho family ROP GTPases in leaf epidermal pavement cells. Here we found that auxin activation of the ROP2 signaling pathway regulates the asymmetric distribution of PIN1 by inhibiting its endocytosis. ROP2 inhibits PIN1 endocytosis via the accumulation of cortical actin microfilaments induced by the ROP2 effector protein RIC4. Our findings suggest a link between the developmental auxin signal and polar PIN1 distribution via Rho-dependent cytoskeletal reorganization and reveal the conservation of a design principle for cell polarization that is based on Rho GTPase-mediated inhibition of endocytosis.
| Formation of cell polarity is a process of distributing cellular structures or molecules in an asymmetric manner. This process plays an important role in the generation of diverse cell forms and types. In plants, the quintessential hormone auxin is important for diverse physiological functions, including growth and development of cells and organs. To perform these functions, auxin must be transported and localized to specific regions within the plant. This is partially mediated by polar distribution of the PIN-FORMED (PIN) auxin efflux transporters, which transport auxin outside of the cell and allow for the directional short- and long-distance transport of auxin throughout plant tissues and organs. Although auxin itself has been implicated as a signal to regulate PIN polar distribution, how auxin does so remains to be elucidated. We previously showed that auxin promotes the generation of “puzzle-piece” polarity in leaf epidermal pavement cells, which contain interdigitated lobes and indentations, by activating the ROP (Rho-like GTPases from plants) members of the conserved Rho family of small GTPases. Here, we find that auxin-dependent local activation of ROP2 in the lobe region inhibits PIN1 internalization into the endosomal compartments (or endocytosis), leaving higher levels of PIN1 polar distribution in the lobe region. PIN1 internalization is inhibited by altering the actin cytoskeleton through the ROP2 effector protein RIC4, a protein involved in cytoskeletal remodeling. On the basis of our findings, we propose that the Rho GTPase-mediated inhibition of endocytosis of PIN1 provides a self-organizing mechanism for the polar PIN1 distribution. Rho GTPase-based inhibition of endocytosis is also important for the formation of cell polarity in animal cells. Thus, we conclude that Rho GTPase signaling to inhibit endocytosis is a common mechanism for cell polarization in multicellular organisms.
| Cell polarity is a conserved cellular property that is necessary for the generation of diverse forms and types of cells in both uni- and multicellular organisms [1],[2]. The general design principles that govern the formation of polarity and how they are used to generate diverse forms of polarity is a fundamental issue of developmental mechanisms. In the unicellular yeast, Rho family GTPase-mediated activation of endocytosis is required for cell polarization [3]–[5]. In contrast, emerging evidence suggests that Rho family GTPase-mediated inhibition of endocytosis is essential for the polarization of cells in some multicellular tissues as shown in cultured epithelial cells from rat [6] and neuroectodermal epithelial cells from Drosophila [7]. It is unclear whether Rho-mediated inhibition of endocytosis is a common design principle for polarity establishment in multicellular systems and how the inhibition of endocytosis is regulated.
In multicellular plants, coordinated polarization of the proposed auxin efflux carriers PIN-FORMED (PIN) proteins within a plant tissue is required for polar auxin transport and formation of auxin gradients, which regulate a wide range of morphogenetic and growth patterns in plants [8]–[11]. Asymmetric endocytosis and recycling of plasma membrane (PM)-localized PINs have been shown to contribute to the polar PIN localization [12],[13], and auxin has been implicated as a self-organizing signal to polarize PIN proteins through its inhibition of clathrin-dependent PIN endocytosis in root cells, which is mediated by the auxin-binding protein 1 (ABP1) putative cell surface auxin receptor [14],[15]. We studied auxin regulation of cell polarity formation and PIN1 polarization in Arabidopsis leaf epidermal pavement cells (PCs), which display multipolarity by forming the puzzle-piece appearance with interdigitated lobes and indentations [16]–[20]. Recently we showed that ABP1-dependent auxin signaling promotes the formation of multipolarity in PCs by activating Rho-like GTPases from plants (ROPs) that are associated with the plasma membrane [19],[21]. ROPs also regulate other processes mediated by auxin such as root hair development, lateral root formation, and root gravitropic responses [22]–[24]. In addition, auxin activation of ROPs is associated with auxin regulation of gene expression in the nucleus [25],[26].
We found that polar PIN1 localization to the tip of lobes in PCs is dependent upon ROP2, which is activated by auxin in the same PM region where PIN1 is localized [19]. PIN1 is required for ROP2 activation and lobe formation, supporting a role for auxin in self-organizing PIN1 polarization in PCs [19]. How auxin-activated ROP2 regulates PIN1 polarization is unknown. One possible mechanism would be the inhibition of PIN1 endocytosis by activated ROP2, because inactivation of ROP2 leads to PIN1 internalization in PCs [19]. This finding is consistent with the report showing that the expression of constitutively active ROPs inhibited internalization of the endocytosis tracer dye FM-64 in roots and guard cells [27]–[29]. ROP2 regulates the formation of the multipolarity in PCs by activating RIC4 [17], a member of the ROP INTERACTIVE CRIB MOTIF-CONTAINING proteins (RICs) family of ROP effector proteins [30]. RIC4 promotes the local accumulation of fine cortical actin microfilaments in the tip of PCs and pollen tubes [17],[31], and actin dynamics has been implicated in the regulation of auxin transport and PIN endocytosis [32]–[34]. These observations raise an interesting possibility that the ROP2-RIC4 pathway could regulate PIN1 polarization through endocytic trafficking and the actin cytoskeleton.
In this report we show that PIN1 endocytosis is preferentially inhibited in the PM region of lobes and that auxin activation of ROP2 in this region inhibits clathrin-dependent PIN1 endocytosis, allowing PIN1 to be polarized to the ROP2-active region. We further demonstrate that ROP2 promotion of F-actin accumulation via its effector protein RIC4 is responsible for its inhibition of PIN1 endocytosis. Our results reveal the conservation of a new design principle for cell polarization, which is based on localized inhibition of endocytosis by Rho GTPase signaling in multicellular plants and animals, and provide new insights into the mechanisms by which Rho GTPases inhibit clathrin-dependent endocytosis of polarity proteins. Our results establish an auxin signaling pathway leading to the polarization of PIN proteins that is essential for pattern formation and morphogenesis in multicellular plants.
To test the auxin-mediated self-organizing PIN1 polarization, we investigated how auxin-activated ROP2 signaling regulates PIN1 localization to the lobe tip. We first utilized PIN1-green fluorescent protein (GFP) transient expression in leaves of Nicotiana benthamiana (tobacco) plants by the agrobacterium infiltration method [35]. This system allows determining the effect of mutant ROP2 on PIN1-GFP localization independent of PC shape changes, which occur in Arabidopsis rop2 mutants [16],[17]. Within 3 d after infiltration, PIN1-GFP was detected in PCs of tobacco leaves and localized to the PM with stronger accumulation at the tips of lobes as in Arabidopsis PCs (Figure 1A, arrow). PIN1-GFP signal was also observed in the cytoplasm as endosome-like vesicles (Figure 1A, arrow). Time-lapse imaging showed that PIN1-GFP appeared to be internalized preferentially in the indentation region but not in the lobe region where stronger PM accumulation of PIN1-GFP was observed (Figures 1A, S1A, and S1B). Both PIN1-GFP and FM4-64 were internalized simultaneously and became colocalized in the same vesicles, confirming that GFP-PIN1 was internalized through endocytosis (Figure S1C).
Because PIN1 internalization appears to occur preferentially in the indentation region but not in the lobe region where ROP2 is activated [17],[19], we hypothesized that ROP2 activation may inhibit endocytosis of PIN1, allowing PIN1 to be polarized in that region. To visualize PIN1 internalization, we utilized PIN1 fused with the dendra2 photo-convertible fluorescent protein (Figures S2 and S3) [36]. Photo-conversion of PIN1-dendra2 transiently expressed in tobacco or Arabidopsis leaves was conducted using transient high dosage of irradiation with 405-nm laser (Figures S2A and S3A). To confirm whether PIN1-dendra2 expressed in leaves was internalized from the PM, PIN1-dendra2 cells were treated with Brefeldin A (BFA), which inhibits ADP ribosylation factor (ARF) GEF and arrests endosomal recycling, causing internalized PIN1 to accumulate in an aggregate known as BFA bodies in plant cells [14],[32]. PIN1-dendra2 at the PM was photo-converted from green to red emission. 30 min after photo-conversion, converted PIN1-dendra2 was observed in BFA bodies, which demonstrated the occurrence of PIN1-dendra2 endocytosis (Figure S2A). To test the effect of ROP2 on PIN1-dendra2 endocytosis, we coexpressed a dominant-negative mutant of ROP2 (DN-ROP2) with PIN1-dendra2 and observed the internalization of the photo-converted signal at the PM. In the lobe regions of PCs expressing PIN1-dendra2 only, PIN1-dendra2 vesicles were rarely formed from the PM (Figure 1B). In contrast, in cells expressing both PIN1-dendra2 and DN-ROP2, numerous PIN1-dendra2 vesicles were formed and pinched off from the PM (Figure 1B, arrowheads). Furthermore, time-lapse imaging showed that DN-ROP2 expression greatly accelerated the decrease in the photo-converted PM signal, which was quantified by changes in the relative intensity (Figure 1C) or in the absolute intensity (Figure S2B) of the converted signal. In cells coexpressing DN-ROP2 and PIN1-dendra2, the PM PIN1-dendra2 signal was generally weaker compared to cells expressing PIN1-dendra2 alone (Figure 1B). This finding was likely due to the DN-ROP2–mediated induction of endocytosis, but not its general toxic effect, because DN-ROP2 expression did not affect the expression and localization patterns of several endosomal markers (Figure S4). Thus, these results show that DN-ROP2 expression promoted PIN1-dendra2 internalization.
To confirm that the effect of DN-ROP2 on PIN1 endocytosis in tobacco cells reflected the function of ROP2 in Arabidopsis, we transiently expressed PIN1-dendra2 in the PCs of wild type (WT) or the rop4-1 rop2 RNAi line, in which ROP2 is down-regulated by RNAi and the functionally redundant ROP4 is knocked out (rop4 R2i) [17]. As expected, photo-converted signal was found in vesicles budding from the PM and decreased rapidly from PM in rop4 R2i cells but not in WT control cells (Figures 1D, 1E, and S3B). Moreover, expressing the constitutive active form of ROP2 (CA-ROP2) in rop4 R2i cells suppressed PIN1-dendra2 internalization (Figures 1E and S3B). These results indicate that ROP2/ROP4 suppresses PIN1 internalization, which supports our hypothesis that active ROP2 inhibits PIN1 endocytosis in the lobe region.
We next tested the identity of the PIN1 vesicles induced by DN-ROP2 expression by examining the colocalization with known endocytic markers in plants. Coexpression of DN-ROP2 with PIN1-GFP in tobacco leaves greatly increased the number of PIN1-GFP vesicles in the cytoplasm (Figure 2A and 2B), similar to the PIN1-dendra2 vesicles. Previous studies showed that endocytic trafficking mediated by the Rab5 family of GTPases plays an essential role in various developmental processes including PIN polarization [13],[37],[38]. Ara7, a Rab5 homolog, resides in an endosomal compartment from which various internalized proteins, such as PIN1, are sorted for targeting to vacuoles or recycling to the PM [39]. In cells coexpressing Venus-Ara7, PIN1-GFP, and DN-ROP2, most PIN1-GFP vesicles overlapped with Venus-Ara7 (Figure S5). Thus, most PIN1 vesicles induced by DN-ROP2 were localized to the endosomal compartment containing Ara7. Taken together our results suggest that activated ROP2 in the lobe region inhibits PIN1 endocytosis in that region.
Several types of endocytosis have been characterized in yeast or animals [40]. We speculated that the clathrin-dependent endocytic pathway contributed to the PIN1 internalization in PCs because this pathway has been reported to modulate the internalization of PIN proteins in other tissues [15],[41],[42]. To test this notion, we inhibited clathrin-dependent endocytosis by coexpressing the C-terminal region of AtAP180 protein (ANTH[C]) with PIN1-GFP. The conserved AP180 protein contains both the PIP2-binding domain and the clathrin-binding domain and is essential for the early stage of clathrin-dependent endocytosis [43],[44]. ANTH(C), which contains the clathrin-binding domain (ANTH domain), has a dominant-negative effect on the function of AP180 protein and inhibits the clathrin-mediated endocytosis [43]. Overexpression of ANTH(C) greatly reduced the number of PIN1-GFP–associated vesicles and suppressed DN-ROP2 induction of the PIN1-GFP vesicles (Figure 2A and 2B). ANTH(C) did not have a general toxic effect, because its expression did not affect the expression and localization of several other endosomal markers (Figure S4). Treatment with Ikarugamycin (Ika), a specific inhibitor of clathrin-dependent endocytosis [45], produced the same effect as ANTH(C) overexpression (Figure 2A and 2B). These results suggest that ROP2 activation suppressed clathrin-dependent endocytosis of PIN1.
Because we previously showed that auxin activates ROP2 in the regulation of PC shape formation [19], we next sought to test whether ROP2-mediated inhibition of endocytosis is also regulated by auxin. We first monitored the uptake of FM1-43 in the PCs of WT or rop4 R2i plants. BFA treatment for 2 h resulted in the accumulation of the FM dye in aggregated structures (BFA bodies) in WT cells (Figure S6A). Treatments with 5–10 µM auxin inhibits the internalization of FM dyes in root cells [14]. We found that application of naphthalene acetic acid (NAA) as low as 100 nM prevented the accumulation of FM1-43 in BFA bodies (Figure S6A). In rop4R2i PCs, FM1-43 accumulated in BFA compartments as in WT cells (Figure S6A). However, NAA did not prevent the accumulation of FM1-43 in these structures in rop4 R2i cells (Figure S6A). Furthermore, expression of CA-ROP2 suppressed FM1-43 accumulation in BFA bodies in PCs treated with BFA (Figure S6B). Thus, these results suggest that ROP2 is required for the auxin-induced inhibition of endocytosis.
Auxin-induced inhibition of PIN1 internalization has been well documented in roots [14],[15],[46]. We next asked whether PIN1 internalization in Arabidopsis PCs is also inhibited by auxin in a ROP2-dependent manner by transiently expressing PIN1-GFP in rop4 R2i cells (Figure 3) [19]. The BFA-induced PIN1-GFP structures were similar to the BFA compartments containing FM1-43 (Figures 3A and S6A). Treatments with NAA (100 nM) inhibited PIN1-GFP accumulation in the BFA compartments (Figure 3A and 3C). Thus, auxin suppresses PIN1 internalization in PCs as in other tissues. However, NAA treatments did not reverse PIN1-GFP accumulation to endosomal vesicles in rop4 R2i cells (Figure 3A and 3C). When CA-ROP2 was coexpressed with PIN1-GFP in rop4 R2i cells, the accumulation of PIN1-GFP vesicles was suppressed (Figure 3B and 3C). Therefore, ROP2/4 is required for the inhibitory effect of auxin on PIN1 endocytosis. In contrast to WT cells treated with BFA, PIN1-GFP remained in the endosomal vesicles in rop4 R2i cells upon BFA treatment (Figure 3A and 3C), implying that ROP2 may also regulate the PIN1 trafficking from or the transition of these endosomosal vesicles into recycling PIN1 vesicles, which BFA acts on.
Actin microfilaments have been implicated in the regulation of polar trafficking of PIN proteins [32],[34],[47],[48], but the exact nature of F-actin and the mechanism by which this F-actin modulates PIN polarization remains elusive. Because ROP2/ROP4 promotes the accumulation of fine cortical actin microfilaments through its downstream target protein RIC4 [17], we assessed whether RIC4-dependent F-actin mediates ROP-dependent PIN1 localization in PCs. We first analyzed the localization of PIN1 in ric4-1 mutants. Reduction of RIC4 level in ric4-1 mutants results in abnormalities in the PC shape that is less profound than but similar to those in the loss-of-function ROP2/ROP4 mutants [17]. In PCs of ric4-1, PIN1-GFP was internalized into endosomal vesicles (Figure 3A and 3C) as in rop4 R2i cells [19], and NAA treatment did not reverse PIN1-GFP accumulation in the endosomal vesicles in ric4-1 PCs (Figure 3A and 3C). Unlike rop4 R2i cells, however, coexpression of CA-ROP2 did not suppress the internalization of PIN1-GFP in ric4-1 cells (Figure 3A–3C). As shown for rop4 R2i cells, NAA treatments did not suppress FM dye accumulation in BFA compartments in ric4-1 PCs (Figure S6A). Taken together, these results suggest RIC4 acts downstream of ROP2/ROP4 in the suppression of PIN1 endocytosis in PCs.
Given a role for RIC4 in promoting the accumulation of cortical F-actin in the lobe region, we hypothesize that RIC4 inhibits PIN1 endocytosis through the RIC4-dependent F-actin. We tested this hypothesis by using a combination of F-actin–modifying chemicals and genetically modified Arabidopsis plants with both loss of and gain of RIC4 function. Stabilization of F-actin by treatments with chemicals such as TIBA or Jasplakinolide is reported to inhibit PIN endocytosis in roots of Arabidopsis [33]. Similarly, these chemicals stabilized cortical fine F-actin and inhibited endocytosis in WT Arabidopsis PCs, because treatment with TIBA or Jasplakinolide induced accumulation of cortical fine F-actin (Figure S7) and inhibited uptake of FM1-43 (Figure S8). Time-lapse imaging of photo-converted PIN1-dendra2 showed that loss of RIC4 function (ric4-1) greatly accelerated the internalization of photo-converted PIN1-dendra2 as expected (Figure 4A and 4B), whereas TIBA treatments completely reversed the acceleration of PIN1-dendra2 induced by the ric4-1 mutation or DN-ROP2 expression (Figures 4A, 4B, S9A, and S9B). In contrast, RIC4 overexpression suppressed FM1-43 internalization, as did CA-ROP2 expression (Figures 4C and S6B). In RIC4-overexpressing plants treated with the actin-depolymerizing drug Latrunculin B (100 nM), the accumulation of internalized vesicles was restored (Figure 4C and 4D; arrows). The same concentration of Latrunculin B greatly reduced the amount of the cortical fine F-actin, but not that of cytoplasmic actin cables (Figure S7). These results suggest that the accumulation of the cortical fine F-actin, which is activated by the ROP2-RIC4 pathway in the lobing region, inhibits the endocytosis of PIN1, and consequently promoting PIN1 polarization in the lobing region of the PM in PCs (Figure 4E).
Our findings here have established an auxin-activated ROP2-signaling pathway that regulates PIN1 protein polarization to the PC lobe through the localized inhibition of PIN protein endocytosis. Given the requirement of PIN1 for the ROP2 activation at the lobe region of the PM [19], this signaling pathway underscores a positive feedback loop leading to PIN1 polarization, which provides strong support for the hypothesis that auxin acts as a self-organizing signal in the control of PIN-dependent auxin efflux [14],[15]. Furthermore, we have demonstrated that auxin signaling links the Rho GTPase-dependent accumulation of the cortical fine F-actin to PIN polarization. This finding provides an important insight into the mechanism for the modulation of F-actin reorganization in its regulation of PIN endocytosis and polarization [32],[47],[49]. Several recent studies implicate actin dynamics in the regulation of PIN endocytic recycling. By using transgenic rice plants that express different levels of mouse talin protein, Nick et al. recently showed that dynamics of actin organization and auxin transport efficiency are coupled [34]. Auxin transport inhibitors such as TIBA were shown to induce bundling of actin filaments and inhibit endocytosis, and thus were suggested to affect auxin transport through actin-mediated vesicle trafficking of auxin transport-related proteins [13]. Our data show that the ROP2/RIC4-dependent auxin signaling pathway induces the accumulation of the cortical fine F-actin, which inhibits clathrin-dependent PIN1 endocytosis that leads to PIN1 polarization. The mechanism by which the ROP2-dependent F-actin accumulation inhibits endocytosis needs to be investigated in the future. In yeast, clathrin-dependent endocytosis requires not only Cdc42 GTPase-dependent polymerization of cortical actin patches but also their dynamics. Similarly in pollen tubes both ROP1 GTPase-dependent polymerization and dynamics of tip F-actin are critical for polarized pollen tube growth [50],[51]. Thus it will be interesting to know whether the polymerization of ROP2-mediated F-actin is also important for clathrin-dependent PIN1 endocytosis.
Importantly our findings show that Rho GTPase inhibition of endocytosis is a conserved design principle for the establishment of cell polarity in plants and animal cells. Rac and Cdc42 inhibition of endocytosis has also been shown to be required for cell polarization in cultured epithelial cells from rat [6]. Rho-GTPase mediates the developmental process of neuroectodermal epithelial cells in Drosophila, in which endocytosis of apical proteins are inhibited and their trafficking from early endosome to late endosome is promoted by CDC42 [7]. In plants, auxin inhibition of PIN endocytosis has been implicated in the regulation of PIN polarization that is required for auxin gradient formation and auxin flow and the formation of various developmental patterns [13]–[15],[52]. ROPs have been implicated in the regulation of similar developmental processes [22],[53]. ROP2 appears to regulate PIN2 polarization required for gravitropic responses [24]. ABP1 regulates auxin-induced inhibition of PIN1 endocytosis in roots [15], and acts upstream of ROP2 in the activation of the formation of the multipolarity of PCs in leaves [19]. It is reasonable to speculate that the ABP1/ROP2-based auxin signaling modulates PIN endocytosis in various developmental processes in plants. Thus, Rho GTPase regulation of PIN endocytic trafficking may provide a common mechanism for the regulation of PIN protein polarization.
Apart from the localized inhibition of PIN endocytosis, PIN polarization requires polar recycling of internalized PIN proteins [12],[54]–[56]. In addition to its activation of the RIC4-actin pathway that inhibits PIN1 endocytosis, ROP2 signaling may also promote polar recycling of PIN1. In support of this notion, we previously found that rop4 R2i PCs show stronger defects in cell shape formation and PIN1 distribution compared to ric4-1 PCs [17]. Interestingly, mutations in the ICR1 ROP effector protein induce strong defect in PIN polarization in Arabidopsis roots and embryos by affecting PIN recycling [55]. ICR1 is structurally unrelated to RICs and was shown not to affect PIN endocytosis [55],[57]. ICR1 interacts with the Arabidopsis homolog of SEC3, a component of the conserved exocyst complex that regulates the docking of exocytic vesicles to the PM site of exocytosis [28],[57]. Loss of ICR1 function also induces a strong defect in PC shape formation. Future work should determine whether ICR1 acts as a ROP2 effector to promote PIN1 recycling into the lobe region of the PM in PCs.
Our data suggest that the ROP2-RIC4-actin pathway participates in other aspects of endosomal trafficking in addition to its inhibition of PIN1 endocytosis. In this work, we found that defects in this pathway cause PIN1 to accumulate in an endosomal compartment containing Ara7 but not in BFA bodies. This finding implies that the ROP2-RIC4-actin pathway either is required for PIN1 trafficking to recycling vesicles or inhibits PIN1 trafficking to vacuolar compartments for degradation [54],[58]. In pollen tubes, the ROP1-RIC4-actin pathway regulates exocytosis required for tip growth [51]. It is possible that RIC4 also contributes to exocytic trafficking of PINs through actin-based targeting of recycling vesicles. Further studies will be needed to determine whether auxin activation of ROP signaling coordinates various downstream pathways leading to PIN polarization, such as the RIC4- and ICR1-dependent pathways.
Seeds of Arabidopsis or N. benthamiana (tobacco) were surface sterilized by 50% bleach with 0.1% triton X-100 and washed three times with distilled water. Arabidopsis plants were grown at 22°C on MS agar plates or in soil with 16-h light/8-h dark cycles. Tobacco plants were grown in soil with the same light cycles. The double-mutant ROP2RNAi rop4-1, ric4-1, and CA-ROP2 lines were described previously [16],[17],[19]. For chemical treatment, BFA, NAA, Ika, TIBA, JASP, and LatB were used from 50 mM, 100 µM, 5 mM, 50 mM, 2 mM, and 500 µM stock solutions dissolved in DMSO. DEX applications for induction of gene expression were done by spraying leaves with 3 µM DEX solution.
Plasmids used for balistics-mediated transient expression were constructed in pBI221. pBI221-CA-ROP2, pBI221-DN-ROP2, and pBI221-PIN1-GFP were described previously [16],[19]. Plasmids for DEX-inducible expression were constructed in derivative of pTA7002 [59], containing gateway cassettes kindly provided by Yuichiro Watanabe in the university of Tokyo. CDS of DN-ROP2, ARA7, or partial CDS of an AP180 protein (residue 991–1,959 of At1g05020) were introduced to pENTR or pDONR. LR reactions with obtained entry vectors and pTA7002-Venus-GW (ARA7), pTA7002-mCherry-GW (C-ANTH), or pTA7002-SECFP-GW (DN-ROP2) were performed to obtain pTA7002-Venus-ARA7, pTA7002-mCherry-C-ANTH, or pTA7002-SECFP-DN-ROP2. Entry vector for photo-convertible PIN1 (pDONR-PIN1-dendra2) was constructed by replacing GFP in PIN1-GFP amplified from pPIN1-PIN1-GFP [60] to dendra2 [36]. LR reactions with obtained entry vector and pGWB2 [61] or pBI221-sGFP-gateway were performed to obtain pGWB2-PIN1-dendra2 or pBI221-PIN1-dendra2.
Subcellular localization analysis in Arabidopsis PCs was done by ballistics-mediated transient expression as described previously [17]. We used 1 µg pBI221-PIN1-GFP or pBI221-PIN1-dendra2 and 0.5 µg pBI221-CA-ROP2 for particle bombardment. GFP signal was observed 24 h after bombardment by confocal microscopy (Leica SP2 confocal microscope or Zeiss 710 confocal microscopy). Conditions for imaging were set as 488-nm excitation, collecting bandwidth at 500–570 nm for GFP. For quantification of the number of PIN1-GFP vesicles per area, each cell area or vesicle size was measured using ImageJ.
Subcellular localization analysis in tobacco PCs was done by agrobacterium-mediated transient expression in leaf epidermal cells. Infiltration of agrobacterium for transient expression was performed as a standard protocol [62]. Leica SP2 or Zeiss 710 confocal microscopy was used for observation. Conditions for imaging were set as 488-nm excitation, collecting bandwidth at 495–515 nm for GFP, 514-nm excitation, collecting bandwidth at 560–640 nm for YFP, 442-nm excitation, collecting bandwidth at 450–490 nm for CFP, and 560-nm excitation, collecting bandwidth at 600–720 nm for mCherry. Any bleach-through signal among each channel was removed by adjusting the gain in the each channel using the signal in cells expressing single construct infiltrated at the same experiment.
For photo-converting PIN1-dendra2 expressed in tobacco PCs, regions of interest were illuminated by 405-nm laser at 5% power and speed set at 5 using Zeiss 710. For photo-converting PIN1-dendra2 expressed in Arabidopsis PCs, regions of interest were illuminated by 405-nm laser at 4% power and speed set at 7 using Zeiss 710. Conditions for imaging photo-converted signal were set as 560-nm excitation, collecting bandwidth at 600–720 nm. Quantification of PM signal at lobe region was performed by measuring intensity of PM along outermost cell outline in lobe sites using ImageJ.
FM1-43 dye uptake experiment was performed in liquid MS medium containing 10 µM FM1-43 using 2-d-old seedlings. Conditions for imaging were set as 488-nm excitation, collecting bandwidth at 500–570 nm.
|
10.1371/journal.pgen.1000076 | ATM Promotes the Obligate XY Crossover and both Crossover Control and Chromosome Axis Integrity on Autosomes | During meiosis in most sexually reproducing organisms, recombination forms crossovers between homologous maternal and paternal chromosomes and thereby promotes proper chromosome segregation at the first meiotic division. The number and distribution of crossovers are tightly controlled, but the factors that contribute to this control are poorly understood in most organisms, including mammals. Here we provide evidence that the ATM kinase or protein is essential for proper crossover formation in mouse spermatocytes. ATM deficiency causes multiple phenotypes in humans and mice, including gonadal atrophy. Mouse Atm−/− spermatocytes undergo apoptosis at mid-prophase of meiosis I, but Atm−/− meiotic phenotypes are partially rescued by Spo11 heterozygosity, such that ATM-deficient spermatocytes progress to meiotic metaphase I. Strikingly, Spo11+/−Atm−/− spermatocytes are defective in forming the obligate crossover on the sex chromosomes, even though the XY pair is usually incorporated in a sex body and is transcriptionally inactivated as in normal spermatocytes. The XY crossover defect correlates with the appearance of lagging chromosomes at metaphase I, which may trigger the extensive metaphase apoptosis that is observed in these cells. In addition, control of the number and distribution of crossovers on autosomes appears to be defective in the absence of ATM because there is an increase in the total number of MLH1 foci, which mark the sites of eventual crossover formation, and because interference between MLH1 foci is perturbed. The axes of autosomes exhibit structural defects that correlate with the positions of ongoing recombination. Together, these findings indicate that ATM plays a role in both crossover control and chromosome axis integrity and further suggests that ATM is important for coordinating these features of meiotic chromosome dynamics.
| Meiosis is the specialized cell division that gives rise to reproductive cells such as sperm and eggs. During meiosis in most organisms, genetic information is exchanged between homologous maternal and paternal chromosomes through the process of homologous recombination. This recombination forms connections between homologous chromosomes that allow them to segregate accurately when the meiotic cell divides. Recombination defects can result in reproductive cells with abnormal chromosome numbers, which are a major cause of developmental disorders and spontaneous abortions in humans. Meiotic recombination is tightly controlled such that each pair of chromosomes undergoes at least one crossover recombination event despite a low average number of crossovers per chromosome. Moreover, multiple crossovers on the same chromosome tend to be evenly and widely spaced. Mechanisms of this control are not well understood, but here we provide evidence that ATM protein is required for normal operation of this process(es) in male mice. ATM has long been known to be involved in cellular responses to DNA damage. Our studies reveal a new function for this protein and also provide new insight into the mechanisms by which meiotic cells ensure accurate transmission of genetic material from one generation to the next.
| Crossing-over between homologous chromosomes in conjunction with sister chromatid cohesion provides physical connections necessary for accurate chromosome segregation during the first meiotic division [1]. Due to their central role in meiosis, crossovers are tightly controlled in most organisms such that each chromosome pair gets at least one crossover, and multiple crossovers on the same chromosome tend to be evenly and widely spaced [2],[3]. One example of this control is the fact that non-exchange chromosomes are very rare even though the average number of crossovers per chromosome pair is low (often only 1–2 per pair). This observed tendency for at least one crossover to form per pair of homologous chromosomes is often referred to as the “obligate” crossover [3]. (The obligate crossover is viewed as one of the outcomes of the process(es) through which most crossovers form, not as a special type of crossover.) An especially striking example of this phenomenon is the sex chromosomes in males of many mammalian species, for which recombination between the X and Y is restricted to a relatively short region of homology, the pseudoautosomal region or PAR, which is ∼700 kb in some mouse strains [4]. Because a crossover must be formed to ensure segregation of the X and Y, the crossover rate per Mb of DNA is orders of magnitude higher in the PAR than in other regions of the genome.
A second manifestation of the regulation of crossing-over is interference, in which crossing-over in one genomic region makes it less likely that another crossover will be found nearby [2],[3],[5],[6]. A third manifestation is crossover homeostasis, documented in budding yeast as a tendency for crossover numbers to be maintained despite reduction in the number of recombination initiation events [7].
The number and distribution of crossovers are thus subject to multiple layers of regulation, which include both crossover-promoting (e.g., the obligate crossover and crossover homeostasis) and crossover-suppressing (e.g., interference) aspects. The term “crossover control” is often used as a catchall phrase to encompass these distinct aspects [3]. The various manifestations of crossover control may reflect a single underlying mechanism or closely interrelated set of mechanisms, although this remains to be experimentally verified ([2],[3],[5],[8] but see also [9]). The biochemical and genetic factors that govern crossover number and distribution are not well understood in most organisms, including in mammals.
Although key for chromosome segregation, crossing-over between homologous chromosomes is just one outcome of meiotic recombination, since noncrossovers also occur. Meiotic recombination initiates with DNA double-strand breaks (DSBs) introduced by the SPO11 transesterase [10]. DSBs are nucleolytically processed and the strand exchange proteins RAD51 and its meiotic homolog DMC1 act on the resulting single-stranded DNA ends to promote strand invasion into intact homologous DNA. Evidence from Saccharomyces cerevisiae, for which the mechanisms of meiotic recombination are best understood, suggests that the crossover versus noncrossover decision is made at or about this step during meiotic prophase [11]. In mouse spermatocytes, noncrossovers are estimated to outnumber crossovers approximately 10 to 1, as inferred from the ratio of RAD51 foci to MLH1 foci, which apparently mark sites of crossing-over [12],[13].
In male mice, several different molecular defects cause apoptosis of spermatocytes at the same point in meiotic prophase, equivalent to mid-pachynema in normal males [14],[15]. These defects include failure to initiate meiotic recombination (Spo11−/−) [16],[17] and failure to repair SPO11-generated DSBs (Dmc1−/−) [18],[19]. Despite the similar timing of apoptosis, spermatocytes from these mutants appear to arrest at different stages of meiotic progression, such that Spo11−/− spermatocytes express markers of early to mid-pachynema, whereas Dmc1−/− spermatocytes primarily express earlier markers (mid to late zygonema) [20]. Epistasis analysis with Spo11−/− revealed that the apparently earlier arrest in Dmc1−/− spermatocytes is a response to unrepaired DSBs [20]. Although the timing of apoptosis is quite different in females, oocytes also display distinct DNA damage- dependent and independent responses, such that Spo11−/− oocytes progress further than Dmc1−/− oocytes [21].
Loss of the serine/threonine kinase ATM also causes defects in meiotic progression during prophase I [22]–[24]. ATM activates cell cycle checkpoints in response to DSBs in somatic cells [25], and orthologs of ATM and the related kinase ATR also serve checkpoint monitoring functions for defects in meiotic interhomolog recombination in several organisms, including budding yeast and Drosophila (reviewed in [26]). However, phenotypes of Atm−/− spermatocytes and oocytes in mice are similar in many ways to those of Dmc1−/− meiocytes, and epistasis analysis with Spo11 mutation further reinforces this similarity [20],[21]. These findings strongly indicate that the loss of ATM impairs the repair of meiotic DSBs, suggesting that ATM plays a role in promoting meiotic recombination rather than only serving a monitoring function. This interpretation is consistent with other studies that demonstrate that ATM and/or ATR orthologs promote normal recombination patterns in unperturbed yeast and Drosophila meiosis [26]–[29], and also promote repair of DNA damage [25],[30] as well as basic chromosomal events [31] in non-meiotic mammalian and yeast cells. Precisely what meiotic processes are influenced by ATM in mammalian cells has been difficult to uncover, however, in part because progression through meiotic prophase I fails so catastrophically in Atm−/− mutants.
During our investigation of the epistatic relationship between Spo11 and Atm, we found that the testis cellularity of ATM-deficient mice was markedly increased by Spo11 heterozygosity, accompanied by significantly improved chromosome synapsis. A similar finding with a different Spo11 mutation was recently reported [32]. Spo11+/−Atm−/− spermatocytes can progress to meiotic metaphase I, although most cells undergo apoptosis at this stage. The rescue of meiotic progression to this stage allowed us to further explore the role of ATM in meiosis. Our analysis provides evidence for involvement of ATM in several aspects of crossover control and chromosome axis integrity.
Testis cellularity of ATM-deficient mice is markedly increased by Spo11 heterozygosity [32]. To characterize the increase, we performed a histological analysis of testis sections. Seminiferous tubules contain germ cells at various stages of spermatogenesis, with mitotic and early meiotic cells at the base of the tubule and later meiotic and post-meiotic stages displaced toward the lumen. Tubule cross sections can be classified into stages, referred to as I–XII, based on the particular set of germ cells present [33]. Spo11+/− testes show the normal pattern of these various stages (Figure 1A and data not shown), whereas tubules in Atm−/− mice are severely depleted of cells as a result of apoptosis of pachytene spermatocytes at stage IV [20],[23] (Figure 1B). In contrast, Spo11+/−Atm−/− mice presented morphologically normal pachytene cells in tubules at stage IV and beyond (Figure 1C and Figures S1A and S1B). Although some apoptosis at stage IV was still observed (data not shown), most Spo11+/−Atm−/− spermatocytes appeared to reach metaphase (stage XII tubules, Figure 1C). Round and elongating spermatids and sperm were also observed, although post-meiotic stages were severely reduced in number compared to wild-type mice, and in some cases appeared abnormal (Figures 1C and S1B; data not shown). Meiotic progression is dependent on Spo11 heterozygosity, as Spo11−/−Atm−/− mice undergo a stage IV apoptosis, like Atm−/− mice [20].
To further evaluate meiotic progression, testis sections were stained for phospho-histone H3 (p-H3), which is normally detected in spermatocytes from diplonema through the second division, as well as in dividing spermatogonia [34] (Figure 1D). Atm−/− spermatogonia were positive for p-H3 but spermatocytes were not, as expected because of apoptosis during prophase I (Figure 1E and data not shown). By contrast, p-H3-positive spermatocytes were observed in Spo11+/−Atm−/− mice, verifying progression to metaphase I (Figure 1F). These metaphase I cells of Spo11+/−Atm−/− mice often showed relatively darkly stained cytoplasm characteristic of apoptosis (Figure 1F and data not shown). TUNEL staining confirmed that most spermatocytes were eliminated at metaphase I by apoptosis [32] (Figure S1C). Thus, Spo11 heterozygosity sufficiently rescued defects associated with ATM loss to allow progression to metaphase I, but spermatogenesis was for the most part halted at this stage.
Loss of ATM also leads to germ cell depletion in females [22],[24]. We examined ovaries of Spo11+/−Atm−/− mice to determine if meiotic progression could also be rescued in oocytes. Ovaries were examined between 17 and 29 dpp, at a time when wild-type or Spo11+/− ovaries contain several thousand oocytes (Figure 1G). In Atm−/− females, only one oocyte was found in four mice examined (0.13 oocytes/ovary) (Figure 1H), whereas in four Spo11+/−Atm−/− females, there was a small but significant increase to 7.9±3.7 follicular oocytes/ovary (Figure 1I; mean±sd, p<0.0001, t test). Thus, Spo11 heterozygosity partially suppresses Atm−/− meiotic defects in both males and females, although to a different extent in females.
To further characterize the metaphase I defect of Spo11+/−Atm−/− spermatocytes, we examined meiotic spindles in testis sections (Figure 2A and 2B). Well-developed bipolar spindles were apparent in both Spo11+/− and Spo11+/−Atm−/− mice, with chromosome congression at the metaphase plate. However, one or two lagging chromosomes were often evident in the Spo11+/−Atm−/− mice (arrowheads, Figure 2B). Specifically, two of 11 Spo11+/− spindles (18%) showed a single lagging chromosome, whereas 11 of 13 Spo11+/−Atm−/− spindles (85%) showed lagging chromosomes (eight with one laggard, two with two laggards, one with three laggards) (p = 0.0031, Fisher's exact test). One possible explanation of these results is the frequent presence of achiasmate chromosomes (i.e., which have not undergone crossing over), because crossing-over is required for chromosome congression at metaphase I [35].
To determine if a particular chromosome pair was more likely to be achiasmate, we performed spectral karyotyping on meiotic chromosome spreads. In three Spo11+/− metaphases examined, each chromosome pair was present as a single unit (a bivalent) including the XY pair (Figure 2C). Autosomes also formed bivalents in five Spo11+/−Atm−/− metaphases examined, but the sex chromosomes were separated (formed univalents) in four of the cells (Figure 2D). To specifically examine the sex chromosomes in a larger number of metaphases, we performed FISH with probes for the X and Y. The sex chromosomes were always joined in Spo11+/− spermatocytes, as expected (n = 10; Figure 2E), but were univalents in 80% of Spo11+/−Atm−/− spermatocytes (Figure 2F) (n = 20; p<0.0001, Fisher's exact test). This behavior contrasted with autosomes: a single FISH signal was observed in both genotypes for Chromosome 10 (Chr10) (n = 15 for each genotype) (Figure 2G and 2H) and for Chr3 (data not shown). These results indicate that chiasma formation is not globally defective, but suggest instead that the XY pair is uniquely sensitive to defects caused by lack of ATM.
During meiotic prophase, homologous chromosomes are juxtaposed along their length via the synaptonemal complex (SC), which is fully assembled by pachynema. The SC comprises several proteins, including the axial element protein SYCP3, which assembles beginning in leptonema, and the central element protein SYCP1, which assembles along chromosome axes as homologous chromosomes synapse beginning in zygonema. Crossing-over is intimately associated with SC formation (reviewed in [36]). We therefore tested whether the XY crossover defect is accompanied by a defect in synapsis. As expected, the X and Y were always adjacent to each other in Spo11+/− pachytene spermatocytes (n = 38 cells) (Figure 3A), and the chromosome axes were closely juxtaposed at one end (Figure 3A insets), indicative of synapsis within the context of the SC. In contrast, the X and Y were far apart in 10.1% of Spo11+/−Atm−/− spermatocytes (n = 89 cells; Figure 3B). Furthermore, even though the X and Y were adjacent to one other in the remaining Spo11+/−Atm−/− cells, they were frequently not synapsed, as judged by the absence of intimate contact between their respective axes (Figure 3C inset). Separate immunostaining experiments with anti-SYCP1 and anti-SYCP3 confirmed that PAR synapsis occurred normally in Spo11+/− spermatocytes but was frequently defective in Spo11+/−Atm−/− cells (data not shown).
The sex chromosomes share homology only within the PAR, which is where the obligate XY crossover occurs (see Introduction). We used FISH to more precisely characterize PAR pairing and synapsis. In every Spo11+/− spermatocyte examined (n = 50 cells), a single, merged PAR signal that overlapped intimately juxtaposed axes was observed for the X and Y, consistent with synapsis in this region (Figure 3D). In contrast, three XY configurations were observed in Spo11+/−Atm−/− spermatocytes (Figure 3E and 3F; summary in Figure 3G). In 27% (n = 78 cells), there was a single PAR signal that overlapped intimately juxtaposed axes, consistent with XY synapsis (Figure 3E). However, in the major class (53.8%), PAR signals were separated even though the X and Y were adjacent (Figure 3F). Thus, even though the X and Y were usually juxtaposed, the PARs usually failed to synapse. The remaining 19.2% of spermatocytes had well separated X and Y (similar to Figure 3B; data not shown). These findings reveal that ATM is required for efficient pairing and/or synapsis of the sex chromosomes. As described further in Discussion, we consider it likely that the small size of the available region of homology within the PAR makes this genomic region uniquely sensitive to defects in these processes.
Importantly, the absence of ATM did not significantly reduce the total number of RAD51 foci in leptotene and zygotene spermatocytes in a Spo11+/− background. We observed 144.0±31.0 in Spo11+/− (14 cells) versus 123.5±78.1 in Spo11+/−Atm−/− in leptonema (25 cells) (mean±sd, p = 0.354, t test); and 173.8±23.8 in Spo11+/− (20 cells) versus 202.7±57.2 in Spo11+/−Atm−/− (23 cells) in zygonema (p = 0.041). This result suggests that the increased frequency of asynaptic and/or achiasmate sex chromosomes cannot be attributed simply to a reduction in overall DSB frequencies in Spo11+/−Atm−/− compared to Spo11+/−.
Sex chromosomes in spermatocytes are transcriptionally silenced during prophase through meiotic sex chromosome inactivation (MSCI), during which the X and Y are included in the sex body, a heterochromatin domain that excludes the active (phosphorylated) form of RNA polymerase II (reviewed in [37],[38]) (Figure 4A–4C). Phosphorylated RNA polymerase II was excluded from the sex chromatin of both Spo11+/− and Spo11+/−Atm−/− spermatocytes (Figure 4D–4F), indicating that ATM is dispensable for MSCI. Importantly, MSCI occurred even when the X and Y did not synapse within the PAR (Figure 4), consistent with recent studies suggesting that MSCI is driven at least in part by asynapsis per se [39].
The sex body is enriched for numerous proteins and protein posttranslational modifications [40]. One of these modifications is phosphorylation of the histone variant H2AX (γH2AX), which has also been implicated in MSCI [41]. ATM is dispensable for γH2AX formation in the sex body [32]. Using FISH, we confirmed that the sex chromosomes were included within a γH2AX-positive domain even when they were not synapsed (Figure S2A and S2B). When the sex chromosomes were separated, the X and Y were contained within separate γH2AX domains (Figure S2C and S2D). Similarly, neither ATM nor PAR synapsis were required for localization of two additional sex body components, NBS1 and TOPBP1 (Figure S2E and S2F). Thus, ATM is dispensable for formation of apparently bona fide sex bodies.
As described above, the XY pair frequently failed to generate a crossover in the absence of ATM, whereas crossing over on autosomes appeared grossly normal, at least insofar as ensuring formation of bivalents. This pattern could indicate that ATM is required specifically for recombination on the sex chromosomes, but the numerous structural defects on autosomes demonstrate that consequences of ATM loss are not confined to the sex chromosomes. We therefore considered the possibility that ATM deficiency alters crossing over more generally. To test this idea, we examined autosomal MLH1 foci, which localize to crossover-designated sites at pachynema [12],[13],[45] (Figure 6A). Autosomal MLH1 foci in Spo11+/−Atm−/−spermatocytes appeared grossly normal in that nearly all bivalents had at least one focus (Figure 6B), consistent with the metaphase I analysis indicating that ATM is not required for crossover formation per se. However, a close examination revealed several unusual characteristics consistent with a small but significant defect in crossover control on autosomes. These findings are in general accord with the recent demonstration of crossover control defects associated with mutations of Mre11 and Nbs1 that attenuate ATM signaling in mouse [46].
In the absence of ATM, mouse spermatocytes and oocytes die by apoptosis during prophase of meiosis I, exhibiting profound defects in meiotic chromosome behavior [22],[24]. Remarkably, most of the spermatocyte defects are eliminated simply by halving Spo11 gene dosage: instead of dying by apoptosis at pachynema (like Atm−/− and Spo11−/− single mutants), most Spo11+/−Atm−/− spermatocytes progress to metaphase I, and sometimes beyond [32]. Homologous synapsis, sex body formation, and crossing over are substantially, albeit incompletely, rescued. In this study, we took advantage of this intriguing phenotype to analyze the role of ATM in meiotic recombination. As discussed further below, our findings suggest previously undefined roles of ATM in crossover control and in promoting integrity of higher order chromosome structures.
How does Spo11 heterozygosity rescue Atm−/− meiotic progression? Cytological and other evidence suggest that Spo11+/− spermatocytes form fewer DSBs than wild type (F. Cole, S. Keeney, and M. Jasin, unpublished observations). Given that ATM has an established role in meiotic DSB repair [20],[21],[32], a straightforward interpretation is that a reduced number of SPO11-generated DSBs is responsible for suppression of the meiotic DSB repair defects arising from ATM deficiency. Perhaps there is a threshold amount of DSBs below which another kinase (e.g., ATR) can partially substitute for ATM; above this threshold, the number of DSBs may exceed the capacity for this kinase to substitute. It is also possible that DSBs are formed in wild-type numbers in Spo11+/− mice but are delayed such that induction of ATR later in prophase is able to substitute for ATM [32]. Current findings do not allow us to distinguish between these and other possibilities.
Atm−/− spermatocytes, similar to many other mutants including Spo11−/−, Dmc1−/− and Msh5−/−, show pronounced defects in forming a bona fide sex body and transcriptionally silencing the X and Y chromosomes [20],[37]. Studies of these and other mouse mutants defective for MSCI strongly support the hypothesis that failure to silence the sex chromosomes is sufficient to trigger apoptosis of pachytene spermatocytes in stage IV tubules (reviewed in [37],[38]). Thus, the substantial restoration of sex body formation and MSCI in Spo11+/−Atm−/− spermatocytes may account for the suppression of Atm−/− pachytene apoptosis.
Although Spo11+/−Atm−/− spermatocytes progress further than Atm−/− single mutants, they are substantially eliminated at or prior to the first meiotic division. It is possible that apoptosis is triggered by a spindle checkpoint responding to the lagging chromosomes observed at metaphase I, which are likely to be the frequently achiasmate X and Y. Indeed, metaphase I spermatocyte apoptosis has been observed in several instances where one or a few achiasmate chromosomes are present because of chromosomal abnormalities [53]–[55], as well as in Mlh1−/− mice in which most chromosomes lack chiasmata [56]. Alternatively, or in addition, metaphase I apoptosis of Spo11+/−Atm−/− spermatocytes may be a response to unrepaired DSBs, whose presence is indicated by persistent γH2AX, RAD51, and RPA (data not shown).
The rescue of meiotic progression in Spo11+/−Atm−/− females was much less pronounced than in males. We have shown that ovarian follicle formation is particularly sensitive to the presence of unrepaired DNA damage [21]. Thus, even if meiotic prophase events were rescued to the same extent as in spermatocytes, it is possible that persistent DNA damage would preclude rescue of oocytes at this stage.
Spo11+/−Atm−/− spermatocytes exhibited numerous defects in chromosome axes. It is possible that the structural flaws reflect defects in axis morphogenesis, but as discussed below there is also reason to consider that the lack of ATM causes defects in axis stability. Previous studies noted chromosome fragmentation in Atm−/− spermatocytes but were unable to distinguish whether this defect was an indirect effect of arrest and apoptosis in early to mid prophase [22]. Since progression through meiotic prophase I is substantially rescued in Spo11+/−Atm−/− spermatocytes, our results indicate that axis defects are more directly tied to the lack of ATM.
The crossover and chromosome axis defects in Spo11+/−Atm−/− spermatocytes may be separate. However, considerations about the relationship between meiotic recombination and higher order chromosome structures lead us to speculate instead that these defects may be manifestations of a single underlying problem. In many organisms, mutations affecting chromosome structure proteins perturb meiotic recombination and, conversely, mutations affecting recombination factors perturb chromosome structures [reviewed in 36,49]. Moreover, cytological and molecular studies reveal that meiotic recombination occurs in close spatial coordination with chromosome axes (reviewed in [49]). Taken together, these observations reveal functional connections between recombination and axes. It has been argued that these connections are important for establishing a functional chiasma, because a chiasma is more than just a crossover at the DNA level—a chiasma also involves higher order chromosome structure changes, including exchange of the chromosome axes and local separation of sister chromatids [5],[49],[59].
In order for chromosome structures and recombination events to develop in parallel, signals coordinating these processes must be transduced in both directions between the axes and the recombination machinery. Moreover, chromosome axes are likely to participate directly in crossover control by providing a conduit for an interference signal that governs distribution of crossovers [5],[60]. We propose that ATM kinase activity generates or transduces one or more of these signals. Consistent with this interpretation, mutations of Mre11 and Nbs1 that attenuate ATM signaling also cause crossover control defects in mouse spermatocytes [46]. Relevant phosphorylation targets remain to be identified, but might include histones, structural components of the axes, and/or recombination proteins (see also [27],[58]). Non-catalytic (i.e., kinase-independent) functions of ATM are also possible [61].
This model suggests how axis and recombination perturbations could both arise from absence of ATM. Sites of ongoing recombination are also places where axes are locally destabilized, for example showing buckling or twisting of the axes (reviewed in [49],[62]). If Atm−/− mutants are defective for interactions between recombinosomes and the axes (e.g., if ATR is only partly effective as a substitute), then correlated defects would be expected in all of the processes that depend on these interactions. If correct, this model predicts that axial interruptions in Spo11+/−Atm−/− spermatocytes occur specifically at sites where DSBs have occurred. The observed correlation between chromosomal anomalies and persistent γH2AX, RAD51, and RPA foci at pachynema in these mice is consistent with this prediction. Moreover, we found that axis defects that result in overt chromosome fragmentation in the absence of ATM are spatially correlated with chromosomal regions where crossover control is known to play an important role—the short SC fragments in Spo11+/−Atm−/− spermatocytes were usually derived from the distal tips of chromosomes, and there is a known preference in spermatocytes for one (or the only) crossover on a bivalent to be located distally [12],[51]. This nonrandom positioning is thought to be another manifestation of crossover control [51],[63]. Thus, the position of fragmentation is consistent with our hypothesis that axis and crossover control defects are functionally connected.
The meiotic cell's ability to coordinate multiple molecular processes spanning size scales that differ by orders of magnitude is truly remarkable. The unexpected rescue by Spo11 hemizygosity of meiotic prophase progression in Atm−/− spermatocytes has allowed us to identify ATM as a prime candidate to be directly involved in this unique feature of meiotic chromosome dynamics.
Spo11−/− and Atm−/− mice were as previously described [16],[22] on a C57Bl/6×129/Sv mixed background. To minimize variability from strain background, experimental animals were compared to controls from the same litter or from the same matings involving closely related parents. Each analysis was done with 2–4 Spo11+/−Atm−/− experimental animals (except for TOPBP1 staining, Figure S2). In each case, experimental animals were matched with 2–4 Spo11+/− controls, except for RAD51 focus counts (text); phospho-Pol II staining (Figure 4A), STAG3 staining (Figure 5D), evaluation of chromosome continuity by combined FISH/immunofluorescence (Figure 5E), and TOPBP1 staining (Figure S2). Importantly, all of the patterns described above for XY synapsis/chiasma defects, autosomal MLH1 numbers and distributions, and chromosome axis defects were reproducibly observed in multiple sib-pair comparisons. No significant variations were observed in between-individual or between-litter comparisons of animals with the same Spo11/Atm genotype. Genotyping was performed by PCR of tail tip DNA as previously described [21]. Experiments conformed to relevant regulatory standards and were approved by the MSKCC Institutional Animal Care and Use Committee.
Testis cell preparations were prepared for surface spreading and sectioning as described [20] from 2–4 month-old mice unless otherwise stated. Indirect immunofluorescence analysis of spread chromosomes was performed using described methods and antibodies [20]. Additional primary antibodies were rabbit anti-MLH1 (Calbiochem PC56T), 1∶75 dilution; rabbit anti-RAD51 (Oncogene), 1∶250; rabbit anti-TRF1 (generously provided by T. de Lange, Rockefeller Univ.), 1∶200; guinea pig anti-STAG3 (generous gift of C. Höög, Karolinska Institute), 1∶30; and CREST serum to detect centromeres (generous gift of P. Moens, York University), 1∶500. For RAD51 and MLH1 focus counts, nuclei were staged according to the extent of SYCP3 staining and synapsis. Leptonema was defined as having short, unsynapsed SYCP3 fragments. Early zygonema was defined as <50% synapsis and late zygonema was defined as >50% but less than 100% synapsis. Only nuclei with at least 19 autosomal MLH1 foci were considered for MLH1 counts, and only RAD51 and MLH1 foci that co-localized with SYCP3 staining were counted.
Detailed methods for testis sectioning and immunohistochemistry are described elsewhere [20]. Briefly, for histological analysis, sections were stained by periodic acid-Schiff (PAS) and hematoxylin. Spermatogenic staging of PAS-stained seminiferous tubule sections was as described [33]. For immunohistochemistry, anti-phospho-histone H3 antibody (Upstate Cell Signaling) was used at 5 μg/ml and detected with HRP-conjugated secondary antibodies using DAB as a substrate; slides were counter-stained with hematoxylin.
For analysis of meiotic spindles, 30-μm testis sections were placed on poly-L-lysine (Sigma) coated slides, and dried 2 hr, followed by a 2 hr incubation at 37°C. Slides were post-fixed 10 min in cold methanol, washed twice with PBS, then blocked for 20 min with antibody dilution buffer [20], and washed three times with PBS before incubation overnight at 4°C with anti-β-tubulin antibody (Sigma T4026) at 1∶200 dilution. Coverslips were mounted using ProLong Gold (Molecular Probes) containing DAPI. Images were analyzed using a confocal imaging system (Zeiss).
Combined immunofluorescence/FISH was performed as described [20] using FITC-conjugated X chromosome paint (Cambio, UK) and coumarin-conjugated (ENZO) Y-specific repetitive BAC probe Ct7-590p11 (Invitrogen). SpectrumGreen-conjugated (Vysis, Abbott Labs) PAR-specific probe was prepared from mouse BAC RP24-500I4 (CHORI). Paints for chromosomes 10 and 3 were from Cambio. SKY was performed as described [64]. Metaphase cells were documented with a Nikon E800, and images were analyzed using the SKYview 2.1.1 software.
Autosomal SC lengths and MLH1 focus positions were recorded using MicroMeasure, version 3.01 (http://www.colostate.edu/Depts/Biology/MicroMeasure). MLH1 position was measured from the centromeric end of the chromosome as revealed by the brighter DAPI staining of pericentromeric heterochromatin. Once SC length and the position of each MLH1 focus were obtained, the SCs in each spread were rank-ordered based on their absolute length, from rank 1 (longest) to rank 19 (shortest). Similarly ranked SCs were grouped to allow comparison of similar chromosomes between cells and genotypes. Based in part on published analyses of autosomal SC sizes using combined FISH and immunofluorescence [52], we chose the following groupings: ranks 1–2 (i.e., the two longest SCs in each spread), ranks 3–5, ranks 6–11, ranks 12–16, and ranks 17–19. For SCs containing gaps, the length of the gap was subtracted from the total length of the bivalent. Whenever possible for SC fragments, the lengths of fragments that appeared to originate from the same bivalent were combined to estimate the complete SC length for ranking purposes. Best fits of frequency distributions of MLH1 inter-focus distances to the gamma distribution were calculated using the GenStat software package (VSN International Ltd, Hemel Hempstead, UK), as described [51]. Correction was applied as described [51] to adjust ν values for the limited number of interfocus distances that can be measured (see Table 2). Other statistical tests were as specified in the text. We applied the non-parametric Mann-Whitney U Test to total numbers of MLH1 foci per cell to avoid the need to assume that the data were normally distributed. However, similar conclusions as to statistical significance were drawn if a t test was used instead (data not shown). Two-by-two contingency tables were subjected to two-tailed Fisher's exact tests. For larger contingency tables, a log-likelihood test for heterogeneity (G test) was applied. |
10.1371/journal.ppat.1002374 | CD11b+, Ly6G+ Cells Produce Type I Interferon and Exhibit Tissue Protective Properties Following Peripheral Virus Infection | The goal of the innate immune system is containment of a pathogen at the site of infection prior to the initiation of an effective adaptive immune response. However, effector mechanisms must be kept in check to combat the pathogen while simultaneously limiting undesirable destruction of tissue resulting from these actions. Here we demonstrate that innate immune effector cells contain a peripheral poxvirus infection, preventing systemic spread of the virus. These innate immune effector cells are comprised primarily of CD11b+Ly6C+Ly6G- monocytes that accumulate initially at the site of infection, and are then supplemented and eventually replaced by CD11b+Ly6C+Ly6G+ cells. The phenotype of the CD11b+Ly6C+Ly6G+ cells resembles neutrophils, but the infiltration of neutrophils typically occurs prior to, rather than following, accumulation of monocytes. Indeed, it appears that the CD11b+Ly6C+Ly6G+ cells that infiltrated the site of VACV infection in the ear are phenotypically distinct from the classical description of both neutrophils and monocyte/macrophages. We found that CD11b+Ly6C+Ly6G+ cells produce Type I interferons and large quantities of reactive oxygen species. We also observed that depletion of Ly6G+ cells results in a dramatic increase in tissue damage at the site of infection. Tissue damage is also increased in the absence of reactive oxygen species, although reactive oxygen species are typically thought to be damaging to tissue rather than protective. These data indicate the existence of a specialized population of CD11b+Ly6C+Ly6G+ cells that infiltrates a site of virus infection late and protects the infected tissue from immune-mediated damage via production of reactive oxygen species. Regulation of the action of this population of cells may provide an intervention to prevent innate immune-mediated tissue destruction.
| During a natural virus infection, small doses of infectious virus are deposited at a peripheral infection site, and then a “race” ensues, in which the replicating virus attempts to “outpace” the responding immune system of the host. In the early phases of infection, the innate immune system must contain the infection prior to the development of an effective adaptive response. Here we have characterized the cells of the innate immune system that move to a site of peripheral virus infection, and we find that a subset of these cells display atypical expression of cell surface molecules, timing of infiltration, and function. These cells protect the infected tissue from damage by producing reactive oxygen molecules, which are widely accepted to increase tissue damage. Therefore our findings indicate that during a peripheral virus infection, the typical rules governing the function of the innate immune system are altered to prevent tissue damage.
| Typically, the acute innate immune response to a peripheral challenge involves rapid infiltration of Ly6C+Ly6G+ neutrophils, followed by Ly6C+Ly6G- monocytes, in a process that involves chemoattraction mediated by arachidonic acid metabolites, cytokines, and chemokines [1]. Both neutrophils and monocytes mediate inflammation, but monocytes are also thought to play a major role in clearance of apoptotic neutrophils and restoration of tissue homeostasis [2], [3]. Neutrophils and monocytes are not, however, homogeneous populations of cells, and subtypes of these cells have been described based on their expression of surface markers or production of cytokines. A full understanding of the phenotype and function of each of these cell populations is required in order to understand (and manipulate) the mechanisms that clear pathogens, prevent systemic spread, and prevent or reduce immune–mediated tissue damage at the site of infection.
The majority of studies investigating the role of innate immune effector cells have been conducted using either sterile inflammation models or bacterial infections. Here we have examined the role of innate immune effector cells in protection against peripheral infection with virus. Many investigations studying antiviral immunity have utilized systemic routes of infection (intraperitoneal or intravenous) or examined infections in the respiratory tract. However, numerous viral infections are transmitted through breaks in the skin, and the dermal route of inoculation is favored for delivery of viral vaccine vectors [4], [5], [6]. Following infection of the skin with a pathogenic virus, replication occurs locally unless controlled by the innate immune system, and subsequently the virus spreads systemically to cause disease. After intradermal infection with vaccinia virus (VACV), a natural peripheral route of infection [7], the immune system prevents systemic spread of the virus [8]. A large number of the infiltrating cells at the site of infection are F4/80+, likely representing monocytes/macrophages [9], [10]. Although CD4+ T cells and antibodies have been implicated in the control of VACV infection following systemic challenge [11], the cells responsible for preventing systemic spread of VACV following an intradermal infection have not been identified.
Several recent studies have described important roles for monocytic cells in the immune responses to various intracellular pathogens [12], [13], including viruses [14], [15], [16]. In respiratory infection, depletion of alveolar macrophages enhances the spread of VACV to peripheral sites such as the ovaries, indicating that these cells may play an important role in anti-VACV immunity [17]. Following systemic infection with VACV, TLR2-mediated recognition of uncharacterized viral components causes both IL-6 production [18] and Type I interferon production by Ly6C+CD11b+ cells [19], resulting in a reduction in virus titers. However, the role of these innate immune effector cells and molecules in control of virus spread from a natural peripheral site of infection as well as their role in tissue regeneration following infection, has not been addressed. In addition, the role of Ly6C+Ly6G+ cells in protective immunity and tissue protective responses following VACV infection is unknown.
Here we describe the role of Ly6C+Ly6G- monocytes in preventing systemic spread of virus from a dermal site of infection without a requirement for T cell infiltration. In addition, we describe the role of Ly6C+Ly6G+ cells that infiltrate the site of infection subsequent to the accumulation of monocytes. These Ly6C+Ly6G+ cells produced Type I interferon and mediated tissue repair via the production of reactive oxygen species. These findings demonstrate the complexity of the cellular innate response to peripheral virus infections, and how the response differs from innate responses to a sterile inflammatory stimulus. Thus, our results demonstrate the plasticity of innate immune effector compartments, describe a previously unknown role for a subset of Ly6C+Ly6G+ cells, and show that reactive oxygen species (ROS) production by these cells allows the resolution, rather than the exacerbation, of tissue damage during an acute infection.
To gain insight into the mechanisms deployed by the immune system to prevent the systemic spread of virus following a peripheral infection, we infected mice in the ear with VACV and monitored both virus replication and the infiltration of populations of immune effector cells to the ear at various times post infection. Virus replicated exponentially in the ear pinnae until day 5, at which point the titer began to plateau (Fig. 1A). Virus titers then dropped until virus was finally cleared when the scab that formed at the site of infection fell off between day 12 and 15 post infection. At day 5 post infection, the time point at which virus replication plateaus, no significant accumulation of αβ TCR T cells was observed at the site of infection (Fig. 1B). However, the plateau of virus replication coincided with the peak in numbers of CD11b+ cells (Fig. 1B). The CD11b+ population contained a small but reproducible number of CD11clo cells that expressed CD11b, but not B220 (CD45RA), and likely represent monocyte-derived DC (Fig. S1). We did not observe infiltration of CD11c+ B220+ plasmacytoid DC, proposed to be the primary producer of Type I IFN, to the site of infection up to day 11 post infection (data not shown).
To determine the role of the infiltrating CD11b+ cells, we intravenously injected clodronate liposomes to induce apoptosis of the phagocytic cells that internalize the liposomes [20]. We typically observed a 70–80% depletion of CD11b+ cells at the site of infection throughout the time course of the experiment upon repeated injection of clodronate liposomes (data not shown). Liposome injection resulted in a minor increase in VACV titers in the ear pinnae of the phagocyte-depleted mice on days 5–9 as compared to control mice (Fig. 1C). VACV replication is completely confined to the ear pinnae following intradermal infection with a dose of 104 pfu [8]. The confinement of virus replication to the ear pinnae contrasts with systemic routes of infection, such as intraperitoneally or intravenously, after which virus can be found in multiple organs and tissues, including the ovaries, the primary site of VACV replication in a female mouse. Because the spread of VACV to the ovaries is accelerated following the depletion of alveolar macrophages after an intranasal challenge [17], we examined the role of phagocytes in the confinement of viral replication to the ear pinnae. The presence of replicative VACV in the ovaries was analyzed using a plaque assay following depletion with clodronate liposomes. Replicative VACV was observed in the ovaries of mice undergoing clodronate liposome treatment, but not in control mice (Fig. 1D). VACV was detected in the ovaries beginning on day 5 and peaked on day 7 at a level 1000-fold in excess of the original inoculum. Notably, depletion of T cells systemically and in the ear using an anti-Thy1 antibody did not increase virus titers in the ear through day 9 post infection (Fig. S2) or allow detection of VACV in the ovaries (data not shown). Therefore phagocytes, but not T cells, play a vital role in controlling virus replication at the site of infection and in preventing the systemic spread of VACV following a peripheral infection.
CD11b is a broadly expressed integrin subunit that is found on neutrophils, monocytes, macrophages and some DC subsets. In an attempt to specifically address the role of monocyte/macrophages in control of virus replication at the site of infection and the spread of VACV systemically, we utilized Macrophage Fas-Induced Apoptosis (MAFIA) mice which encode both a suicide gene and Green Fluorescent Protein (GFP) driven by the c-fms (CD115/MCSF receptor) promoter [21]. Injection of the drug AP20187 leads to dimerization of the suicide protein, activation of the Fas pathway, and subsequent apoptosis of cells expressing CD115. Previous publications have shown that monocyte/macrophages in MAFIA mice are GFP+ and are depleted upon AP20187 treatment, whereas neutrophils express very little GFP and are unaffected by AP20187 treatment in these mice [21]. In our studies, repeated injections of AP20187 produced >80% depletion of CD11b+ cells at the site of infection (data not shown). Similar to clodronate liposome treatment, AP20187 treatment of MAFIA mice produced a minor but reproducible increase in VACV titers in the ear of infected mice (Fig. 1E). AP20187 administration to MAFIA mice also allowed systemic spread of virus in the majority of mice (Fig. 1F), but the VACV titers found in the ovaries of treated MAFIA mice were 105-fold lower than titers seen following clodronate liposome treatment (Fig. 1D). These data indicated that the primary cell type required for preventing systemic spread of VACV is a subset of phagocytic cells that was depleted effectively in mice treated with clodronate liposomes, but less effectively depleted in MAFIA mice treated with AP20187.
VACV infection in the ear causes significant pathology that can be quantified by measurement of swelling, lesion size and tissue damage [8]. The pathology we observed following VACV infection differed slightly from published reports, a finding that is likely caused by differing sources of laboratory animals and different housing conditions between institutions. When we depleted mice of phagocytes using either clodronate or AP20187 treatment, the pathology at the site of VACV infection was dramatically increased at late times. To quantify this pathology, we measured the lesion size and tissue damage, which represents the loss of necrotic tissue from the ear [8], in infected mice that were untreated or depleted using clodronate liposomes or AP20187 . Both lesion size and tissue loss was significantly greater in mice depleted of phagocytes using either methodology (Fig. 1G, H). Notably, both the increase in lesion size and the tissue loss in phagocyte-depleted mice occurred primarily at time points after the minor increase in virus replication had been controlled. These data indicate a dual role for phagocytes during VACV infection, namely blockade of systemic spread and reduction of host-mediated tissue pathology.
In order to investigate which cell population was required to prevent spread of VACV from the site of infection, we needed to further characterize the phenotypes of cells within the infiltrating CD11b+ population at the site of infection. We examined infiltration of CD11b+ cells to the ear of MAFIA mice at various times post infection, taking care to ensure that the cells that we observed were not doublets (which could confound their identity) as outlined in our gating strategy (Fig. S3). Virtually all of the CD11b+ cells infiltrating the site of infection were GFP+ in MAFIA mice (Fig. 2A), until day 7 post infection when a CD11b+ GFP- population began to accumulate [9]. Thus, either all CD11b+ cells infiltrating the site of infection were of monocytic origin or the use of a VACV infection model led to expression of GFP within cells in MAFIA mice that were not of monocytic origin. Therefore it was necessary to distinguish monocyte/macrophages from neutrophils, which comprise another likely major population of infiltrating phagocytes.
There are few phenotypic markers that can distinguish neutrophil and monocytes/macrophage populations. One of these markers is CD115, but we were unable to detect infiltrating CD11b+GFP+CD115+ cells at the site of VACV infection (data not shown) although similar digestion protocols in different tissues in uninfected mice did produce CD115 staining. As surrogate GFP expression driven by the CD115 promoter did not distinguish between these cells types (Fig. 2A), we examined expression of cell surface Ly6C and Ly6G, as published work indicates that Ly6C is expressed by monocytes [22], [23], whereas both Ly6C and Ly6G are expressed by granulocytes [24], [25]. Using antibodies specific for Ly6C and Ly6G to analyze the CD11b+ cell infiltrate by flow cytometry, we observed several different cell populations accumulating in the ear over the course of the infection (Fig. 2B). The accumulation of immune cell populations, including Ly6C+Ly6G+ neutrophils, did not occur until day 2–3 post infection (Fig. 2A–B).
A population of Ly6C+Ly6G- monocytes was detected in the ear at numbers above uninfected tissue on day 3 post infection and represented the predominant CD11b+ cell at the site of infection through day 7 (Fig. 2B). Small numbers of Ly6C+Ly6G+ cells could be identified in the ear at early times after infection, but the numbers of these cells increased significantly after 5 days of infection and became the major population beyond day 7. Typically, this Ly6C+Ly6G+ population would be classified as neutrophils. However, the timing of infiltration of these Ly6C+Ly6G+ cells is not consistent with the infiltration of neutrophils that infiltrate a site of insult early, before monocytes, and die rapidly unless replaced [2]. Similar populations of infiltrating Ly6C+Ly6G+ cells were observed at the site of infection 5 days after dermal infection with other viruses (Fig. S4), indicating that the later infiltration of Ly6C+Ly6G+ cells is not specific to VACV infection. Closer analysis revealed that the Ly6C+Ly6G+ cells expressed several proteins considered to be monocyte/macrophage markers, including CD68, F4/80, CD200 and the scavenger receptor CD163 (Fig. 2C–F).
To investigate whether the Ly6C+Ly6G+ cells that we observed were monocytes, we infected CX3CR1+/GFP mice in which one of the copies of the CX3CR1 chemokine receptor has been replaced by GFP [26]. All monocytes in these mice express GFP at high levels [26]. Following VACV infection, the CD11b+Ly6C+Ly6G- cells in the ear pinnae expressed GFP (Fig. 2G). However, the CD11b+Ly6C+Ly6G+ cells did not express GFP driven by the CX3CR1 promoter, indicating that they are unlikely to be derived from the monocyte population (Fig. 2G). To further investigate the phenotype of infiltrating cell populations from the ear of VACV infected mice, we isolated Ly6G+ and Ly6G- cells from ear pinnae 5 days post-infection by magnetic separation and visualized their nuclear morphology by microscopy. Both Ly6G- and Ly6G+ fractions were comprised primarily of mononuclear cells (Fig. 2H). Therefore it appears that the Ly6C+Ly6G+ cells that infiltrated the site of VACV infection in the ear are phenotypically distinct from the classical description of both neutrophils and monocyte/macrophages.
To identify whether Ly6C+Ly6G+ cells migrate to the site of infection or expand in situ, we blocked infiltration of these cells to the infection site with a non-depleting antibody targeting the integrin subunit CD11b. Anti-CD11b antibody treatment reduced infiltration of both Ly6C+Ly6G- and Ly6C+Ly6G+ cells equally (Fig. 3A), indicating a similar requirement for transport across the endothelial cell layer for each of these cell populations. To study whether the Ly6C+Ly6G+ cell population we observed at the site of VACV infection is derived from the circulation, we injected fluorescent fluorospheres i.v. and examined the migration of cells that had ingested the fluorospheres to the site of VACV infection. Prior to fluorosphere injection, we depleted circulating monocytes with clodronate to ensure that we labeled cells that were mobilized from the bone marrow [27]. Similar profiles of Ly6C+Ly6G+ and Ly6C+Ly6G- cells in the ear contained fluorospheres after internalization in the blood (Fig. 3B), suggesting that both populations repopulate the circulation following clodronate depletion of monocytes. Taken together, these data indicate that neither population of cells is likely to be derived from resident cells, but rather each moves into the site of infection from the blood.
The true measure of cellular specialization is best demonstrated by the dedicated function of a cell type. Thus, we analyzed the Ly6C+Ly6G+ and Ly6C+Ly6G- populations for differences in function. The Ly6C+Ly6G- population at the site of infection expressed inducible nitric oxide synthase (iNOS) and, when exposed to CpG oligonucleotides, a portion of these Ly6C+Ly6G- cells produced TNF-α (Fig. 4A, B). In contrast, none of the Ly6C+Ly6G+ cells expressed iNOS above the level of CD11b- cells, nor did the Ly6C+Ly6G+ cells produce TNF-α either directly ex vivo or following stimulation with CpG oligonucleotides. Although both TNF-α [28] and iNOS [29], [30] have been reported to be required for efficient control of VACV replication in mice, we did not observe any difference in VACV replication in mice lacking iNOS compared to wild-type mice (Fig. S5). Production of Type I interferons is also essential to control VACV replication in vivo [31], [32] so we examined production of IFN-α and IFN-β by Ly6C+Ly6G- and Ly6C+Ly6G+ cells at the site of VACV infection. We found that Ly6C+Ly6G+ cells produced significant levels of Type I IFN when compared to either CD11b- cells or Ly6C+Ly6G- cells when directly isolated from the site of VACV infection (Fig. 4C, D). Notably, this production of Type I IFN occurred without the need for any additional stimulation. Ly6C+Ly6G+ neutrophils typically produce large quantities of ROS, so we incubated cells isolated from VACV-infected ears with a dye that becomes fluorescent upon exposure to ROS (CM-H2DCFDA). Ly6C+Ly6G- cells stained with the ROS substrate at a higher level than CD11b- cells, but Ly6C+Ly6G+ cells produced much higher levels of ROS (up to a 2 log10 shift in fluorescence) without additional stimulation (Fig. 4E). These data indicate that the Ly6C+Ly6G- and Ly6C+Ly6G+ cells are functionally distinct, and demonstrate that both cell types provide functions important in the control of VACV replication.
To ascertain whether production of Type I IFN by Ly6C+Ly6G+ cells is required for efficient control of VACV replication at the site of infection, we depleted Ly6G+ cells using 1A8 antibody specific for Ly6G. This antibody depletes Ly6G+ cells while leaving Ly6C+Ly6G- cells unaffected, in contrast to antibodies used to deplete Gr1+ cells, such as RB6-8C5 [25] (Fig. 5A). Mice receiving the anti-Ly6G 1A8 antibody displayed a modest 2.5-fold increase in virus replication in the ear pinnae on days 5 and 7 post-infection (Fig. 5B) comparable to that observed following clodronate liposome treatment or MAFIA-dependent depletion. In contrast to treatments that globally deplete monocyte/macrophages, such as clodronate treatment, administration of the 1A8 antibody resulted in undetectable levels of replicative VACV in the ovaries (data not shown). Virus replication in the ear was controlled by day 9 post-infection. Therefore, Ly6G+ cells, including the population of Ly6C+Ly6G+ cells we observed at the site of infection, likely play only a minor role in reducing virus replication and spread.
When we depleted mice of Ly6G+ cells, the pathology at the site of VACV infection was dramatically increased and large areas of infected ears became necrotic and eventually fell off (Fig. 5C). To quantify this pathology, we measured the lesion size and tissue damage, which represents the loss of necrotic tissue from the ear [8], in infected mice that were vehicle treated or depleted of Ly6G+ cells. Both lesion size and tissue loss was significantly greater in mice depleted of Ly6G+ cells (Fig. 5D, E) than in those treated with isotype control antibody (Fig. S6). Notably, both the increase in lesion size and the tissue loss in anti-Ly6G antibody-treated mice occurred primarily at time points after the minor increase in virus replication had been controlled. To ensure that the tissue protective function of Ly6G+ cells occurred after the control of VACV replication we injected anti-Ly6G antibody or isotype control on day 10 post-infection and monitored lesion size (not shown) and tissue damage (Fig. 5F). As above, depletion of Ly6G+ cells following control of virus replication enhanced tissue damage. These data indicate a tissue protective role for Ly6G+ cells during VACV infection.
Because production of ROS is the major functional phenotype of Ly6C+Ly6G+ cells following VACV infection, we investigated the role of ROS production in control of virus replication in the ear, control of virus spread to the ovaries, and tissue protection. We infected gp91-/- mice that lack the membrane component of the phagocyte NADPH oxidase, and therefore cannot generate ROS [33]. Gp91-/- mice displayed slightly (0.5-fold) enhanced replication of VACV in the ear at day 5 post-infection (Fig. 6A) and, similar to wild-type mice, no replicating virus could be detected in the ovaries (data not shown). However, the ears of gp91-/- mice displayed very similar characteristics upon infection to those of mice depleted of Ly6G+ cells, namely that large portions of the ear became necrotic and were eventually shed (Fig. 6B). When we quantified lesion size and tissue loss as outlined above, we observed that a lack of ROS significantly increased tissue damage at time points when there was no effect upon virus replication (Fig. 6C-D). There was no significant difference between the infiltration of CD11b+ or Ly6C+Ly6G+ cells between wild-type and gp91-/- mice, indicating that a difference in chemotaxis of tissue protective cells did not account for the difference in tissue damage observed in gp91-/- mice (Fig. S7). In addition, depletion of Ly6G+ cells in gp91-/- mice did not exacerbate or ameliorate tissue damage (Fig. S8) in contrast to wild-type mice, where treatment with anti-Ly6G dramatically increased tissue damage (Fig. 5 D, E). These data demonstrate that production of ROS, likely by Ly6C+Ly6G+ cells, prevents tissue damage following VACV infection.
The mechanisms responsible for the pathology observed at the site of VACV infection remain unknown, but a great deal of recent work has focused on the ability of myeloid cell population to suppress T cell activity. Therefore we examine the ability of T cells to induce tissue damage in mice lacking ROS. We depleted T cells with anti-Thy1 antibody (Fig. S2) and measured lesion size and tissue damage as above. If T cells were responsible for tissue damage and their function was modulated by ROS we would expect to observe a reduction in the tissue damage in gp91-/- depleted of T cells. We did not observe a decrease in lesion size (Fig. 6E) or tissue damage (Fig. 6F) in mice treated with anti-Thy1 antibody, and in some, but not all, experiments we observed an increase in damage in T cell depleted mice. Therefore ROS-mediated modulation of tissue damage is not achieved via an effect on T cell activity.
In this study, we describe two populations of innate immune effectors, identified as CD11b+Ly6C+Ly6G- and CD11b+Ly6C+Ly6G+ that migrate to a peripheral site of virus infection. These populations are phenotypically distinct, and mediate multiple functions that control virus replication, prevent systemic spread of virus, and simultaneously reduce tissue damage. We demonstrate the recruitment of a non-typical Ly6C+Ly6G+ population that appears to mediate both effector (Type I interferon production) and immunomodulatory (reduction of tissue damage) functions following virus infection. Depletion of this population of cells reveals that their function is vital in protection of tissue from catastrophic damage mediated by the inflammatory response. Manipulation of their function may allow the generation of tissue protective responses during infection to prevent immune-mediated pathology.
Our observation that Ly6C+Ly6G+ cells produce Type I interferons is in contrast to previous publications in which production of antiviral Type I IFN by innate immune effectors is typically held to be the role of plasmacytoid DC [34]. We did not observe CD11c+B220+ plasmacytoid DC at the site of infection, and we were unable to attribute an effector function such as cytokine or other inflammatory mediator production to the small number of CD11c+ cells with the phenotype of monocyte-derived DC at the site of infection. These “inflammatory DC” may play a role in antigen presentation to CD4+ T cells that migrate to the site of infection at later time points [35]. Following systemic VACV infection, TLR2-mediated recognition by CD11b+Ly6C+Ly6G-, but not Ly6C+Ly6G+, cells leads to the production of Type I IFN [19]. This apparent discrepancy could be explained by the ability of systemically administered VACV to reach lymphoid resident macrophage populations that have a differential ability to produce Type I interferon. The natural route of infection with VACV appears to be via touch [36], and dermal infection reveals a role for immune evasion molecules that other routes of infection do not [7]. If natural infection is via the dermal route, then only cells migrating to the site of infection may be exposed to the virus, explaining the difference in cell type producing Type I interferons in our study.
It is clear from our data that Ly6C+Ly6G+ cells are required to modulate the immune response and reduce tissue damage following infection with VACV. The enhanced damage observed following depletion of Ly6G+ cells is unlikely to result from the minor increases in VACV titers in the ear. A two-fold increase in virus titer is minor, as each infected cell will produce 102–103 progeny virions, so major changes in control of virus replication would likely produce log10 changes in titer. We show that Ly6C+Ly6G+ cells are present after clearance of virus, presumably to aid in recovery of the tissue from immune-mediated pathology. Modulation of tissue damage requires the production of ROS, which are often associated with tissue damage. Oxygenation is often required for tissue repair, however, and the presence of oxygen may allow greater production of oxygen radical that are required for tissue repair, or modulation of the immune response to reduce tissue damage [37], [38]. The production of ROS can modulate both T cell responses and innate immune responses [39], [40], [41], [42]. Our data indicate that T cell depletion in gp91-/- mice does not reduce the lesion size or tissue damage following VACV infection, suggesting that production of ROS by Ly6C+Ly6G+ cells modulates tissue damage in a T cell-independent manner. The exact mechanisms responsible for the profound damage found in the ear of mice infected with VACV remains unknown, and is a focus of our ongoing studies. The late time point at which damage occurs may reflect a role for antibody-mediated mechanisms that act through innate effector cells to initiate damage. Ly6C+Ly6G+ cells express several other molecules capable of suppressing immune responses, including CD163 and CD200. CD200-mediated suppression targets any cell expressing CD200R, including monocytes, macrophages, granulocytes, and T cells [43]. The activity of CD163, a scavenger receptor involved in the clearance of hemoglobin, leads to the up-regulation of the enzyme HO-1 [44]. In turn, this enzyme is both anti-inflammatory [45] and tissue-protective [46], [47], [48] through pathways involving CO, bilirubin, and Fe2+. In addition, ROS are critical mediators of signaling by cytokine and hormone receptors that may be required for tissue repair, such as insulin, platelet-derived growth factor, fibroblast growth factor and angiotensin [49]. Thus, Ly6C+Ly6G+ cells possess many mechanisms capable of modulating the immune response in order to provide tissue protection and repair.
The role of the Ly6C+Ly6G- monocyte population in immunity to a peripheral virus infection is less defined, as there is currently no method available to specifically deplete these cells without affecting the Ly6C+Ly6G+ cells. Ly6C+Ly6G- monocyte recruitment occurs prior to the recruitment of αβ T cells and coincides with the time point at which virus replication is controlled. By subtractive reasoning we are able to gain an insight into the role of these cells. The Ly6C+Ly6G- monocyte population may be required for control of virus replication at the site of infection, but systemic depletion of populations including these cells does not substantially increase virus titers in the ear. Clearly cells that are depleted by treatment with clodronate liposomes do prevent systemic spread of the virus via an unknown mechanism that does not involve iNOS. Subcapsular sinus macrophages have been proposed as a gatekeeper cell type that prevents systemic spread of viruses [50], [51] and these cells are known to be infected following VACV infection [52]. However we observed no depletion of subcapsular sinus macrophages following systemic depletion with clodronate liposomes, ruling out these cells as the ones responsible for controlling systemic spread of VACV following intradermal infection. It is possible that clodronate-mediated depletion of Ly6C+Ly6G- monocytes (or other cells) at sites other than the ear pinnae or subcapsular sinus is responsible for allowing systemic spread of VACV following peripheral infection, but we have been unable to identify specific populations of cells that are depleted by systemic administration of clodronate and definitely prevent virus spread.
The derivation of the Ly6C+Ly6G+ cell population we have described remains unknown. Ly6C+Ly6G+ cells are recruited from the blood at a time point after infection that is not normally associated with neutrophil recruitment. Ly6C+Ly6G+ cells have CD115 promoter activity at some point during their differentiation and display a mononuclear morphology but do not express the monocyte marker CX3CR1. The latter observation could be related to infection with VACV, as this virus expresses an immune modulator that causes the production of glucocorticoids [53], and in vivo administration of glucocorticoids can induce production of a monocytic cell type that downregulates expression of CX3CR1 [54]. These glucocorticoid-induced cells express many of the surface markers of myeloid derived suppressor cells (MDSC), a heterogeneous cell population described as suppressors of T cell responses in a tumor microenvironment [55], [56]. The Ly6C+Ly6G+ cells found at the site of VACV infection share expression of many surface markers with granulocytic MDSC, and produce large quantities of ROS, which MDSC use to modulate T cell activity [57], [58], [59]. However, MDSC are typically identified by expression of CD1b and Gr-1, a phenotype shared with neutrophils and inflammatory macrophages, and a full pathway of differentiation of these cells during resting or pathological conditions is yet to be published. The Ly6C+Ly6G+ cells we describe in this study share expression of some surface markers and some functions with neutrophil and monocyte populations, as well with MDSC. Without the definition of a widely agreed upon panel of markers that identify MDSC it is therefore impossible to conclude that the Ly6C+Ly6G+ cells are MDSC. Indeed, in the absence of definitive evidence that the Ly6C+Ly6G+ cells derive from a discrete lineage, their immunomodulatory function thus appears insufficient at this time to define the existence of a novel cell population.
In summary, we have identified and described the role of a distinct population of Ly6C+Ly6G+ cells that adds to the complexity of the phagocyte compartment [60]. These Ly6C+Ly6G+ cells are innate immune cells that regulate the destructive action of the innate immune system, reducing tissue damage and allowing wound healing. Modulation of the activity of these cells presents an attractive therapeutic strategy for preventing tissue damage in a wide range of infections and other pathologies.
All mice were housed in the specific pathogen free animal facility of the Hershey Medical Center. C57BL/6 mice were purchased from Charles River Laboratories (National Cancer Institute, Frederick, MD). iNOS-/- mice [61] were purchased from Taconic. Gp91-/- mice [33] and MAFIA mice [21] were purchased from Jackson Laboratory. All transgenic or knockout mouse strains were on the C57BL/6 background after a minimum of 12 backcrosses to this strain. Mouse strains, with the exception of C57BL/6, were subsequently bred at the Hershey Medical Center. All animals were maintained in microisolator cages and treated in accordance with the National Institutes of Health and American Association of Laboratory Animal Care (AAALAC International) regulations. All animal-related experiments and procedures were approved by the Penn State Hershey Institutional Animal Care and Use Committee.
Mice were injected in each ear pinna with 104 pfu VACV strain Western Reserve in a volume of 10 µl [8]. To analyze the presence of replicating virus, the ear pinnae or ovaries were harvested, subjected to three freeze/thaw cycles, homogenized and sonicated. Lysate was then placed on a monolayer of 143B cells, and a plaque assay was used to determine viral titer [8].
To deplete phagocytes, 200 µl of clodronate liposomes in PBS were injected i.v. on days 0, 1, 3, and 4 of infection [20]. This injection scheme effectively depletes monocytes and many macrophage populations [27], [62]. Cl2MDP (or clodronate) was a gift of Roche Diagnostics GmbH, Mannheim, Germany. Liposomes were prepared using Phosphatidylcholine (LIPOID E PC, Lipoid GmbH) and cholesterol (Sigma). Depletion of GFP+ cells in MAFIA mice was accomplished using AP20187 (Ariad Pharmaceuticals), as previously described [21]. AP20187 was diluted to a working concentration of 0.55 mg/ml in sterile water containing 4% ethanol, 10% PEG-400, and 1.7% Tween immediately before injection. AP20187 (10mg/kg) was injected i.v. daily for 5 days. Depletion was maintained with injections AP20187 (1mg/kg) every 3 days. Antibody-mediated depletion of Ly6G+ cells was by injection of 0.5 mg of anti-Ly6G (1A8, BioXCell) i.p. every 4 days [25].
Monocyte extravasation was partially blocked using the anti-CD11b antibody clone 5C6 [63]. Mice were injected i.v. with 0.5 mg/mouse of 5C6 antibody or an isotype control (rat IgG2b) every 24 hours. The clone 5C6 recognizes a epitope distinct from the anti-CD11b antibody clone M1/70 [63], so there is no 5C6-mediated interference in the detection of CD11b+ cells by flow cytometry.
Immune cells were isolated as previously described [9], [10]. Briefly, ear pinnae were microdissected to increase surface area and incubated in 1 mg/ml collagenase XI (Sigma) for 30 min at 37°C. The tissue was then passed through metal screens to create a single cell suspension. RBC were lysed using ACK lysis buffer (Invitrogen). Cells were then incubated in Fc Block (BD) and stained in a solution of antibodies diluted in Fc Block and 10% mouse serum (Sigma). To detect intracellular markers or cytokines, cells were first fixed in 1% paraformaldehyde and stained and washed in the presence of 0.5% saponin (Sigma). Antibodies were from eBioscience unless noted otherwise and included anti-CD11b (M1/70), -CD11c (N418), -CD19 (eBio1D3), -CD68 (FA-11), -CD90 (53-2.1), -F4/80 (BM8), -IFN-α (RMMA-1), -IFN-β (RMMB-1), -iNOS (6), -Ly6C (AL-21), -Ly6G (1A8), -NK1.1 (PK136), -TCRβ (H57-597, BD Bioscience), -CD163 (ED2, Serotec), and -TNF-α (MP6-XT22). Streptavidin (BD Bioscience) was used to label biotin-conjugated antibodies. For detection of production of IFN-α, IFN-β and TNF-α, ex vivo cells were incubated in the presence of 5 µg/ml brefeldin A (Sigma) for 4 hours prior to staining. To detect TNF-α production, cells were incubated in the presence of 20 µg/ml CpG (Invivogen) for 2 hours prior to the addition of brefeldin A. For flow cytometry analysis all sample acquisition was with a FACsCanto or LSRII (BDBiosciences) in the Hershey Medical Center Flow Cytometry Core Facility. Data were analyzed using FlowJo software (Treestar) as outlined in Fig. S3.
Cells isolated from infected ear pinnae, as described above, were incubated in 20 mM CM-H2DCFDA in PBS for 30 min at 37°C and developed for an addition 4–6 hours in the dark. Fluorescence was detected by flow cytometry.
Tracking of classical monocytes was accomplished using fluorosphere labeling as previously described [27]. A 2.5% solids solution of 0.5μm FITC-labeled latex beads (Polysciences, Inc) was diluted 1∶25 in sterile PBS. A 250 µl dose of this dilution was injected i.v. alone or 1 day following an i.v. injection of 200 µl clodronate liposomes. Phagocytes were then isolated from ear pinnae, stained, and analyzed using flow cytometry as described above.
Isolated cells were stained with a Phycoerythrin (PE)-conjugated anti-Ly6G antibody (clone 1A8), then incubated with anti-PE beads and separated using an AutoMACS sorter. The Ly6G+ and Ly6G- fractions (which contained large amounts of debris that were disregarded) were analyzed using Romanowsky staining.
|
10.1371/journal.pcbi.1002228 | Comparative Microbial Modules Resource: Generation and Visualization of Multi-species Biclusters | The increasing abundance of large-scale, high-throughput datasets for many closely related organisms provides opportunities for comparative analysis via the simultaneous biclustering of datasets from multiple species. These analyses require a reformulation of how to organize multi-species datasets and visualize comparative genomics data analyses results. Recently, we developed a method, multi-species cMonkey, which integrates heterogeneous high-throughput datatypes from multiple species to identify conserved regulatory modules. Here we present an integrated data visualization system, built upon the Gaggle, enabling exploration of our method's results (available at http://meatwad.bio.nyu.edu/cmmr.html). The system can also be used to explore other comparative genomics datasets and outputs from other data analysis procedures – results from other multiple-species clustering programs or from independent clustering of different single-species datasets. We provide an example use of our system for two bacteria, Escherichia coli and Salmonella Typhimurium. We illustrate the use of our system by exploring conserved biclusters involved in nitrogen metabolism, uncovering a putative function for yjjI, a currently uncharacterized gene that we predict to be involved in nitrogen assimilation.
| Advancing high-throughput experimental technologies are providing access to genome-wide measurements for multiple related species on multiple information levels (e.g. mRNA, protein, interactions, functional assays, etc.). We present a biclustering algorithm and an associated visualization system for generating and exploring regulatory modules derived from analysis of integrated multi-species genomics datasets. We use multi-species-cMonkey, an algorithm of our own construction that can integrate diverse systems-biology datatypes from multiple species to form biclusters, or condition-dependent regulatory modules, that are conserved across both the multiple species analyzed and biclusters that are specific to subsets of the processed species. Our resource is an integrated web and java based system that allows biologists to explore both conserved and species-specific biclusters in the context of the data, associated networks for both species, and existing annotations for both species. Our focus in this work is on the use of the integrated system with examples drawn from exploring modules associated with nitrogen metabolism in two Gram-negative bacteria, E. coli and S. Typhimurium.
| It is now routine to have genomics data for multiple organisms of interest. For example, data may be available for both an organism of primary relevance to a specific study, as well as data for related species. Tools and algorithms for comparative analysis of multi-species datasets are therefore in high demand. Comparative analysis of gene sequences is a mainstay in computational biology [1], but comparative methods for genomics and transcriptomics data analysis are relatively new, primarily due to the fact that only recently have researchers had access to large-scale datasets from multiple species [2], [3], [4], [5], [6], [7].
A number of tools are being developed for interpreting and exploring large-scale biological networks, such as: PathSys [8], NAViGaTOR [9], BIOZON [10], [11], BN++ [12], ONDEX [13], Cytoscape, and Osprey [14]. For a review of visualization tools for systems biology see [15]. Most tools focus on automated methods for integrating interaction datasets and displaying them graphically as network diagrams. Some contain novel data structures and data models, connections to databases, and many can incorporate additional data such as, abundance, sequence, literature derived and text mining derived data. These tools often contain functions for highlighting differences in the collected datasets. While the CMMR overlaps and encompasses many of the functionalities of these previously described tools, for example visualizing network graphs from a collected data compendium, its primary function is displaying the results of multiple-species integrated biclustering analysis.
Several recent studies have shown that comparative genomics analysis improves our ability to learn regulatory interactions, co-regulated groups, and to delineate the conserved components of fundamental pathways and modules [2], [16], [17], [18], [19], [20]. In particular, multiple-species clustering and biclustering can be used to detect conserved co-regulated gene groups and serve as a foundation to begin characterizing key differences in the regulatory programs of related species. In this work we present a data visualization system that enables the visualization and exploration of integrative multi-species biclustering analysis [20]. Our interface is built on a loosely coupled system architecture that connects multiple tools and databases using the Gaggle [21], Sungear [22], and Cytoscape [23]. This interface provides coordinated access to multiple-species clusters, biclusters and networks derived from comparative genomics analysis tools such as multi-species cMonkey (MScM) [20].
The analysis of multiple species datasets presents several challenges not encountered when analyzing single species datasets. In addition to the display and exploration of multiple datatypes, such as interaction networks, cis-regulatory sequences, transcriptome and proteome data, we add the challenge of tracking connections between orthologous groups of genes. In this work we focus on exploring sets of multi-species biclusters generated with MScM. A typical multi-species biclustering (set of biclusters) will consist of:
Our system to navigate this analysis enables exploration of both conserved biclusters, in the context of both species, and species specific additions to conserved biclusters, in the context of each individual species dataset, and illustrates general strategies for building loosely coupled systems for exploring other multi-species genomics analysis.
High-throughput data exists for many microbial organisms on multiple information levels (i.e. genome sequences, transcriptomics, proteomics, metabolomics, networks of pathways and interactions). Collecting and integrating diverse and heterogeneous datasets from disparate databases is not trivial and poses a number of barriers to automating the process. One of the most significant barriers to automation of data-import is the inconsistency among the naming schemes for loci, mRNA and protein products that are employed by the major public repositories such as NCBI, Uniprot and EMBL. Versioning can also be an issue if a given data source is delayed in updating their annotations. Our resource integrates diverse data from microarray experiments, genomic sequences, and various functional associations. It utilizes a database for translating gene names across datatypes and disparate resources and ortholog names across species, and is linked to the Gaggle. We will focus our examples on two closely related γ-Proteobacteria: E. coli and S. Typhimurium.
Clustering and biclustering are typically used to identify groups of co-expressed genes that, ideally, represent true regulatory modules and co-functional groups such as pathways and complexes. Biclustering groups genes into condition-specific gene clusters, and can allow genes to participate in more than one bicluster. Many biclustering methods have been previously described, for example, SAMBA [24], QUBIC [25], ISA [26], BIMAX [27], and NNN [28], and other algorithms [4], [29], [30], [31], [32]. Recent integrative biclustering methods, such as MATISSE [33], the recent version of SAMBA [19], and cMonkey [18], [20] have shown that incorporating additional datatypes, such as protein interactions and cis-acting regulatory sites, improves the performance of identifying of co-functional putative co-regulated modules. There are many benefits to comparing elements among species considering that a high fraction of co-regulated modules are conserved, in whole or in part, across species [3], [34]. Recent access to multiple genomics datasets from multiple species has allowed for new comparative analyses of genomics data, for example discovering regulatory elements [35] and the MScM algorithm [20] used here. MScM learns coregulated modules by integrating expression data across subsets of experimental conditions, co-occurrence of putative cis-acting regulatory motifs in the regulatory regions of bicluster members, functional associations and physical interactions. The output consists of condition dependent conserved modules of orthologous gene groups as well as species-specific elaborations of these conserved groups. The method is a true biclustering method: a typical conserved bicluster is typically supported by a subset of the input data for each species.
To enable exploration of a multi-species integrative biclustering result, we have constructed a system using the Gaggle and MScM (Figure 1). The Gaggle is a Java program that integrates tools by broadcasting gene, network and data selections between tools. For example, nodes selected in Cytoscape are sent to the Gaggle, which then sends the selections to all tools, which then automatically mirror those selections. The Gaggle has been shown to enable efficient creation of multi-tool systems to explore complex datasets and associated analysis [36]. Also, the loosely coupled visualization systems the Gaggle enables have several advantages including: systems-performance advantages – one tool crashing does not disable the whole system, development advantages – existing tools need not be reengineered and can be incorporated with small development costs, and maintenance advantages – due to the modularity of the resulting systems. We have extended the gaggle tools and built a corresponding database to give the user the ability to mirror gene selections in tools populated with results for one organism with the corresponding selection of the correct orthologs in the network, data, and bicluster views of another organism. Several component tools and databases are compatible with the Gaggle, or have been made compatible as part of this work, including: Sungear, Cytoscape, Cytoscape plugins such as BioNetBuilder [37], a Global Synonym/Ortholog Translator, and several tools designed to enable exploration of the genomics data available for each species (e.g. the data matrix viewer (DMV) and annotations viewer). Interactions among the CMMR, Gaggle tools and several online public databases containing annotations and genomic sequence is accomplished via a FireFox browser plugin called, the FireGoose [38] (available: http://gaggle.systemsbiology.net/docs/geese/firegoose). Selections in any tool are sent to the Gaggle which broadcasts both those gene selections to all tools for the organism in which the original selection was made and the orthologs in the other species of the selected genes. We show that this simple strategy enables effective exploration of this multi-datatype, multi-species integrative analysis.
We present an overview of the MScM algorithm, and the system we have constructed for visualizing the resulting multiple-species biclusters. Further methodological detail, additional validation of our method, and a full description of the dataset used to demonstrate our resource can be found in the supplemental section (Text S1).
Microarray data was acquired from several large, public repositories such as the Gene Expression Omnibus (GEO) [39], [40], ArrayExpress [41], [42], Stanford Microarray Database (SMD) [43], [44], Many Microbes Microarray database (M3D) [45], and KEGG Expression [46], with newer datasets manually obtained from individual publications. Genomic sequences corresponding to the upstream promoter regions of each predicted gene in each genome were retrieved from Regulatory Sequence Analysis Tools (RSAT) [47], [48]. Lastly, functional associations, in the form of interaction networks, were automatically acquired from multiple sources including Prolinks [49], Predictome [50], STRING [51], [52], and MicrobesOnline [53]. We have created a data compendium containing all publicly available data for a number of microbial species including several Gram negative species Escherichia coli, Salmonella Typhimurium, Vibrio cholerae, Helicobacter pylori, Desulfovibrio vulgaris; three related Gram positive species Bacillus subtilis, Bacillus anthracis, Listeria monocytogenes, and the archeon Halobacterium salinarum; within this compendium all name translations have been curated to minimize error due to incorrect translation of gene synonyms. In selecting this group of microbial species, we decided to start with the two most extensively studied bacterial model organisms, E. coli and B. subtilis, and included several closely related species and some representatives from important clades of the microbial tree of life. Additional species will be included in future versions of the database, as a sufficient amount of large-scale data becomes available for those species. A full listing of all datasets used in this study for both species, including references to papers describing both original collection and several databases that aided the import and curation of the datasets, are provided in the supplemental materials (Text S1).
The MScM algorithm consists of four main steps. Beginning with step 1, putative orthologous relationships between genes in each species are identified using InParanoid [54]. InParanoid identifies not only single gene pair relationships (one-to-one) but also families of homologous genes (one-to-many, many-to-many). This allows for flexibility when considering which orthologous gene pairs to cluster. Genes are often in several putative orthology relationships and selection of orthologous pairs, one pair per putative orthology relationship, is driven by the genomics data (see Text S1 for details). After defining the set of gene pairs spanning the two species, or orthologous core, step 2 identifies the conserved biclusters via an iterative Monte Carlo optimization of the MScM score. To determine the likelihood of an orthologous gene pair belonging to a bicluster, we first simultaneously compute single-species cMonkey scores for each gene supported by each organism's individual data space (expression, common sequence motif, and connected subnetwork). Then, we compute a single, multi-species score based on the combined single-species scores. The putative-orthology based gene coupling between species is removed in step 3, where each detected conserved bicluster is split into its two constituent single-species biclusters, then species-specific additions are made separately for each species using the single-species cMonkey score. The conserved core of the bicluster detected in step 2 is kept static while species-specific additions, including both non-orthologous and orthologous genes, to the conserved biclusters are discovered via this iterative optimization. An optional step 4, not carried out in this study, identifies purely species-specific biclusters for each organism using the original cMonkey algorithm applied to genes not yet in any conserved bicluster.
We have made the cMonkey and MScM code available including tools for automating many of the data acquisition and processing steps required for assembling an integrated dataset [55]. These tools facilitate automatic queries to online biological databases for association and upstream data, such as BioNetBuilder, MicrobesOnline [53], Prolinks [49], STRING [51], [52] and RSAT [47], [48]. All input and output are stored in a MySQL database to facilitate use of the integrated dataset and MScM results by other tools. We also include a manual mode with example inputs for the algorithm both as flat files and as R data objects for those wishing to use data not in public databases. These key changes to how data is imported and stored in the MScM database and the core data-object for cMonkey and MScM are critical novel changes to the code that are required for multi-species integration and scaling of the code to much larger datasets and organisms.
We created a database containing the MScM biclustering analysis data compendium for a number of microbial species. Our pipeline begins with several post-processing steps to convert cMonkey output to Gaggle compatible formats. Enrichment of functional annotations within biclusters is determined for each bicluster and the bicluster is assigned any significant annotations (p-values<0.05). A score is computed from the statistical components of each bicluster (e.g. residual, functional enrichment significance values). Specifically, the bicluster score is computed using Stouffer's z-score method for meta-analysis from a collection of bicluster statistics. Data files are generated for the complete bicluster network and the subnetwork of related biclusters before the website for a result is generated. Lists of orthologous genes between each species are generated as part of the analysis and loaded into the synonym/ortholog database.
To mirror selections simultaneously in several tools that visualize different aspects of the data, the results and the comparison between species we utilize the Gaggle, a loosely coupled system of web applications (geese) [21]. The Gaggle is a software framework that integrates independent application tools and biological data into an environment that allows the exchange of data among tools. All of the tools employed in our resource are Java web-starts or directly integrated into the web interface, thus removing any barrier to use based on tool compatibility, installation or data-transfer. The Gaggle also serves to coordinate the deployment and interoperation of these Java Web Start tools. Each individual application, or goose, can be launched with the click of a button on the BiclusterCard. The geese included in the resource are: a Global Synonym/Ortholog Translator, BioNetBuilder (Cytoscape plug-in), the FireGoose, Data Matrix Viewer, Annotations viewer, Cytoscape – bicluster network and gene network viewers, and Sungear. All the tools are connected through a communication hub called the Gaggle Boss, which passes simple messages among the geese, called broadcasting, summarized in Figure 1. When a broadcast is received, the goose will display the relevant information for the data. BiclusterCards and online databases (e.g. STRING, KEGG, etc.) connect to the tools through the FireGoose, a browser plug-in for Firefox adding the capability to communicate with the Gaggle. Embedded in each BiclusterCard is microformat code containing metadata for properties such as gene names, bicluster nodes, and condition names that can be broadcasted to other geese. The Bicluster Network viewer is a Cytoscape goose that displays a network of bicluster interactions, where nodes are biclusters, and edges are any shared properties (e.g. functional annotation, gene overlap, etc). Similarly the Gene Associations viewer is a Cytoscape goose that displays the gene associations from the data compendium. A Data Matrix Viewer goose acts as a spreadsheet program that can display and plot gene expression values. The Annotations goose displays a table of the genes and their various annotations specific to a single organism, for example, locus tag, gene name, protein id, and gene id accession. There is a Global Synonyn/Ortholog Translator that, given a list of genes from one species, can display the orthologous genes from another species. Lastly, the MScM output showing gene expression, gene subnetwork, sequence motifs, and motif locations in promoter sequence, can be displayed in the ClusterInfo Viewer.
A web interface was implemented to facilitate exploration of the multi-species biclusters. The starting page allows users to create several types of queries and contains a text box to input a gene name or group of genes, select boxes to choose bicluster sets from single and, core or elaborated MScM analyses, and a submit button to begin the search for biclusters containing the gene or genes of interest from the selected biclustering analyses (Figure 2A). Any biclusters returned from a search are presented as a list ranked by bicluster score. A first step in organizing the diverse information contained in, and supporting each bicluster was to create a system for generating bicluster summaries that link to online tools and source data. To this end, for each bicluster, our system creates a ‘BiclusterCard’. Each BiclusterCard provides the following information in the form of expandable/collapsible tabs (Figure 2B):
Each element of the bicluster card is generated automatically by our system, is compatible with outputs from other widely used biclustering tools, and provides links to descriptions/tutorials for using the linked tools or databases.
Visualizing the entire multiple-species dataset and integrative biclustering analysis at once, in a single view or tool, is cumbersome and ineffective at conveying biologically useful information due to the scale and multitude of different relationships in the data and analysis. Therefore, a main goal of our resource is to design an interface that provides access to the MScM results and collected data compendium via multiple queries (e.g. query by pathway, gene, network neighborhood, bicluster or ontology term). Although multiple queries are possible it is envisioned that a user will typically begin by querying for a gene or group of genes and browse MScM gene modules. A user can then begin exploring relationships between datasets for individual genes, subnetworks of genes, among modules, or among modules with particular shared attributes, such as, functional annotation. The system also allows high-level manipulation of queries, i.e. queries and operation on results of past queries, via Sungear. Examining the intersections, complements, and unions of module gene memberships, or identifying common promoter elements among genes in a module or among modules can be performed using Sungear following several broadcasts of gene lists. Gene lists are typically the results of queries, neighbors in a network loaded into the cytoscape goose, or the members of biclusters. These are just few examples of how a user can use the resource. Moreover, all of this functionality is automatically performed (mirrored) across species multiple species datasets.
To demonstrate our resource's capabilities, we explore nitrogen metabolism associated multi-species biclusters with the specific biological goal of identifying new genes functionally associated with nitrogen metabolism in E. coli and S. Typhimurium. For a global validation of our multi-species biclustering method and a detailed comparison of our method to several other methods, as well as a detailed description of the complete dataset used in this study see the supplemental section (Text S1) provided in the electronic version of this article. The CMMR is available at http://meatwad.bio.nyu.edu/cmmr.html.
Nitrogen is an essential input into several metabolic pathways including amino acid and nucleotide biosynthesis, and can act as a terminal electron acceptor in dissimilatory nitrate reactions [56]. It is common for some microbes including E. coli to use nitrogen for energy-harvesting purposes in anaerobic and nutrient depleted conditions [56]. A central component of nitrogen assimilation and metabolism is nitrate reductase, a membrane bound enzyme that catalyzes the conversion of nitrate to nitrite. The narGHJI operon encodes the multiple subunits of nitrate reductase A in E. coli. The following section sequentially guides the reader through using our system to explore biclusters containing genes in the nar operon and other nitrogen metabolism associated genes. A web tutorial for the use of our system can also be found at: http://meatwad.bio.nyu.edu/psbr/index.php/Tutorials
We begin our exploration of identifying conserved biclusters containing nar genes by searching for “narG” in the core set of genes from an E. coli and S. Typhimurium MScM bicluster set (Figure 2A). Explicitly, typing ‘narG’ into the gene-name textbox, selecting the core checkbox and clicking ‘submit’ on the CMMR start page, will retrieve any biclusters containing narG in the core set of genes. The results page returned following our “narG” query includes a header with links to the CMMR wiki, links to tutorials, a description of the search query and a list of any retrieved biclusters, in this case 3 biclusters were found (Figure 2B). There is a button for each bicluster that will display its BiclusterCard (see materials and methods). Looking at the first BiclusterCard for E. coli bicluster-57 (eco57), we will click on the ‘Coupled Bicluster’ button to open the BiclusterCard for S. Typhimurium bicluster 57 (stm57). Expanding the ‘Statistics’ tab shows that eco57 contains 75 genes (51 core genes, 24 elaborated genes), 226 experiments, whereas stm57 contains 66 genes (51 core genes, 15 elaborated genes) and 43 experiments (Figure 3A). This first table highlights differences in gene membership of the two biclusters. The ‘Enrichment Summary’ shows similar but not identical annotations involved in various metabolic activities related to anaerobic respiration and energy production from nitrogen for both biclusters (Figure 3B). The ‘Experiments’ tab shows that expression of these genes changes under a variety of conditions including: stress, growth on minimal media, anaerobic metabolism, and DNA damage. Expanding the ‘Enrichment Analysis’ tab displays tables containing significant COG, GO and KEGG annotations. We can see that eco57 and stm57 differ in the ranking of the KEGG pathway annotations and stm57 includes an additional pathway (Figure 3C). This could reflect slightly different uses of these modules in these organisms or discrepancies in the gene annotations.
Then, looking at the gene GO, KEGG and COG annotations by expanding the ‘Core Genes’ tab we see many genes have the same or similar annotations and some have either none or different annotations such as narG and yjjI (Figure 3D). Finally, under the ‘Plots’ tab we can view plots for gene expression profiles, bicluster mean expression, and an expression heatmap – to visualize differences in clustering bicluster gene members (Figure 4A).
Expanding the ‘Bicluster Motifs’ tab displays the motifs detected in the bicluster. Two of the detected motifs for eco57 show similarity to known nitrate/nitrite response transcriptional regulator binding motifs (Figure 4B). Motif #1 matches the E. coli FNR (fumarate and nitrate reduction) binding consensus sequence (TTGAT N4 ATCAA) [57] and eco57 motif #3 corresponds to the NarP binding sequence [58], [59]. The sequence motifs of stm57 show no notable similarity to known motifs. The FNR homolog in S. Typhimurium, oxaR, has a similar but less defined consensus sequence [60], which could account for the lack of association with stm57 motif #1. The promoter motif patterns display which gene members share common motifs and the location in the gene's upstream sequence. Identical motif patterns indicate they are an operon, such as operon narGHJI (Figure 4C). MScM and MicrobesOnline [61], [62] predict yjjI and yjjW to be in an operon, which is reflected in eco57 (yjjW is present in the elaborated gene set) but not stm57 (Figure 4C). Exploring the correspondence of the MScM detected motifs with known nitrogen metabolism motifs increases our level of confidence that this bicluster is truly coregulated in both organisms.
Among the core gene list for this bicluster, yjjI is described only as encoding a conserved protein with no functional annotation (Figure 3D). To examine this gene in the context of multiple network-types, the original data, and the biclustering, we now open several Gaggle tools, including the bicluster and gene network Cytoscape geese, Data Matrix Viewer and BioNetBuilder. First, we explore associations between core gene members of eco57 and stm57. For the 51 genes in the core gene member subnetworks, eco57 has 518 associations and stm57 has 420 edges, with no associations for yjjI (Figure 5A; associations shown are operon edges, metabolic pathway edges, phylogenetic profile edges, and protein interaction edges between genes in different biclusters). Next, we explore the expression profiles of the bicluster gene members and conditions by broadcasting them to the Data Matrix Viewer. Selecting yjjI, we can see that it has similar expression to other bicluster gene members (Figure 5B). Thus, the data (sequence motifs, associations, expression) supports eco57 and stm57 as coherent, putatively coregulated gene groups, and gene yjjI, while lacking associations, is supported by common motifs and correlated expression. We can use more Gaggle tools to search for additional information characterizing the bicluster gene members, particularly yjjI. For example, broadcasting the gene members to BioNetBuilder, we can browse protein structure and functional predictions. YjjI is predicted to have a domain structure that matches a “Class III anaerobic ribonucleotide reductase NRDD subunit” [63] and a function prediction of oxidoreductase activity [64], [65]. If we broadcast yjjI to other online databases such as Entrez Gene [66], we find that yjjI is adjacent to yjjW, but no information that they are in an operon. As mentioned above, both MScM and MicrobesOnline have predicted them to be in an operon. There is further information from EcoGene [67] reporting yjjI as an ortholog of H. influenzae hi0521, which is a pflB homolog and coding for a formate acetyltransferase [68]. Taken together, this information suggests a role for YjjI in nitrogen metabolism. It is important to note that a corresponding single-species bicluster in E. coli was not found (in the E. coli single species cMonkey run we find no bicluster with significant gene overlap to this significant conserved bicluster), further illustrating the importance of the MScM method. However, the species-specific elaborations of the bicluster may display additional information, such as, individual adaptations to this metabolic process.
Another possible use of our system is the exploration of collections of biclusters to identify novel interactions among modules. In the context of this example we can extract the subnetwork of biclusters related to the nar bicluster described above from a network that displays associations among biclusters by broadcasting the list of biclusters related to the orthologous core from the BiclusterCard to the Bicluster Network Viewer (Figure 5C). Biclusters are nodes with width and height proportional to the number of genes and conditions, respectively, and shared significant KEGG pathway, COG function, and GO function annotations are edges. The subnetwork shows 38 related biclusters for E. coli and 33 biclusters for S. Typhimurium. In this subnetwork there are several biclusters containing gene modules highlighting complementary interactions such as: amino acid biosynthesis/metabolism pathways and glutamate metabolism (bicluster-61); NADH dehydrogenase, succinate dehydrogenase (bicluster-43), citrate fermentation (bicluster-147), and amino acid ABC-type transporters (bicluster-148). This highlights the presence of conserved core interactions among eco57 and stm57 with other modules and independent species-specific modifications within these modules.
We can further explore nitrogen metabolism in the context of V. cholerae. First, we launch the Sungear goose and the Global Synonym/Ortholog Translator. From the subnetwork of related biclusters we select bicluster 57 and the top 3 overlapping biclusters (based on the ‘Related Biclusters -> Core Related’ table: 12, 83, 90). We then broadcast these 4 biclusters to the Sungear goose, select all groups and create a Sungear plot (Figure 6A). Next we select the vessels that have intersections with bicluster 57, yielding 39 genes. These E. coli genes are then broadcast to the Global Synonym/Ortholog Translator where we obtain 24 orthologs in V. cholerae (Figure 6B). Now, we launch the V. cholerae Bicluster Network Viewer by clincking the ‘B’ button on the CMMR start page next to the E. coli – V. cholerae MScM analysis. After the network has loaded, we highlight any biclusters containing those genes by broadcasting the translated orthologs to the bicluster network. This reveals 27 biclusters, of which only 3 are enriched for genes involved in nitrogen metabolism. Further investigation of the E. coli – V. cholerae MScM analysis shows that bicluster 109, a highly significant bicluster enriched for nitrogen metabolism in E. coli (eco109) but not V. cholerae (vch109), is absent from this list. Rather, vch109 is enriched for genes involved in molybdate ion transport and sulfur metabolism. The genes involved in nitrogen metabolism in eco109 are found in the elaborated set and not in the conserved core. This could represent a possible species-specific difference between these two organisms.
Using the CMMR, much knowledge was uncovered from the search of just a single gene, narG. In one case, for a currently uncharacterized gene, yjjI, the gathering of diverse information such as: putative orthology between two species, co-expression and common putative regulatory motifs with other bicluster genes, and a prediction for the protein's structure and function, was facilitated by the various BiclusterCards and Gaggle tools.
We have developed a publicly accessible web resource for comparative genomics studies of several prokaryotic organisms, with plans to expand this resource over time. As described above, in our example with coupled E. coli – S. Typhimurium bicluster 57, the combination of our method for simultaneously biclustering multiple datasets from multiple species and easy to use exploration system quickly led to novel biological insights and generate an informed hypothesis about the involvement of gene yjjI, a currently uncharacterized gene, in nitrogen metabolism. The complexity and richness of the results of comparative genomics data analysis requires a system like the one presented here. We present specific examples of the use of our system in the hopes of sparking discussion about what the next generations of comparative genomics analysis and visualization systems should look like. Our paper focuses on the combined, multi-tool interface required by biologists wishing to explore the biological significance and function of multi-species, multi-datatype biclusters and their species-specific elaborations and deletions. An important aspect of our system is the ability to submit new data for analysis and integrate the results into the resource for public access. We provide multiple avenues for researchers wishing to build this system for their species of interest, such as publicly available tools and code, and/or we will run our analysis and build this system for researchers without computational resources.
The CMMR wiki is intended to be a platform for information exchange, encouraging the contributions of researchers who use the resource, whether via curation or suggestions of new tools. Improvements to the resource could be made 1) in method development, for example, further optimization of the MScM algorithm and inclusion of additional analysis methods, 2) as datasets become available, increasing the number of included species, and 3) as further development and invention of intuitive visualization and exploration tools manifest. This effort could also serve as a framework for applications to comparative biclustering of eukaryotic organisms.
|
10.1371/journal.pcbi.1003016 | Synthetic Lethality between Gene Defects Affecting a Single Non-essential Molecular Pathway with Reversible Steps | Systematic analysis of synthetic lethality (SL) constitutes a critical tool for systems biology to decipher molecular pathways. The most accepted mechanistic explanation of SL is that the two genes function in parallel, mutually compensatory pathways, known as between-pathway SL. However, recent genome-wide analyses in yeast identified a significant number of within-pathway negative genetic interactions. The molecular mechanisms leading to within-pathway SL are not fully understood. Here, we propose a novel mechanism leading to within-pathway SL involving two genes functioning in a single non-essential pathway. This type of SL termed within-reversible-pathway SL involves reversible pathway steps, catalyzed by different enzymes in the forward and backward directions, and kinetic trapping of a potentially toxic intermediate. Experimental data with recombinational DNA repair genes validate the concept. Mathematical modeling recapitulates the possibility of kinetic trapping and revealed the potential contributions of synthetic, dosage-lethal interactions in such a genetic system as well as the possibility of within-pathway positive masking interactions. Analysis of yeast gene interaction and pathway data suggests broad applicability of this novel concept. These observations extend the canonical interpretation of synthetic-lethal or synthetic-sick interactions with direct implications to reconstruct molecular pathways and improve therapeutic approaches to diseases such as cancer.
| Organizing gene functions into molecular pathways is a major challenge in biology. The observation that two viable gene mutations become lethal when combined as a double mutant has been developed into a major genetic tool called synthetic lethality. The classic interpretation of synthetic lethality stipulates that the two mutations identify genes that work in parallel, mutually compensatory pathways that together perform an essential function. However, a significant number of negative interactions are caused by defects affecting a single molecular pathway. Here, we recapitulate by mathematical modeling recent experimental data that demonstrate synthetic lethality between mutations in genes acting in a single, non-essential molecular pathway. We propose a novel mechanism involving reversible pathways steps and trapping of an intermediate. The modeling also predicts that overexpression of certain genes functioning in reversible pathways will lead to synthetic lethality with gene defects in the same pathway. Our results significantly broaden the interpretation of synthetic lethal and synthetic dosage effects, which fundamentally impacts the assignment of genes to pathways. The concept of synthetic lethality has been applied to cancer therapy, and our modeling results suggest new approaches to how to target a single pathway to induce synthetic lethality in cancer cells.
| Synthetic interactions between two mutations in different genes were first revealed in Drosophila by Dobzhansky in the 1940s [1]. Synthetic lethality (SL) describes that two viable single gene mutations lead to lethality (synthetic-lethal) or severely impair growth (synthetic-sick) when combined as a double mutant. This concept was implemented as a powerful research tool for molecular pathway analysis in yeast [2]–[5]. Further refinement introduced more quantitative measures of genetic epistasis [6] and lethality induced by gene overexpression in a mutant background (synthetic dosage-lethality [7]). A genetic interaction is negative or aggravating, when the combined effect of two gene defects is more severe than it is expected from a simple multiplicative model. In a positive or alleviating interaction the effect is less severe than expected. These approaches and measures are now increasingly used in mammalian cells exploiting RNA-mediated gene knockdown technologies [8], [9].
Following a proposal by Hartwell and colleagues [10], SL has been utilized as a therapeutic approach in cancer treatment employing a combination of genetic ablation (loss of tumor suppressor) and chemical inhibition. The first paradigm was set in BRCA1/2-deficient tumor cells, which are synthetic-lethal with inhibition of PolyADP-Ribose Polymerase (PARP) [11]–[13]. Small molecule PARP inhibitors are currently being evaluated in clinical trials in BRCA1- and BRCA2-deficient cancers (e.g. [14]).
The canonical interpretation of SL stipulates two mutually compensatory, parallel pathways capable of performing the same essential function [2]–[4]. Thus, disrupting a single pathway is viable, while disrupting both pathways is lethal. This concept of between-pathway synthetic lethality (bpSL) (Figure 1A) led to the creation of computational approaches aiming at reconstructing interaction networks from pair-wise gene deletions or siRNA-induced gene knock-down screens in yeast and mammals [15]–[17]. However, recent genome-wide genetic interaction data revealed multiple negative interactions between mutations affecting the same molecular pathway or complex [3], [5],[16],[18]–[21]. For example, it was estimated that ∼9% [19] and in another study 14% [17] of all negative genetic interaction clusters belong to the same biological pathway. Several mechanistic models were suggested to explain within-pathway SL (wpSL) [6], [16], [20], [22]. The deletion of a gene might lead only to a partial degradation of an essential pathway which might be tolerable, whereas the double mutation leads to complete pathway degradation and lethality (Figure 1B1). This is especially relevant for the interpretation of siRNA-based screens where the efficiency of a particular gene knock down is uncertain. A second possible mechanism suggests that steps in an essential pathway are internally redundant (Figure 1B2). Lastly, two mutations may cumulatively degrade an essential protein complex, whereas they are individually viable (Figure 1B3). This mechanism is consistent with the observations that molecular complexes are frequently characterized by the dominance of negative over positive genetic interactions between their components [18]. wpSL interactions between defects in components of a single protein complex are highly enriched for complexes with an essential component [16], [22]. It was estimated that the contribution from within-complex interactions to the total number of within-pathway negative interactions does not exceed 7% [19]. Common to these mechanistic explanations of wpSL is that they involve either an essential pathway or an essential protein complex.
Here, we highlight a novel scenario of wpSL involving two components of a non-essential pathway. The view of molecular pathways as unidirectional, linear reaction cascades is too simplistic. Pathway steps can be reversible which leads to forward and backward propagation of molecular events along the pathway increasing robustness and fidelity of the overall process [23]–[28]. Affecting both forward and reverse steps of the pathway by abrogating the corresponding enzymes creates scenarios in which the pathway flow can be trapped in an intermediate state that may be toxic to the cell or deprive the cell of a limiting resource (Figure 1C). This can create a genetic scenario we define as within-reversible-pathway synthetic lethality (wrpSL), which is the subject of this study.
Here, we study bpSL and wrpSL scenarios using mathematical modeling to better understand the system properties of these genetic relationships. We present a simplified model of the pathway applicable for its formal analytical study and performed in silico simulations for bpSL and wrpSL as well as synthetic dosage effects. Our main experimentally confirmed examples of wrpSL are in the homologous recombination DNA repair pathway. Homologous recombination (HR) is an important mechanism to maintain genome integrity [29] (Section S1 and Figure S1 in for more discussion). Analysis of yeast gene interaction and pathway data suggests broad applicability of this novel concept.
In order to assess the importance of within-pathway negative interactions we ranked all pathways from the KEGG database [30] according to their normalized proportion of negative interactions [5] within each KEGG pathway (Figure 2). This analysis confirms the previous conclusion [19] that only a minority of within-pathway negative interactions can be explained by negative interactions within a complex (Figure 1B3). In our analysis only 12% of all negative interactions were of this type (compared to 7% in [19]). Interestingly, HR ranks at the top with 27 within-pathway negative interactions between 20 KEGG pathway components (Figure 2), of which only a single one affects components of the same protein complex.
Recent studies show that individual reaction steps in HR are reversible [23]–[25] (Figure S1 in Text S1). The Rad51-ssDNA filament is a key intermediate in HR, as it performs the signature reactions of homology search and DNA strand invasion. The formation of this filament is catalyzed by specific co-factors (see Section S1 and Figure S1 in Text S1). The helicase Srs2 specifically targets the Rad51-ssDNA filament for disruption to reverse filament formation [31], [32], [33]. The reversibility of the Rad51-ssDNA filament sets a new paradigm and draws attention to additional reversible steps and their mechanisms in HR, other DNA repair processes, and unrelated molecular pathways.
We derived the simplest linear mathematical model of a main DNA repair pathway with reversible steps and a toxic intermediate, and a compensatory pathway, which can recapitulate bpSL and wrpSL (Figure 3). Each state transition is catalyzed by an abstract enzyme, which may correspond to several biological entities (compare Figure S1 in Text S1 with Figure 3). In wrpSL trapping of toxic intermediate I is caused by defects in the first backward reaction (I→S, R1, k−1) and the second forward reaction (I→P, F2, k2). The reversibility of the second step (reaction P→I, R1) is not essential for wrpSL to occur, but might be important for quantitative pathway characteristics. Introduction of a final irreversible step (Figure S2 in Text S1) would result in a kinetic proofreading mechanism [34] (see Figures S2, S3 in Text S1 and discussion there). Such a mechanism increases the robustness of DNA repair, as it avoids a futile P↔I cycle. However, in this simplest model we eliminated the final irreversible step to allow us analyzing the most essential features of wrpSL (Section S2 in Text S1). Figure 4 explores conditions for various cellular fates (Normal Robust, NR: no single knockout leads to lethality (Figure S4 in Text S1); Normal Fragile, NF: single knockout can lead to lethality (Figure S4 in Text S1); Compensated, C: repair is performed by compensatory pathway; death due to DNA Damage, DD: steady state probability of DNA damage >50%; and Death due to Toxic intermediates, DT: steady state probability of toxic intermediate >50%). Figure 5 visualizes parametric conditions (see Section S2B in Text S1 for discussion).
Using analytical study and numerical simulations with some characteristic choices of kinetic rate values, we explored the dynamical behaviors of the simplest model (see Figure 6). Here, we discuss the qualitative results and interpretations, while the more formal derivation of these statements is found in Section S3 in Text S1. To illustrate the static and dynamic properties of the toy model, we selected two typical positions (Figure 5 #1, #2) corresponding to NR and NF pathway states, respectively. From these “normal” conditions we simulated a number of single and double mutant/overexpression conditions as shown in Figure 5 (see also Figure S4 in Text S1).
Synthetic lethality/sickness and synthetic dosage lethality are important genetic tools to assign individual gene functions into molecular pathways [2]–[4], [7]–[9]. The canonical interpretation for two mutants found to be synthetically lethal or sick stipulates that the encoded gene products function in different parallel pathways that can mutually compensate (bpSL) [2]–[4], [7]–[9], [15]–[17]. However, computational analysis of genetic interaction data combined with protein interaction data revealed multiple negative interactions between mutations affecting functions in the same molecular pathway or complex (wpSL) [3], [5], [16], [18]–[21]. Several mechanisms of wpSL have been proposed (Figure 1B), and they all involve either essential pathways or essential protein complexes. In extension of this fundamental concept of wpSL, there are several cases of SL between mutants encoding proteins acting in HR, a pathway that is not essential in yeast [35], [46]–[48]. We term this novel genetic interaction within-reversible-pathway Synthetic Lethality (wrpSL; Figure 1C) and provide a novel mechanistic explanation for wpSL, which can create SL within a non-essential pathway or between hypomorphic mutations in an essential pathway that is different from a model invoking sequential pathway degradation by accumulation of partial defects of successive steps (Figure 1B1).
Here, we explore by mathematical modeling the system properties of wrpSL. The modeling must make assumptions about the system properties (state transition rates, relative pathway efficiencies, etc.) and identifies several conditions to be met for wrpSL. 1) Reversibility of pathway steps. In fact, only the first pathway step must be reversible, whereas reversibility of the second pathway steps enables additional genetic scenarios. 2) Possibility of kinetic trapping of an intermediate state of the pathway when both the backward and forward reactions are compromised. The trapping per se can be detrimental due to blockage of cell signaling, sequestering an essential compound, or toxicity. We have assumed lethal toxicity in our model. 3) The possibility of rescue by a parallel compensatory pathway may not be strictly required, but highlights the applicability of this concept to non-essential pathways.
The mathematical model is validated by the experimentally observed recombination-dependent SL of the srs2 rad54 double mutant in budding yeast [35] (Figure 3, Figure S1 in Text S1, Figure 6, row 5). Srs2-defective cells are unable to reverse Rad51-ssDNA filaments. These Rad51-ssDNA filaments represent toxic intermediates that accumulate in the cell due to kinetic trapping and interfere with cell viability. The key functions of the Rad54 protein are to assist in DNA strand invasion and allowing DNA synthesis off the invading 3′-end [36]. Hence, in the srs2 rad54 double mutant Rad51-ssDNA filaments and/or D-loops may accumulate forming a toxic intermediate that leads to cell death (Figure S1 in Text S1; Figure 3 green pair and Figure 6, row 5). This interpretation is supported by the observation that lethality in this double mutant is suppressed by a defect in Rad51-ssDNA filament formation (mutations in RAD51, RAD55, RAD57, or RAD52) [49] (see Figure S1 in Text S1), what has been termed recombination-dependent lethality. Preventing Rad51-ssDNA filament formation allows bypass of recombination by alternative means of DNA repair (for DSBs: Nonhomologous endjoining or single-strand annealing; for gaps: Translesion synthesis or fork regression [23]; see Figure S1 in Text S1). The recombination-dependent lethality of srs2 rad54 is not unique and is also found in additional double mutants in recombinational repair genes including the double mutants mph1 mus81, mph1 mms4, srs2 sgs1 and sgs1 (or top3 or rmi1) and mus81 (or mms4) which likely reflect additional examples of wrpSL possibly involving different toxic intermediates [35], [46]–[53]. As discussed in detail in Section S1 in Text S1, the synthetic lethalities involving sgs1 are more complex, because of the multiple roles of Sgs1-Top3-Rmi1 in HR, and could be caused also by other mechanisms of SL.
Further modeling revealed additional genetic conditions including overexpression of specific pathway enzymes that are predicted to lead to wrpSL (Figure 6). The mathematical modeling also reveals the importance of reversible pathway steps, which are validated by genetic and biochemical experiments in yeast [23]–[25]. First, the existence of reversible pathway steps does not affect normal pathway progression (Figure 6, rows 1–3). Second, reversible pathway steps allow much more efficient and timely use of compensatory pathways (Figure 6, row 6). Third, reversible pathways coupled with compensatory pathways avoid lethality of single mutations (Figure 6, row 7). The existence of reversible intermediates in HR, and possibly other molecular pathways, has been proposed to increase the robustness of the overall DNA repair system [23]–[25], and here we provide quantitative modeling evidence and formal analysis of this assertion.
An important question is how general wrpSL might be or whether it is an idiosyncrasy of the recombinational repair pathway. Even if wrpSL were restricted to HR, this concept provides significant potential application in anti-cancer therapy. However, there is considerable evidence that many molecular pathways include reversible steps catalyzed by different enzymes in the forward and backward directions (see Figure 8). Any of those processes can be theoretically trapped into one of their intermediate states if two regulators of forward and backward steps are inactive. In these cases, the accumulating intermediate might be toxic, block proper signal propagation or prevent resource recycling. Focusing on three examples of reversible protein modifications (phosphorylation by Cdc5/dephosphorylation by Cdc14, sumoylation by Slx5–Slx8/desumoylation by Ulp1, Nup60, ubiquitylation by Rad6–Rad18/deubiquitylation by Bre5, Ubp3 or degradation dependent on Doa1, Rpn6; see Figure S6 in Text S1 for details), we found ample evidence in published genetic interaction data that are consistent with the wrpSL mechanism. These examples have not been fully explored, but are consistent with the wrpSL concept and amenable to test specific predictions.
In summary, genetic and biochemical data strongly support the significance of the wrpSL mechanism in HR, and existing data are consistent with the notion that wrpSL could be a general, widely applicable type of genetic interaction. This may refine our understanding of relationships between gene products and will help to improve pathway reconstruction. In particular, our mathematical modeling provides a conceptual framework for guiding systematic exploitation of mutations and changes in the expression profiles of HR genes and potentially genes of other pathways to induce SL.
The simplest mathematical model of Figure 3 was converted into a set of linear ordinary differential equations using the standard chemical kinetics formalism. The steady state model properties were analyzed analytically and exemplified with numerical simulations. Classification of the pathway states according to the extreme (large or small) values of the control parameters and the corresponding asymptotic solutions follow the methodology of the asymptotology of reaction networks [54].
All numerical simulations were performed using SBTOOLBOX package for Matlab (Section S4 in Text S1).
|
10.1371/journal.pntd.0000874 | Cost-Effectiveness of a Chemoprophylactic Intervention with Single Dose Rifampicin in Contacts of New Leprosy Patients | With 249,007 new leprosy patients detected globally in 2008, it remains necessary to develop new and effective interventions to interrupt the transmission of M. leprae. We assessed the economic benefits of single dose rifampicin (SDR) for contacts as chemoprophylactic intervention in the control of leprosy.
We conducted a single centre, double blind, cluster randomised, placebo controlled trial in northwest Bangladesh between 2002 and 2007, including 21,711 close contacts of 1,037 patients with newly diagnosed leprosy. We gave a single dose of rifampicin or placebo to close contacts, with follow-up for four years. The main outcome measure was the development of clinical leprosy. We assessed the cost effectiveness by calculating the incremental cost effectiveness ratio (ICER) between the standard multidrug therapy (MDT) program with the additional chemoprophylaxis intervention versus the standard MDT program only. The ICER was expressed in US dollars per prevented leprosy case.
Chemoprophylaxis with SDR for preventing leprosy among contacts of leprosy patients is cost-effective at all contact levels and thereby a cost-effective prevention strategy. In total, $6,009 incremental cost was invested and 38 incremental leprosy cases were prevented, resulting in an ICER of $158 per one additional prevented leprosy case. It was the most cost-effective in neighbours of neighbours and social contacts (ICER $214), slightly less cost-effective in next door neighbours (ICER $497) and least cost-effective among household contacts (ICER $856).
Chemoprophylaxis with single dose rifampicin given to contacts of newly diagnosed leprosy patients is a cost-effective intervention strategy. Implementation studies are necessary to establish whether this intervention is acceptable and feasible in other leprosy endemic areas of the world.
| In 2008, 249,007 new leprosy patients were detected in the world. It therefore remains necessary to develop new and effective interventions to interrupt the transmission of M. leprae. We assessed the economic benefits of single dose rifampicin (SDR) for contacts as chemoprophylactic intervention in the control of leprosy. The study is based on a large trial including 21,711 contacts of 1,037 patients with newly diagnosed leprosy. We gave a single dose of rifampicin or placebo to contacts and followed them up for four years. The main outcome measure was the development of clinical leprosy. The cost effectiveness was expressed in US dollars per prevented leprosy case. Chemoprophylaxis with SDR for preventing leprosy among contacts of leprosy patients is cost-effective at all contact levels and thereby a cost-effective prevention strategy. In total $6,009 was invested and 38 leprosy cases were prevented after 2 years, costing $158 per prevented leprosy case. Implementation studies are necessary to establish whether this intervention is acceptable and feasible in other leprosy endemic areas of the world.
| Leprosy is a chronic infectious disease, caused by the bacillus Mycobacterium leprae, which affects the skin and peripheral nerves leading to skin lesions, loss of sensation, and nerve damage. This in turn can lead to secondary impairments or deformities of the eyes, hands and feet. For treatment purposes, leprosy is classified as either paucibacillary (PB) or multibacillary (MB) leprosy. The standard treatment for leprosy is multidrug therapy (MDT) [1]. PB patients are treated for 6 months with dapsone and rifampicin; MB patients are treated for 12 months with dapsone, rifampicin and clofazamine.
The World Health Organisation (WHO) had set a goal in the early 1990s to eliminate leprosy as a public health problem by the year 2000. Elimination was defined as reducing the global prevalence of the disease to less than 1 case per 10 000 population [2]. The WHO elimination strategy was based on increasing the geographical coverage of MDT and patients' accessibility to the treatment. The expectation existed that reduction in prevalence through expanding MDT coverage would eventually also lead to reduction in incidence of the disease and ultimately to elimination in terms of zero incidence of the disease. An important assumption underlying the WHO leprosy elimination strategy was that MDT would reduce transmission of M. leprae through a reduction of the number of contagious individuals in the community [3]. Unfortunately, there is no convincing evidence for this hypothesis [4].
With a total of 249 007 new patients detected globally in 2008 [5], it remains necessary to develop new and effective interventions to interrupt the transmission of M. leprae. BCG vaccination against tuberculosis offers some but not full protection against leprosy and in the absence of another more specific vaccination against the bacillus other strategies need to be developed, such as preventive treatment (chemoprophylaxis) of possible sub-clinically infected people at risk of developing leprosy. Recently, the results were published of randomised controlled trial into the effectiveness of single dose rifampicin (SDR) in preventing leprosy in contacts of patients [6]. It was shown that this intervention is effective at preventing the development of leprosy at two years and that the initial effect was maintained afterwards.
In order to assess the economic benefits of SDR as an intervention in the control of leprosy, we performed a cost-effectiveness analysis. We provide an overview of the direct costs of this new chemoprophylaxis intervention and calculate the cost-effectiveness compared to standard MDT provision only.
This study was based on the results of the prospective (sero-) epidemiological study on contact transmission and chemoprophylaxis in leprosy (COLEP; ISRCTN 61223447), which was conducted in the Rangpur and Nilphamari districts of northwest Bangladesh between 2002 and 2007 by the Rural Health Program (RHP) of The Leprosy Mission Bangladesh. The population of the two districts was 4.4 million in 2002, and the number of newly detected leprosy cases in 2002 was 1 317 [7]. Of these, 1 037 patients were included in the COLEP study; 400 with single lesion PB leprosy, 342 with PB leprosy of 2–5 lesions, and 295 with MB leprosy. Intake started in June 2002. Contacts were categorised according to their physical distance to the index patient [8]. For physical distance we defined six categories on the basis of the local housing situation: shares a house only (R) or a house and a kitchen (KR), shares a kitchen only (K), next-door neighbours (N1), neighbours of the neighbours (N2), and social (S) contacts (business contacts and colleagues staying in the same room for at least four hours a day, five days a week).
The COLEP study was a single centre, double blind, cluster randomised, placebo controlled trial. A complete description of the COLEP trial is given by Moet et al. [6]. In short, at intake - that is, after the index patient had received the second supervised dose of MDT – all contacts of one patient received prophylactic treatment, which included either capsules with 150 mg rifampicin or identical placebo capsules without an active (antibiotic) ingredient. According to bodyweight and age, each contact took two to four capsules under direct supervision of a staff member. Of the 1 037 patients, 517 were allocated to the intervention arm and 520 to the placebo arm of the trial. The number of contacts in the intervention arm was 10 857 and 10 854 in the placebo arm. A follow-up investigation took place two years after intake, starting in June 2004, and completed in February 2006. The primary outcome of the trial was the development of clinical leprosy.
Ethical clearance was obtained from the Ethical Review Committee of the Bangladesh Medical Research Council in Dhaka (ref. no. BMRC/ERC/2001–2004/799). All subjects were informed verbally in their own language (Bangla) about the study and invited to participate. Written consent was requested from each adult. For children consent from a parent or guardian was given.
Cost calculations were done from the health care perspective, in which real medical costs were calculated for the chemoprophylaxis intervention compared to the standard MDT treatment program. All direct medical costs of the general health program and the related indirect costs (e.g. transport) for the period 2002–2004 were included. Real medical costs were calculated by multiplying the volumes of health care use with the corresponding unit prices. The calculations of the full cost of the standard MDT treatment program and the rifampicin chemoprophylaxis were based on real resources. If information on resource use and the full cost were available, bottom-up calculations were performed [9]. If detailed information about resource use and unit costs were not available, top-down calculations were performed [10]. Prevention of disabilities, patients' costs and costs caused by loss of production due to absence from work were not taken into account because no reliable data were available. Table 1 provides an overview of the cost calculation method and data source per cost category. All costs were converted into unit prices in US dollars per volume, using the exchange rates of the UN for Bangladeshi taka (BDT) in US dollars ($) [11].
The costs of the standard MDT treatment program consisted of the leprosy costs made by the RHP in 2004, including all program costs such as personnel costs (salaries, allowances and staff benefits for all administrative, financial and field staff), transportation costs, surveys (contact and village), overhead costs (administration, repair and maintenance, health education). All data were adjusted for the leprosy share only of the RHP program and based on costs of 2002 and 2003 [12], [13], and extrapolated to 2004 with the corresponding inflation rate [11] because no reliable detailed information was available for 2004. The treatment for leprosy was based on data of the Novartis Foundation for Sustainable Development, which supplies the MDT free of charge [14]. Costs of surgical interventions were based on reconstructive surgery and corresponding hospital costs of 2002. For the treatment of complications all medical intervention and hospital costs were included, based on the annual report of the RHP. The number of patients needing complication treatment was calculated according to the Bangladesh Acute Nerve Damage (BAND) study [15]–[17]. Information from this study was necessary because no data were available for the number of leprosy patients with complications in the COLEP trial. The BAND study was a prospective cohort study of 2 664 new leprosy cases from the same area and the same population as the COLEP study. Over a period of three years incidence rates were calculated with the number of patients developing the following complications: nerve function impairment, type 1 and type 2 reactions, and silent neuritis. Recorded complication rates in the BAND study were extrapolated to the COLEP cohort. The various complications require different treatment regimens, e.g. mild type 1 reaction requires less corticosteroids than severe type 1 reaction. The appropriate treatment for each complication and the associated costs were based on the guidelines for leprosy treatment current at the time of the study [18]. Expert clinical opinion was taken for the average number of mild or severe reaction types. For reversal reactions (RR or Type 1) a distribution of 50% mild and 50% severe reactions was taken, and for erythema nodosum leprosum (ENL or Type 2) per leprosy patient having a reaction an average of two mild and two severe recurrences were taken, costing $ 3.41 and $ 69.58 per patient involved, respectively.
The chemoprophylaxis intervention is additional to the standard MDT treatment program. Therefore, costs for the intervention consist of two components, the basic health care (represented by the standard treatment) and the intervention costs. Therefore, no extra costs for basic health care were added since the intervention is additional to existing practice. For the cost of chemoprophylaxis, COLEP costs were used based on the cost statements of the COLEP Research Study 2002–2004. The chemoprophylaxis costs were calculated bottom-up from the COLEP data base by multiplying the cost for the mean number of capsules ($ 0.21 for 3.5 capsules) with the number of contacts in the intervention. Other medical costs consisted of MDT for all newly detected leprosy patients and treatment of complications.
Cost effectiveness of the COLEP trial was assessed by calculating the incremental cost effectiveness ratio (ICER) between the standard MDT program with the additional chemoprophylaxis intervention versus the standard MDT program only. We chose the perspective of the program level, calculating the total costs of the program for the standard treatment and added per treatment arm the costs for medical treatment. The difference in treatment costs with and without the intervention was calculated and compared with the difference in the number of newly detected cases among the contacts. Costs and benefits were modelled in a decision tree [9], in which effectiveness was measured by the number of prevented leprosy cases (Figure 1). The ICER was expressed in US dollar per prevented leprosy case as follows:where ‘I’ denotes intervention and ‘ST’ refers to Standard treatment. Costs and effects were discounted annually with 3.5% [19].
Sub-group analyses were also carried out by distance group as defined in studies of Moet et al. The ICER for a specific subgroup was dependent on: i) the differences between the whole program costs, ii) the differences in the number of recipients of additional treatment and consequential costs and iii) differences in the number of new leprosy cases found in the sub-group [6], [7].
To evaluate the uncertainty around the ICER, sensitivity analyses were performed by probabilistic sensitivity analysis (PSA) drawing 500 random samples. The costs and the efficiency of the program were deemed as deterministic but beta distributions were assigned to the complication probabilities to model the uncertainty around it [20]. Parameters of the beta distributions were based on the BAND study.
In total 20 032 contacts of 1037 leprosy patients remained in the trial after 2 years (taking into account loss of follow-up of 7.7%); 10 038 in the standard treatment (placebo) arm and 9 994 in the intervention (chemoprophylaxis) arm of the trial. The distributions of contacts over the two arms of the trial according to the physical distance of the contact to the index patient and the number of new leprosy cases detected in the contact groups after two years are shown in Table 2. The overall reduction of leprosy in the rifampicin arm of the trial compared to the standard treatment arm was 38 cases (57%).
Table 3 shows the total costs of different cost categories in the contact group of standard MDT treatment and of the chemoprophylaxis intervention. Program costs, which consisted of the costs of personnel, transportation and overhead, were higher for the intervention group, with around $ 4 000 due to extra personnel and transportation requirements of the program. The medical costs among the index patients of both groups amounted to approximately $ 5 700, and consisted of the treatment of leprosy with MDT and the treatment of complications. MDT treatment totalled up to $ 4 500, whereas complications estimated to have a burden of $ 1 180. The cost of complications consisted of costs of surgery and the treatment of the two known reaction types in leprosy. This was calculated as follows: unit cost of surgery was estimated to be $ 95 with a probability of need of 4.1% in MB and 1% in PB subgroups. Altogether surgery costs summed up to $ 740 in each treatment arm. The cost of reaction type 1 was $ 8, and the chance to develop such reaction was 31.7% and 2.5% in the MB and PB groups, respectively. The corresponding values for type 2 reaction were $ 71, 2% and 0%. Costs of reaction type 1 amounted to $ 300, whereas type 2 amounted to $ 142. Total cost of standard treatment was estimated to be $ 132 287, whereas chemoprophylaxis intervention needed $ 138 309 investment. The additional cost of the intervention is thus $ 6 022.
The incremental cost effectiveness ratio (ICER) indicates the cost effectiveness of the additional chemoprophylaxis intervention in contacts after 2 years versus the standard MDT treatment for all contacts together and for the three different distance groups. The ICER is expressed in US dollars saved per one prevented leprosy case. In total an incremental of $ 6 009 was invested and 38 incremental leprosy cases were prevented by chemoprophylaxis in contacts on the whole program level, resulting in an ICER of $ 158 (CI: 146–171) per one additional prevented leprosy case (Table 4). Sub-group analyses revealed that chemoprophylaxis was cost-effective for all three contact groups. It was the most cost-effective in neighbours of neighbours and social contacts (ICER $ 214), and slightly less cost-effective in next door neighbours (ICER $ 497) and least cost-effective among household contacts (ICER $ 856). Incorporation of the probabilistic aspect of the complication part into the model did not change the results considerably. The ICERs spread only in a narrow range both at the whole program level and at the sub-group level.
Chemoprophylaxis with single dose rifampicin for preventing leprosy among contacts is a cost-effective prevention strategy. At program level an incremental of $ 6 009 was invested and 38 incremental leprosy cases were prevented, resulting in an ICER of $ 158 per one additional prevented leprosy case.
This is the first report on cost-effectiveness of single dose rifampicin as chemoprophylaxis in contacts of leprosy patients. The analysis is based on the results of a large randomized controlled trial in Bangladesh [6]. For the analysis, the health care perspective was taken because indirect cost data were largely unavailable. The health care perspective excludes indirect costs (patient costs), such as travel costs, loss of income due to illness and clinic visits, and long term consequences of disability. Estimating these costs was beyond the scope of this study, but inclusion would have rendered the intervention even more cost-effective. Another limitation of the study is that a static approach was taken to the analysis, measuring the effect of the intervention after two years only. After these two years, there was no further reduction of new cases in the chemoprophylaxis arm of the trial compared to the placebo arm. Because leprosy is an infectious disease, with person-to-person transmission of M. leprae, one can expect that prevention of primary cases (as recorded in the trial) will lead to further prevention of secondary cases. In time, this would lead to further cost-effectiveness of the intervention. Unfortunately, we could not apply such a dynamic analysis approach because there is insufficient information about the long term effects of the intervention, including the number of secondary cases prevented and the number of primary cases prevented after two years that will eventually develop leprosy after a longer period of time, beyond the 4 years observation period of the trial.
It is also important to understand that the results of the COLEP trial reflect a comparison between the chemoprophylaxis intervention and standard MDT treatment plus contact surveys at 2-year intervals with treatment of newly diagnosed cases among contacts. A contact survey in itself is an intervention that reduces transmission in contact groups and thus new leprosy patients among contacts. The provision of chemoprophylaxis to contacts requires contact tracing, but contact tracing is not part of leprosy control programs in many countries and doing so would increase program costs considerably. WHO however, recognizes the importance of contact tracing and now recommends that it is introduced in all control programs [21]. This would then also lay a good foundation for introducing chemoprophylaxis.
WHO reports regarding cost-effectiveness analyses recommend using disability adjusted life years (DALY) as outcome measure for such studies [22]. In leprosy two measures are common to express disability: WHO grade 1 and 2 [23]. The disability weight for grade 2 disability (visible deformity) has been determined at 0.153 [24], but no weight is available for grade 1. Of all newly detected leprosy cases, a relatively low percentage (2–35%) have grade 2 disability [25]. In our study we chose for the number of leprosy cases prevented as outcome, because there is little information available about survival of patients with grade 2 disability and also because the choice for DALY's would have given a less favourable result due to the low weight of leprosy disability.
There are a number of issues to take into account when relating the outcome of this study to other countries. Firstly, the cost level to conduct leprosy control will differ per country, due to economic standard, budget allocated to primary health care, salaries of health care workers, etc. In our calculation, program costs were similar for both the standard MDT treatment and chemoprophylaxis intervention, but these costs will vary per country. The treatment costs are based on real cost estimates and will vary less between countries and programs. Therefore the actual costs will differ, but the conclusion that the intervention is cost-effective is very likely to remain the same. Secondly, the clinical presentation of leprosy differs between countries and regions. Globally the distribution is around 40% for MB and 60% for PB in newly detected leprosy cases, but with widely varying ratios between countries [25]. Since costs for treating PB and MB leprosy are different, these differences are likely to affect the outcome of the cost-effectiveness analysis. Thirdly, the percentage of newly detected cases that are a household contact of a known leprosy patient differs per country and is possibly determined by the endemicity level of leprosy in a country or area. In Bangladesh, in the high endemic area where the COLEP study was conducted, approximately 25% of newly detected cases had a known index case within the family, whereas in a low endemic area (Thailand) this proportion was 62% [26]. An intervention aimed at close (household) contacts may therefore be more cost-effective in countries where relatively many new cases are household contacts. But the background and implications of such differences on effectiveness of chemoprophylaxis needs further research.
Only few articles have been published about cost-effectiveness analyses of interventions in leprosy [27]. Most articles assess small parts of leprosy control, such as footwear provision [28], MDT delivery costs [29], or the economic aspects of hospitalisation versus ambulatory care of neuritis in leprosy reactions [30]. Only two studies provided a more general cost-effect analysis. Naik and Ganapati included several costs in their economic evaluation, but a limitation of the study is the lack of reference about how they obtained their cost data [31]. Remme et al. based the cost calculations in their study on the limited available published cost data, program expenditure data and expert opinion, and also provide limited insight into how they obtained certain costs and effects [30]. Both studies do not mention well how the costs are obtained, (e.g. real costs, bottom-up or top-down costs). Our current article is basically one of the first structured cost-effective analyses for leprosy presenting an overview of the costs involved and can be used for the assessment of the costs of leprosy control in general.
This report shows that chemoprophylaxis with single dose rifampicin given to contacts of newly diagnosed leprosy patients is a cost-effective intervention strategy. Implementation studies in the field are necessary to establish whether this intervention is acceptable and feasible in other leprosy endemic areas of the world.
|
10.1371/journal.pntd.0001209 | Evidence of Gene Conversion in Genes Encoding the Gal/GalNac Lectin Complex of Entamoeba | The human gut parasite Entamoeba histolytica, uses a lectin complex on its cell surface to bind to mucin and to ligands on the intestinal epithelia. Binding to mucin is necessary for colonisation and binding to intestinal epithelia for invasion, therefore blocking this binding may protect against amoebiasis. Acquired protective immunity raised against the lectin complex should create a selection pressure to change the amino acid sequence of lectin genes in order to avoid future detection. We present evidence that gene conversion has occurred in lineages leading to E. histolytica strain HM1:IMSS and E. dispar strain SAW760. This evolutionary mechanism generates diversity and could contribute to immune evasion by the parasites.
| Gene conversion is a process of recombination that can generate diversity among genes. Gene conversion occurs in some pathogenic species of protozoa to generate diversity among gene families encoding important antigens. The process may contribute to immune evasion by the parasites. Gene conversion, or indeed recombination of any kind, has not previously been demonstrated in human intestinal parasites of the genus Entamoeba. Here, we analysed genes encoding members of an important antigenic protein complex on the surface of Entamoeba parasites which is involved in invasion of the intestinal wall. Three gene families encode heavy-, light- and intermediate-subunits of the complex. We estimated genetic divergence between related genes from two species of Entamoeba, E. histolytica and E. dispar, and compared them to divergence among neighbouring genes and to the average across the whole genome, initially looking for evidence that the genes were evolving under positive selection. However, instead we saw patterns of genetic difference between some of the light- and intermediate-subunit genes indicating the action of gene conversion among members of these gene families. This indicates that recombinational mechanisms may play a part in the molecular evolution of these parasites.
| Entamoeba histolytica causes a significant amount of death and disease, an annual estimate made in the 1980s indicated that 40,000–110,000 people died and 34–50 million people developed severe amoebiasis (dysentery or liver abscess) in 1981 [1]. Infection commonly results from the consumption of contaminated food and water and occurs predominantly among the poor in developing countries. Virulence is a rare outcome of infection, caused by the parasite attacking and crossing the gut wall. It can manifest as dysentery and in some cases as abscesses in the liver and other organs [2]. Most infected people clear their infection within a few months. The related species Entamoeba dispar is not generally believed to cause disease, but rather to live in the gut as a commensal.
A number of genes are implicated in E. histolytica virulence, among them the genes encoding the Gal/GalNAc lectin complex on the parasite's surface. The lectin complex binds galactose and the N-acetyl-D-galactosamine on mucin glycoproteins and on host cell surfaces and mediates both colonisation and contact-dependent cytotoxicity [2]. Anti-lectin immunoglobulin A is associated with protection from amoebiasis [3] and the Gal/GalNAc lectin heavy-chain subunit is a leading vaccine candidate [4]. Immune responses raised against Gal/GalNAc lectin components can protect against virulence [5], [6], although whether this protection is mediated by T-cells or by immuoglobulins is unclear [7], [8]. Immune mediated selection can be a powerful driver of diversity in parasite surface proteins [9]–.
Three components of the Gal/GalNAc lectin complex have been described: the heavy chain subunit, hgl; the light chain subunit, lgl; and the intermediate chain subunit, igl. Each is encoded by a gene family. Heavy- and light-chain lectin subunits are linked by disulphide bonds to form heterodimers. The heavy-chain subunit (hgl) genes contain a transmembrane domain linking a short cytoplasmic and a large extracellular, cysteine-rich domain which appears to mediate binding [12]. Heavy-chain subunit genes show 89–95% amino acid identity. The light-chain subunit is GPI-anchored to the cell membrane. It appears not to mediate adherence but may be associated with virulence, as downregulated virulence is associated with reduced lgl expression [13]. Light-chain subunit genes show more diversity than hgl, with 79–85% amino acid identity among proteins. The intermediate subunit (igl) is GPI-anchored and is non-covalently associated with the other members of the complex [2]. Members of both heavy- and light-chain lectin families have been identified in distantly related Entamoeba species, but igl genes have been identified only in E. histolytica and E. dispar [14], [15].
We reasoned that the difference in virulence between E. histolytica and E. dispar may in part be mediated by adaptive differences in the Gal/GalNAc lectin complex and that immune evasion may drive the evolution of lectin gene families. Therefore, it should be possible to see signatures of this adaptation in the patterns of sequence divergence between the species. However, rather than evidence for positive selection on single nucleotide mutations, we found evidence that gene conversion had occurred among members of the Gal/GalNAc lectin gene families.
Data sets were initially defined using a text search for the term “lectin” on the amoebaDB website (www.amoebadb.org) [16]. Annotated lectin genes were used to query the database of predicted proteins for unannotated gene family members, using BLASTp. Results were assessed by sequence similarity and predicted gene length of the putative lectin. Searches indicated that igl was represented twice in E. histolytica and twice in E. dispar; hgl was represented five times in E. histolytica and twice in E. dispar; lgl was represented seven times in E. histolytica and six times in E. dispar.
The genomic context of each lectin gene was viewed and synteny with E. dispar assessed to define ‘positional orthology’ (i.e. orthology defined by being in the same genomic location). Syntenic genome regions were trimmed such that the lectin was bounded by at least one gene with a putative E. dispar orthologue on each side and aligned using MUSCLE [17] in the SEAVIEW sequence aligment editor [18]. Sequence alignments were checked and manually edited to ensure that nucleotide alignments across coding regions matched the corresponding amino acid alignments. For a small number of genes, the gene models were altered so that both species' gene models matched. The region surrounding the lgl gene EHI_049690 was orthologous to two scaffolds in E. dispar, the ends of which were almost identical when overlapped. A consensus was made of the region, in which the E. dispar genes EDI_071410 and EDI_071300 from scaffold DS548095 were merged with EDI_253210 and EDI_253220 from scaffold DS550857, respectively.
Divergence (d) was estimated across the region for a sliding window (window = 200 bp; step size = 1 bp), using R (Pi was calculated as the number of mismatches per window, rather than per non-gap-site, to reduce spikes caused by very short ‘windows’). In addition, the number of gaps in the alignment (an indicator of alignment quality) was calculated per window. The number of synonymous changes per synonymous site (dS), nonsynonymous changes per nonsynonymous site (dN) and their ratio (dN/dS) were estimated using codeml in the PAML software package [19]. Values were calculated using the maximum likelihood method of Goldman and Yang [20].
To investigate the possibility of gene conversion in the evolution of the lectin gene families, dS values between positional orthologues were compared to a genomic average in order to assess whether they were unusually high. All E. histolytica and E. dispar genes were downloaded from amoebaDB and grouped by orthMCL orthologue group (data from amoebaDB) [21]. To reduce the number of wrongly-aligned non-orthologous genes in the dataset only orthologue groups with exactly one gene from each species were analysed. 4770 orthologue pairs were aligned at the codon level, using PRANK [22] and dS estimated using codeml [19]. To further reduce the effect of misalignments of non-orthologous genes or incorrectly predicted gene models, pairs with extremely high overall divergence (the top 5% of pairwise branch length ‘t’ values from codeml) were removed from the analysis, leaving 4531 orthologue pairs. The frequency distribution of dS values for these gene pairs was plotted in R [23].
Phylogenetic trees were genearted for igl and lgl gene familes. Multiple alignments were generated for each family using MUSCLE [17]. Neighbour-joining phylogenies with bootstrap confidence values were generated using Seqboot, Protdist, Neighbour and Consense programs from the PHYLIP package (http://evolution.genetics.washington.edu/phylip.html), and displayed using the Dendroscope software [24]. A short, possibly truncated, lgl gene (EDI_023210) was removed from the analysis in order to increase the number of sites used to build the phylogeny.
Sequence similarity among members of the igl gene family was assessed. A dotplot was generated for using SEAVIEW [18], in which 40 bp windows were compared across all igl genes and plotted if they showed 100% identity. In addition, a sequence similarity plot was generated across a multiple alignments of the igl genes, using functions from seqinr and base packages of R [23], [25].
Text and BLAST searches of predicted gene sets of E. histolytica and E. dispar in amoebaDB (www.amoebadb.org) defined the members of the heavy-, intermediate- and light-chain subunit gene families hgl, igl and lgl. The hgl family contained E. histolytica genes EHI_042370 , EHI_077500 , EHI_133900 , EHI_012270 and EHI_046650 and E. dispar genes EDI_213670 and EDI_123980. The igl family contained E. histolytica genes EHI_006980 and EHI_065330 and E. dispar genes EDI_276450 and EDI_244250. The lgl family contained E. histolytica genes EHI_049690, EHI_159870, EHI_058330, EHI_148790, EHI_183400, EHI_135690 and EHI_027800 and E. dispar genes EDI_071530, EDI_325130, EDI_131690, EDI_213170, EDI_352500 and EDI_023210.
Visual inspection of aligned scaffolds of E. histolytica and E. dispar in amoebaDB identified six pairs of genes which could be identified as orthologous by surrounding synteny (Figures 1, 2 and Figure S1). These orthologous pairs were: hgl genes EHI_012270:EDI_213670 and EHI_046650:EDI_123980; igl genes EHI_006980:EDI_276450 and EHI_065330:EDI_244250; and lgl genes EHI_049690:EDI_071530 and EHI_159870:EDI_325130. Genomic regions encompassing these genes were aligned (Figures S2, S3, S4, S5, S6, S7) and interspecific diversity estimated across them, under the hypothesis that positive selection in lectin genes should produce a peak of divergence compared to surrounding genes not under such selection. Of the six genes analysed, the two igl genes and an lgl gene (EHI_049690) showed slightly elevated divergence relative to their neighbours. However, on testing for positive selection driving divergence between these genes, the number of synonymous differences per synonymous site (dS) between orthologous genes was particularly high (Figure 1, Figure 2), with dS>1 for two of the genes. We would expect, under positive selection, that the number of nonsynonymous differences per nonsynonymous site (dN) would be high, but that dS would not differ from dS in neighbouring genes. The pattern we observed for the igl genes EHI_006980:EDI_276450 and EHI_065330:EDI_244250 and lgl genes EHI_049690:EDI_071530, where dS was notably higher than for neighbouring genes, suggested a mechanism generating diversity other than positive selection on nonsynonymous mutations. The pattern indicated gene conversion, in which gene regions are changed to match a paralogue that may be more divergent than a gene's true orthologue, hence the elevated dS.
To confirm that dS values between positional orthologues were unusually high, they were compared to a genomic average (Figure 3). The median dS between orthologous genes of E. histolytica and E. dispar was 0.38, and 99% of dS values were below 0.73. The igl genes EHI_006980:EDI_276450 and EHI_065330:EDI_244250 and the lgl genes EHI_049690:EDI_071530 all had dS values in the top 1%, strongly supporting the hypothesis that gene conversion has occurred among these genes.
Phylogenies of multiple alignments of igl and lgl gene familes (Figures S8, S9) further supported gene conversion between igl genes. The expected pattern, given no gene conversion, is that orthologues should cluster together, yet the tree shows strong bootstrap support for the clustering of paralogous igl genes (Figure 4A). In the lgl family, the orthologues EHI_159870 and EDI_325130 cluster together and are quite divergent from other lgl genes (Figure 4B). This pair showed no evidence for gene conversion in Figure 2B and Figure 3. The other pair, EHI_049690 and EDI_071530, occurred in a poorly resolved part of the tree. Amino acid divergence between EDI_071530 and EHI_049690 is 25.2% (dS = 1.17), slightly higher than between EDI_071530 and EHI_035690 (24.8%, dS = 0.98). Between EHI_049690 and EHI_035690, dS was higher (dS = 0.73) than the average between E. histolytica and E. dispar (median dS = 0.38), suggesting that they did not arise from a recent duplication, yet they are more closely related to each other than either is to EDI_071530. Without more information on positional orthology it is difficult to draw any conclusion about the evolutionary history of the genes, other than noting the unexpectedly high synonymous divergence.
To further explore and locate possible gene conversion among the igl genes, we assessed sequence similarity among members of the gene family across the length of the genes. We generated a dotplot showing identical 40 bp windows between genes and calculated sequence similarity across a multiple alignment (Figure S10). Sequence similarity between paralogues is particularly notable at the 3′ end and between bases 1000 and 1500 (Figure 5), covering two of several putative growth factor receptor domains predicted in the genes.
We set out to determine whether members of the Gal/GalNAc lectin complex have evolved under positive selection. However, we were unable to test this due to an unusual pattern of apparent synonymous divergence in some lectin genes. These unusually high values (in some cases >1 synonymous mutation per synonymous site) indicated that the genes of E. histolytica and E. dispar were not in fact orthologous, despite occurring in syntenic genome regions. We showed that dS significantly exceeded the genomic average for igl1, igl2 and an lgl (EHI_049690:EDI_071530). Phylogenetic analysis supported the hypothesis that gene conversion had made paralogous igl genes more similar than orthologous igl genes. In these intermediate-chain lectin genes we saw regions, most notably in the central region and at the 3′ end, where paralogous sequences were highly similar to each other: a pattern expected if gene conversion has occurred.
While it should be noted that the lower sequence coverage of E. dispar, compared to E. histolytica, could result in more errors in its sequence, such errors would inflate the estimated divergence (d) for all genes. Thus, the unusually high dS in lectin genes relative to their neighbours and the the genomic average would not be affected. For the light-chain lectins it is difficult to infer gene conversion from the phylogeny, due to the existence of more gene family members with uncertain orthology. The phylogeny is complicated by the possibility both of incomplete sampling of genes and of recent gene duplications within a species, and by the lack of positional orthology information to compare to the tree. Further sequencing and analysis may help to clarify the picture. In contrast, support for gene conversion among the igl genes was stronger.
Some genes near to lectin genes also showed similarly high dS values. A putative heat shock protein 70 gene (EHI_065320) occurs adjacent to Ehigl2 and shows dS>1. A BLASTp search against E. histolytica and E. dispar protein sequences (data not shown) shows stretches of sequence identity with other genes (EHI_006560 and EDI_169350), suggesting a similar process may have occurred among these genes. A gene encoding an unknown product (EHI_159850) near to a light chain lectin gene (EHI_159870) also shows high dS and a BLASTp search (data not shown) showed high sequence identity with another E. dispar hypothetical protein (EDI_285400). A notable feature of several of the lectin genes is their close proximity to repetitive elements. It is possible that repetitive elements, by creating regions of sequence homology adjacent to non-homologous genes, might promote gene conversion.
Gene conversion is a process of non-reciprocal homologous recombination whereby one region of the genome is ‘converted’ to become identical to another region. Since sequence homology is required, it occurs preferentially among members of multi-gene families. Our results indicate that homologous recombination can occur in E. histolytica. This is significant since the same mechanism is required for sexual reproduction, which has not been demonstrated to occur in E. histolytica despite its genome encoding the necessary genes [26], [27]. Our results do not prove that sexual reproduction (genetic exchange between individuals) occurs.
Although over the long term gene conversion will homogenise sequence, by displacing diversity accumulated during the divergence of paralogous sequences, in the short term it may act as a generator of diversity by creating new haplotypes which may exist alongside ancestral haplotypes in a population. Gene conversion is a mechanism utilised by a number of eukaryotic and prokaryotic pathogen species to generate antigenic diversity within large gene familes and evade the immune response of the host [28], [29]. It has also been identified in a number of smaller gene families in Plasmodium falciparum [30]–[32] and in a large family in Trichomonas vaginalis [33]. The Gal/GalNAc lectin does appear to be a target of protective immunity [3] so it is possible that gene conversion enables the generation of genetic diversity for immune evasion. However, the specific target and vaccine candidate molecule ‘lecA’, a part of the heavy-chain subunit gene EHI_133900, does not have a clear orthologue in E. dispar, so evidence of gene conversion could not be seen in this gene. The hgl genes that were tested did not show clear evidence for gene conversion. It will be interesting to discover whether variant genes arising from gene conversion segregate within E. histolytica and E. dispar populations, to assess the importance of this mechanism in these species.
Genes analysed in this manuscript were: putative heavy chain lectin genes EHI_042370, EHI_077500, EHI_133900, EHI_012270 and EHI_046650 of E. histolytica and EDI_213670 and EDI_123980 of E. dispar; putative intermediate chain lectin genes EHI_006980 and EHI_065330 of E. histolytica and EDI_276450 and EDI_244250 of E. dispar; and light chain lectin genes EHI_049690, EHI_159870, EHI_058330, EHI_148790, EHI_183400, EHI_035690 and EHI_027800 of E. histolytica and EDI_071530, EDI_325130, EDI_131690, EDI_213170, EDI_352500 and EDI_023210 of E. dispar.
|
10.1371/journal.pntd.0007290 | Molecular detection of P. vivax and P. ovale foci of infection in asymptomatic and symptomatic children in Northern Namibia | Knowledge of the foci of Plasmodium species infections is critical for a country with an elimination agenda. Namibia is targeting malaria elimination by 2020. To support decision making regarding targeted intervention, we examined for the first time, the foci of Plasmodium species infections and regional prevalence in northern Namibia, using nested and quantitative polymerase chain reaction (PCR) methods.
We used cross-sectional multi-staged sampling to select 952 children below 9 years old from schools and clinics in seven districts in northern Namibia, to assess the presence of Plasmodium species.
The median participant age was 6 years (25–75%ile 4–8 y). Participants had a median hemoglobin of 12.0 g/dL (25–75%ile 11.1–12.7 g/dL), although 21% of the cohort was anemic, with anemia being severer in the younger population (p<0.002). Most of children with Plasmodium infection were asymptomatic (63.4%), presenting a challenge for elimination. The respective parasite prevalence for Plasmodium falciparum (Pf), Plasmodium vivax (Pv) and Plasmodium ovale curtisi (Po) were (4.41%, 0.84% and 0.31%); with Kavango East and West (10.4%, 6.19%) and Ohangwena (4.5%) having the most prevalence. Pv was localized in Ohangwena, Omusati and Oshana, while Po was found in Kavango. All children with Pv/Pf coinfections in Ohangwena, had previously visited Angola, affirming that perennial migrations are risks for importation of Plasmodium species. The mean hemoglobin was lower in those with Plasmodium infection compared to those without (0.96 g/dL less, 95%CI 0.40–1.52 g/dL less, p = 0.0009) indicating that quasi-endemicity exists in the low transmission setting.
We conclude that Pv and Po species are present in northern Namibia. Additionally, the higher number of asymptomatic infections present challenges to the efforts at elimination for the country. Careful planning, coordination with neighboring Angola and execution of targeted active intervention, will be required for a successful elimination agenda.
| Namibia is a member of the SADC elimination 8 (E8) group with a target to eliminate malaria by 2020. This target stems from years of aggressive interventional strategies that has led to significant reductions in morbidity and mortality. The focus of this strategy is mainly on Plasmodium falciparum as the primary parasite species. Foci of transmission is found in the northern border with Angola and Zambia, which also carries the highest population density. Recently as part of the elimination efforts to predict areas likely to have rebound epidemics, three regions Ohangwena, Kavango and Zambezi were identified. In order to affirm these findings and decision-making process for intervention, we assessed the parasite prevalence in 7 northern regional sites for four Plasmodium species. We identified Pv and Po curtisi parasites in Omusati, Ohangwena and Kavango, as well as a significant number of asymptomatic Pf and Pv infections, part of which may be due to importation from neighboring Angola. As Namibia is targeting elimination by 2020, careful thought and planning will be required to reach the goal.
| Progress in malaria elimination efforts have intensified across the Southern African Development Community (SADC) with the aim of elimination by 2030 for most countries [1], [2], [3], [4]. This means a knowledge of current levels of transmission as part of the malariogenic potential is required. The burden of malaria is still a challenge with sporadic epidemic outbreaks within the region annually, due to continual movements within and between countries [2], [5], [6]. So, the major task for these countries is the increased detection of foci of asymptomatic and symptomatic Plasmodium species infections with sustained surveillance efforts, for targeted intervention and prevention of epidemic outbreaks [7], [8], [9]. We previously reported on the presence of P. vivax asymptomatic infections in Botswana which has impacted the malaria elimination agenda for better [10]. Asymptomatic infections are largely submicroscopic under low transmission settings, so they are not seen either by microscopy or rapid diagnostic test (RDT) and form significant reservoirs for reinfections and transmission of the parasite, making elimination a difficult task [11], [12].
Namibia is a member of the elimination eight group (E8) of SADC, which has targeted malaria elimination by 2020 as part of the vision 2030 agenda to mitigate poverty (WHO, World Malaria Report, 2017). It has a population of 2.2 million over an area of 0.83 km2 [8], [2]. A greater proportion of the population (55%) live in the north, which is traditionally, the main malarious regions in the country [2]. Namibia has achieved significant declines in symptomatic cases of malaria due to effective control methods [8]. However, the continual migration of individuals to and from Angola (a malaria endemic country) in the North, poses a major challenge of importation of parasites, which confounds the efforts at elimination as it creates changes and uncertainties in the disease burden with hidden asymptomatic cases [13], [7]. To assess the risk associated with persistent migrations, continual surveillance to identify the foci and dynamics of asymptomatic and symptomatic Plasmodium species infections in the country is critical for targeted interventional approaches [14]. A previous modelling using the Bayesian Model Based Geostatics (MBG) approach predicted that the three districts most likely to experience rebound or epidemics of malaria in the country were: Ohangwena, Kavango and Caprivi that borders Angola and Zambia in the north [1],[15],[16]. Here we report for the first time on the foci of P.vivax (Pv) and P. ovale curtisi (Po) asymptomatic and symptomatic infections and current levels of P. falciparum (Pf) asymptomatic infections in an active survey in the country using PCR (nested and qPCR) a more sensitive molecular method.
The study sites were selected with the support of the National Malaria Control Program (NMCP) of the Ministry of Health and Social Services (MOHSS) and Ministry of Education, Namibia. The sites were: Kunene, Omusati, Oshikoto, Ohangwena, Kavango West, Kavango East and Zambezi (Caprivi) (Fig 1). The sites are color coded to reflect the malaria incidence Map for these sites in 2015, which had not been updated at the time of sample collection. The respective incidence rates were, more than 5 cases per 1000 people in Kavango East and West, 1–4.99 cases per 1000 people in Zambezi, Ohangwena and Omusati, while Oshikoto and Kunene had less than 1 case per 1000 people. Within each district the Ministry of Education primary school registers and MOHSS Clinic/Health post registries were used in a multi-staged sampling process to randomly select the schools and clinics from which the enrolment was done. Details of the sampling procedure are as previously described [10]. Briefly, in the first stage, districts with known malaria transmission profile based on the recommendations of the NMCP were purposely selected. In the second stage, towns within the district with variable malaria incidence rates were also purposely selected from the NMCP recommendations as shown in Fig 1. This was to ensure that areas of high, moderate, low and sporadic transmissions including those closer to the Angolan border were captured in the sampling process. In the third stage, schools and clinics within each town/village were purposely selected using a two-stage clustering approach based on the population density and adequate cross-sectional representation of the communities to avoid any bias. In the final stage, participants were assigned numbers and enrolled based on informed parental consent and consent of heads of schools and clinics. The total number of samples derived for each district/town was in direct proportion to the estimated population density.
The study was approved by the MOHSS Ethical Committee, Namibia. All parents/guardians provided informed consent on behalf of all participants. Where needed, assent was also obtained from the child before sample collection.
The study enrolled a total of 952 individuals under 9 years old. Subjects enrolled from schools amounted to 591 while 361 were from Clinics and Health posts. Sampling collection was done from September 2016-October 2017 using a multistage sampling strategy as described previously. Fever was defined as subjects with axillary temperature of ≥ 37.2°C at the time of sample collection, while asymptomatic subjects were those without fever (axillary temperature <37.2°C) and without a history of fever in the preceding 72 hours. Sample collections were timed to cover the onset and peak of the malaria transmission season. Prior to the blood sample being drawn, a short questionnaire was administered for travel history and previous malaria illness in the past year, while basic information on age and sex were documented. An aliquot of 1.5–2.5 ml venous blood was collected into EDTA tubes and centrifuged at 3000 rpm for 5 minutes to separate the buffy coat, plasma and red blood cells into separate tubes. These were then stored at -20°C and later transferred to -80°C till analyzed. Hemoglobin (Hb) was measured using Hemocue (Ängelholm, Sweden).
Data were entered in an Excel data sheet and STATA v11.2 (StataCorp, College Station, TX, USA) was used for analysis. Descriptive statistics and appropriate measures of central tendency were provided for relevant demographic covariates. To describe differences between study sub-populations (eg. different regions of residence, presence/absence of Plasmodium infection), continuous covariates were compared using linear regression or the student t-test and categorical variables were compared using logistic regression, the Chi square test, or Fisher’s exact test. For logistic regression analyses, odds ratios (ORs) were provided. For all point estimates, 95% confidence intervals (CIs) were provided. Anemia was defined as Hb < 11.0 g/dL. Statistical significance for all comparisons was set at p<0.05 and adjustment was not done for multiple comparisons in this exploratory study. P-values smaller than 0.001 were reported as p<0.001.
A total of 952 children from 7 districts (Fig 1.) were assessed for Plasmodium species infection. The median participant age was 6 years (25–75%ile 4–8 years). Participants recruited in Kunene (1.60 years younger, 95%ile 0.91–2.29 years younger), Ohangwena (1.24 years younger, 95%CI 0.74–1.74 years younger), and Omusati (1.20 years younger, 95%ile 0.66–1.75 years younger) were all significantly younger than those participants recruited in Kavango East (p<0.001 for all these comparisons). Overall, 52.6% were female. Participants had a median hemoglobin of 12.0 g/dL (25–75%ile 11.1–12.7 g/dL); 21% of the cohort was anemic. There were significant differences in mean participant hemoglobin levels between different regions; those from Kunene (0.72 g/dL less, 95%CI 0.26–1.19 g/dL less, p = 0.002) and Ohangwena (0.65 g/dL less, 95%CI 0.31–0.99 g/dL less, p<0.001) had lower levels than children from Kavango East.
The majority of Plasmodium infections were Pf (n = 41), with 8 Pv infections, 3 Po infections and no Pm infections (Fig 2). Coinfections were common; half of the children with Pv infections were co-infected with Pf and all the children with Po infections had Pf co-infections. Most of children with Plasmodium infection (63.4%) were afebrile. All three children with Pv/Pf coinfections and two with Pf infection in Ohangwena, had previously visited Angola. The mean hemoglobin was lower in those with Plasmodium infection as compared to those who did not (0.96 g/dL less, 95%CI 0.40–1.52 g/dL less, p = 0.0009). There were no differences in the age or gender distributions for those that did and did not have Plasmodium infections. There were clear differences in the proportions of participants infected in the different regions; the highest prevalence was seen in Kavango East (10.4%) and there was a statistically lower prevalence seen in the regions of Kunene (0%, p<0.001), Ohangwena (4.50%, OR 0.40 95%CI 0.17–0.94, p = 0.04), Omusati (2.70%, OR 0.13 95% CI 0.030–0.58, p = 0.007), and Oshana (2.63%, OR 0.17 95%CI 0.039–0.75, p = 0.02). There was no statistical difference between the prevalence of mixed infections in the different regions. Kavango region interestingly, is the most endemic for malaria from the 2015 incidence Map. Pv was localized within Omusati, Ohangwena and Oshana regions, while P ovale was seen in Kavango. The precise number and type of Plasmodium species for each region is presented in Table 1. The overall parasite prevalence was 4.83% with Pv and Po accounting for 0.84% and 0.31% respectively. The two species therefore account for 1.1% of the parasite prevalence.
The study has demonstrated for the first time that the Pv foci of infection in Northern Namibia encompasses three regions: Omusati, Ohangwena and Oshana, while Po focus is in Kavango. The study affirms Kavango as remaining a focus for Pf infections in northern Namibia, as shown in the incidence map in 2015 and epidemic outbreaks in 2016 [7]. The transmission profile of the parasite shows that Pv only infections were all asymptomatic and in a relatively younger population, whereas Pf infections were made up of relatively older children and a high number of asymptomatics, spread across all the regions. A similar pattern of Pv infection was observed in our previous study in Botswana [10]. Reports from Mali, Senegal, Indonesia, Papua New Guinea (PNG), Brazil and others also indicate that asymptomatic infections are commonly observed [22], [23]. Since the population with Pv infections were also anemic compared with those predominantly in Kavango East with Pf infection, the findings reveal some basic differences in the infection dynamics of Pv and Pf in low transmission settings, and in the population in northern Namibia. Pv strictly infects reticulocytes [24], so under anemic conditions where increases in reticulocyte counts occur, Pv infection will be facilitated. In addition, the relapses associated with the hypnozoites stage of Pv infection [25], will increase anti-blood stage immune acquisition more rapidly overtime [26], [27], so that eventually, this will lead to a predominantly asymptomatic population [28], [29]. It is known that all stages of Pv infection can develop in the asymptomatic state [30] to sustain transmission. On the other hand, in Pf infections where hypnozoite stages are absent, if transmission is persistent, individuals acquire immunity with exposure, so the period of immune acquisition is longer and older people become the carriers of asymptomatic infections [22]. The Pf asymptomatic infections point to sustained infections over the years within the population, that has enabled older children to acquire enough immunity to harbor parasites [2]. In a recent MGB modelling method used to predict which regions in Namibia are most likely to experience rebound epidemics, Ohangwena, Kavango and Caprivi (Zambezi) were cited as the major focal areas [7], which is in absolute congruence with the present report. It appears that there is a “quasi-stable” malaria infection scenario within a low transmission setting, where infections although sporadic may be persistent. One should also take cognizance of the fact the there is an inherent genetic variability of parasites for each population that adds to the heterogeneity and so require more tailored interventions, with regards to elimination [31]. The elimination agenda for Namibia has a major challenge of perennial importation of parasite across the border with Angola. This was seen in the present study, with visits to Angola contributing to the Pv infections. The challenge significantly complicates the elimination process [1].
There is a need for active systematic investigations and understanding of the epidemiology of asymptomatic malaria of all species in low transmission settings with an elimination agenda. This can initiate in hotspots and hotpops. Asymptomatic malaria sustains malaria transmission all season and so form a formidable component of transmission as they are not targeted for clearance [32]. This could be a major obstacle for elimination in a scenario where the asymptomatic fraction of the population grows rather than diminish with time, as a result of acquired immunity. Pv and Po hypnozoite forms add to the complexity of dealing with asymptomatic infections with their sequestration in the liver and or bone marrow [33], [22], [34]. So, in low transmission settings, the epidemiology of Plasmodium infections can be heterogeneous, requiring a more thorough active assessment in children towards malaria elimination. It is now no longer valid that Pv infections do not occur in Africa where Duffy antigen negativity is predominant. Several reports indicate that Pv infections occur in Duffy negative individuals in Africa [35], [36]. Clearly, the agenda for malaria elimination should not only be focused on Pf infections but done in parallel with non-falciparum malaria.
We conclude that Pv asymptomatic infections and Po are present in northern Namibia as are asymptomatic infections of Pf. These introduce new paradigms in the elimination agenda for Namibia, that requires careful planning and thought for blocking transmission and aggressive targeting of all populations and species affected.
|
10.1371/journal.pcbi.1000793 | Game Theory of Social Distancing in Response to an Epidemic | Social distancing practices are changes in behavior that prevent disease transmission by reducing contact rates between susceptible individuals and infected individuals who may transmit the disease. Social distancing practices can reduce the severity of an epidemic, but the benefits of social distancing depend on the extent to which it is used by individuals. Individuals are sometimes reluctant to pay the costs inherent in social distancing, and this can limit its effectiveness as a control measure. This paper formulates a differential-game to identify how individuals would best use social distancing and related self-protective behaviors during an epidemic. The epidemic is described by a simple, well-mixed ordinary differential equation model. We use the differential game to study potential value of social distancing as a mitigation measure by calculating the equilibrium behaviors under a variety of cost-functions. Numerical methods are used to calculate the total costs of an epidemic under equilibrium behaviors as a function of the time to mass vaccination, following epidemic identification. The key parameters in the analysis are the basic reproduction number and the baseline efficiency of social distancing. The results show that social distancing is most beneficial to individuals for basic reproduction numbers around 2. In the absence of vaccination or other intervention measures, optimal social distancing never recovers more than 30% of the cost of infection. We also show how the window of opportunity for vaccine development lengthens as the efficiency of social distancing and detection improve.
| One of the easiest ways for people to lower their risk of infection during an epidemic is for them to reduce their rate of contact with infectious individuals. However, the value of such actions depends on how the epidemic progresses. Few analyses of behavior change to date have accounted for how changes in behavior change the epidemic wave. In this paper, I calculate the tradeoff between daily social distancing behavior and reductions in infection risk now and in the future. The subsequent analysis shows that, for the parameters and functional forms studied, social distancing is most useful for moderately transmissible diseases. Social distancing is particularly useful when it is inexpensive and can delay the epidemic until a vaccine becomes widely available. However, the benefits of social distancing are small for highly transmissible diseases when no vaccine is available.
| Epidemics of infectious diseases are a continuing threat to the health of human communities, and one brought to prominence in the public mind by the 2009 pandemic of H1N1 influenza [1]. One of the key questions of public health epidemiology is how individual and community actions can help mitigate and manage the costs of an epidemic. The basic problem I wish to address here is how rational social-distancing practices used by individuals during an epidemic will vary depending on the efficiency of the responses, and how these responses change the epidemic as a whole.
Social distancing is an aspect of human behavior particularly important to epidemiology because of its universality; everybody can reduce their contact rates with other people by changing their behaviors, and reduced human contact reduces the transmission of many diseases. Theoretical work on social distancing has been stimulated by studies of agent-based influenza simulations indicating that small changes in behavior can have large effects on transmission patterns during an epidemic [2]. Further research on agent-based models has argued that social distancing can arrest epidemics if started quickly and maintained for a relatively long period [3]. Compartmental epidemic models have also been used to study social distancing by including states that represent individuals employing specific behaviors. For instance, Hyman and Li [4] formulate and begin the analysis of flu disease transmission in SIR models where some individuals decrease their activity levels following infection. Reluga and Medlock [5] uses this approach to show that while social distancing can resemble immunization, it can generate hysteresis phenomena much more readily than immunization.
Rather than treating behaviors as states, some models treat behaviors as parameters determined by simple functions of the available information. Reluga et al. [6] studies dynamics where contact rates can depend on the perceived disease incidence. Buonomo et al. [7] investigates the impact of information dynamics on the stability of stationary solutions in epidemic models. Chen [8] considers a similar system but allows individuals to learn from a random sample of neighbors. Funk et al. [9] considers the information dynamics associated with social distancing in a network setting by prescribing a reduction in contacts based on proximity to infection. Related work by Epstein et al.[10] explicitly considers the spatial and information dynamics associated in response to an ongoing epidemic.
Building on the ground-breaking work of Fine and Clarkson [11], there has been substantial recent interest in the application of game theory to epidemiology [12]–[17]. The games studied so far have primarily considered steady-state problems, and have not allowed for dynamic strategies. One notable exception to this is the work of Francis [18], which determines the time-dependent game-theoretical solution of a vaccination problem over the course of an epidemic. In another, van Boven et al. [19] studies the optimal use of anti-viral treatment by individuals when they take into account the direct and indirect costs of treatment.
To study the best usage of social distancing, we apply differential-game theory at a population-scale. Differential games are games where strategies have a continuous time-dependence; at each point in time, a player can choose a different action. For instance, a pursuit-game between a target and a pursuer is a two-player differential game where each player's strategies consist of choosing how to move at each successive time until the target is caught by the pursuer or escapes. Geometrically, one might think of differential games as games where strategies are represented by curves instead of points. Two-player differential-game theory was systematically developed by Isaacs [20] as an extension of optimal control theory [21]–[23]. Here, we employ an extension of differential game theory to population games of the form described by Reluga and Galvani [24]. The analysis in this paper will be limited to the simplest case of the Kermack–McKendrick SIR model with strong mixing [25].
In the Model section, we formulate an epidemiological-economics model for an epidemic, accounting for the individual and community costs of both social distancing practices and infection. We then use differential game theory and numerical methods to identify the equilibrium strategies over the course of an epidemic. Numerical methods are used to investigate the finite-time problem where vaccines become available after a fixed interval from the start of the epidemic and the infinite-horizon problem without vaccination. Fundamental results on the value and timing of social distancing are obtained.
In this article, social distancing refers to the adoption of behaviors by individuals in a community that reduce those individuals' risk of becoming infected by limiting their contact with other individuals or reducing the transmission risk during each contact. Typically, social distancing incurs some costs in terms of liberty, social capital, time, convenience, and money, so that people are only likely to adopt these measures when there is a specific incentive to do so. In addition to the personal consequences, the aggregate effects of social distancing form an economic externality, reducing the overall transmission of disease. This externality needs to be accounted for in the determination individuals optimal strategies, but, by definition, depends on the choice of strategy.
To resolve this interdependence, we formulate our analysis as a population game where the payoff to each individual is determined by the individual's behavioral strategy and the average behavioral strategy used by the population as a whole. The model is related to that previously studied by Chen [26]. We will use to represent one specific individual's strategy of daily investment in social distancing. The population strategy is the aggregate daily investment in social distancing by the population. The overbar notation is used to indicate that the aggregate investment should be thought of as an average investment aggregated over all individuals in the population. In the limit of infinitely large populations, and can be thought of as independent because changes in one person's behavior will have little affect on the average behavior. Similarly, the epidemic's dynamics depend on the population strategy but are independent of any one individual's behavior .
The effectiveness of social distancing is represented by a function , which is the relative risk of infection given a daily investment in social distancing practices. If there is no investment, the relative risk . As the daily investment increases, the relative risk decreases, but is bounded below by . We expect diminishing returns with increasing investment, so we will also make the convenient assumption that is convex.
Consider a Susceptible-Infected-Recovered (SIR) epidemic model with susceptible (), infected and infectious (), and removed () states. Suppose an epidemic starts with cases in a community of total individuals (taking ) and proceeds until time , at which point all the individuals in the susceptible state are vaccinated. This epidemic is fast relative to demographic processes and we do not distinguish among the possible states of individuals leaving the infectious state, so the population size can be treated as constant. Between time and time , the dynamics are described by(1a)(1b)(1c)where is the transmission rate and is the removal rate. This SIR model assumes the population is homogeneous, strongly mixed, and that the duration of infections is exponentially distributed. At the start of the epidemic when there are few cases of infection (), the basic reproduction number .
The total cost of the epidemic to the community, , is the sum of the direct costs plus the indirect costs of any economic repercussions from the epidemic. To keep our analysis tightly focused, we will only consider direct costs of the epidemic, including the daily costs from infection, daily investments in social distancing, and the costs of vaccination. Mathematically,(2)where is the daily cost of each infection, is the cost of vaccination per person, and is the discount rate. Note that while the cost of infection is a constant, the investment in social distancing is a function of time. The last term in Eq. (2) is called a salvage term and represents the cumulative costs associated with individuals who are sick at the time the vaccine is made available (). The assumption that the entire remaining susceptible population is vaccinated at time and that vaccination takes effect instantly is, of course, unrealistic, but does provide an approximation to the delayed release of a vaccine.
To simplify our studies, we will work with the dimensionless version of the equations by taking:(3)Under this choice of units, time will be measured in terms of disease generations, social distancing costs will be measured relative to the daily cost of infection, and population sizes will be measured relative to the critical population size necessary to sustain an epidemic.
Epidemics usually start with one or a few index cases, so we focus on scenarios where . The dynamics can be described in terms the shape of , the discount rate , and a single initial-condition parameter(4)From this, it follows that . Since epidemics are often much faster than human demographic processes governing the discount rate [27], we will also take in all calculations. Henceforth, we will drop the hat-notation and work with the dimensionless parameters. The dimensionless equations are(5a)(5b)(5c)with the constraint that . Note that we drop the function notation when necessary to simplify the presentation.
For our further analysis, we will assume(6)with the maximum efficiency of social distancing . Eq. (6) is nicely behaved for numerical solutions because of its relatively fat tail.
We now formulate a differential game for individuals choosing their best social distancing practices relative to the aggregate behavior of the population as a whole. The following game-theoretic analysis combines the ideas of Isaacs [20] and Reluga and Galvani [24]. The premise of the game is that at each point in the epidemic, people can choose to pay a cost associated with social distancing in exchange for a reduction in their risk of infection. The costs of an epidemic to the individual depend on the course of the epidemic and the individual's strategy of social distancing. The probabilities that an individual is in the susceptible, infected, or removed state at time evolve according to the Markov process(7)where is the individual's daily investment as a function of the epidemic's state-variables and the transition-rate matrix(8)Note that both and change over time. Along the lines discussed above, and represent different quantities in our analysis; represents one individual's investment strategy and the population strategy represents an aggregated average of all individual investments. We also note that there are several different ways and can be parameterized. They may be parameterized in terms of time, as and , or in implicit feedback form and , or in explicit feedback form and . The form used will be clear from the context.
Since the events in the individual's life are stochastic, we can not predict the exact time spent in any one state or the precise payoff received at the end of the game. Instead, we calculate expected present values of each state at each time, conditional on the investment in social distancing. The expected present value is average value one expects after accounting for the probabilities of all future events, and discounting future costs relative to immediate costs. The expected present values of each state evolve according to the adjoint equations(9)where . The components , , and represent the expected present values of being in the susceptible, infected, or removed state at time when using strategy in a population using strategy . The expected present values depend on the population strategy through the infection prevalence .
The adjoint equations governing the values of each state are derived from Markov decision process theory. They are(10a)(10b)(10c)with the constraints that for all time . Solution of (10)b and (10)c gives(11)If it is impossible to make a vaccine, the equations must be solved over an infinite horizon. Over an infinite horizon, , assuming becomes constant. In the case of no discounting (), we still have provided for sufficiently large . In the case where a perfect vaccine is universally available at terminal time , the value of the susceptible and removed states differs by the cost of vaccine for . To avoid complications with the choice of whether-or-not to vaccinate, we take so . This is reasonable in scenarios where the cost of the vaccine is covered by the government.
The dynamics are independent of , so we need not consider removed individuals further. Taking and , we need only study the reduced system(12a)(12b)(12c)with boundary conditions(12d)The other conditions must be calculated from the solution of the boundary-value problem and provide useful information. will be the expected total cost of the epidemic to the individual. The final size of the epidemic is given by .
Solving a game refers to the problem of finding the best strategy to play, given that all the other players are also trying to find a best strategy for themselves. In some games, there is a single strategy that minimizes a player's costs no matter what their opponents do, so that strategy can very reasonably be referred to as a solution. In many games, no such strategy exists. Rather, the best strategy depends on the actions of the other players. Any strategy played by one player is potentially vulnerable to a lack of knowledge of the strategies of the other players. In such games, it is most useful to look for strategies that are equilibria, in the sense that every player's strategy is better than the alternatives, given knowledge of their opponent's strategies. A Nash equilibrium solution to a population game like that described by System (12) is a strategy that is a best response, even when everybody else is using the same strategy. i.e. given , is a Nash equilibrium if for every alternative strategy , . A Nash equilibrium strategy is a subgame perfect equilibrium if it is also a Nash equilibrium at every state the system may pass through. I will not address the problem of ruling out finite-time blowup of the Hamilton–Jacobi equation and establishing existence and uniqueness of subgame perfect equilibria. But numerical and analytical analyses strongly support the conjecture that the stategies calculated here are the unique global subgame perfect equilibria to the social distancing game.
The equilibria of System (12) can be calculated using the general methods of Isaacs [20]. The core idea is to implement a greedy-algorithm; at every step in the game, find the investment that maximizes the rate of increase in the individual's expect value . We represent strategies as functions in implicit feedback form. is the amount an individual invests per transmission generation when the system is at state . If is a subgame perfect equilibrium, then it satisfies the maximum principle(13)when everywhere. So long as behaves well, in the sense that it is differentiable, decreasing, and strictly convex, then is uniquely defined by the relations(14)Figure 1 shows the interface in phase space separating the region where the equilibrium strategy will include no investment in social distancing () from the region where the equilibrium strategy requires investment in social distancing ().
Two cases are immediately interesting. The first is the infinite-horizon problem – what is the equilibrium behavior when there is never a vaccine and the epidemic continues on until its natural end? The second is the finite-horizon problem – if a vaccine is introduced at time generations after the start of the epidemic, what is the optimal behavior while waiting for the vaccine? In both of these cases, it is assumed that all players know if and when the vaccine will be available.
The infinite-horizon and finite-horizon problems are distinguished by their boundary conditions. In the finite-horizon case, we assume all susceptible individuals are vaccinated at final time , so , , , while and are unknown. In the limit of the infinite-horizon case (), we solve the two-point boundary value problem with terminal conditions , , and initial conditions , while and are unknown. But these conditions are insufficient to specify the infinite-horizon problem. The plane is a set of stationary solutions to Eq. (12), so we need a second order term to uniquely specify the terminal condition when we are perturbed slightly away from this plane. Using Eq. (12), we can show solutions solve the second-order terminal boundary condition(15)for as .
Most of the equilibria we calculate are obtained numerically. Some exceptions are the special cases where , . Under these conditions, solutions can be obtained in closed-form. First, . While , and(16)When matched to the terminal boundary condition, we find that if we write in feedback form as a function of rather than ,(17)is a solution so long as for all . Inspecting the inequality condition, we find that this holds as long as .
A problem with solving Eq. (12) under Eq. (14) is that it requires to be known from past time and to be known from future time. This is a common feature of boundary-value problems, and is resolved by considering all terminal conditions . Using standard numerical techniques, identifying an equilibrium in the described boundary-value problem reduces to scalar root finding for to match the given . The special form of the population game allows the solution manifold to be calculated directly by integrating backwards in time, rather than requiring iterative approaches like those used for optimal-control problems [23]. Code for these calculations is available from the author on request.
Before presenting the results, it is helpful to develop some intuition for the importance of the maximum efficiency of investments in social distancing. Given for an arbitrary relative risk function , then in the best-case scenarios, where diminishments on returns are weakest, one would have to invest atleast of the cost of infection per disease generation to totally isolate themselves. The units here are derived from dimensional analysis. This could be invested for no more than generations, before one's expenses would exceed the cost of becoming infected. When returns are diminishing, fewer than generations of total isolation are practical. Thus, the dimensionless efficiency can be thought of as an upper bound on the number of transmission generations individuals can afford to isolate themselves before the costs of social distancing outweigh the costs of infection.
For the infinite-horizon problem, an example equilibrium strategy and the corresponding dynamics in the absence of social distancing are shown in Figure 2. We can show that if social distancing is highly inefficient (the maximum efficiency ), then social distancing is a waste of effort, no matter how large . If social distancing is efficient, then there is a threshold value of below which social distancing is still impractical because the expected costs per day to individuals is too small compared to the cost of social distancing, but above which some degree of social distancing is always part of the equilibrium strategic response to the epidemic (Figure 3).
The exact window over which social distancing is used depends on the basic reproduction number, the initial and terminal conditions, and the efficiency of distancing measures. The feedback form of equilibrium strategies, transformed from coordinates to the coordinates of the phase-space is represented with contour plots in Figure 1. Among equilibrium strategies, social distancing is never used until part-way into the epidemic, and ceases before the epidemic fully dies out.
The consequences of social distancing are shown in Figure 4. The per-capita cost of an epidemic is larger for larger basic reproduction numbers. The more efficient social distancing, the more of the epidemic cost can be saved per person. However, the net savings from social distancing reaches a maximum around , and never saves more than % of the cost of the epidemic per person. For larger 's, social distancing is less beneficial.
We can also calculate solutions of the finite-time horizon problem where a vaccine becomes universally available at a fixed time after the detection of disease (Figure 5). If mass vaccination occurs soon enough, active social distancing occurs right up to the date of vaccination. Using numerical calculations of equilibria over finite-time horizons, we find that there is a limited window of opportunity during which mass vaccine can significantly reduce the cost of the epidemic, and that social distancing lengthens this window (Figure 6). The calculations show that increases in either the amount of time before vaccine availability or the basic reproduction number increase the costs of the epidemic. Smaller initial numbers of infections allow longer windows of opportunity. This is as expected because the larger the initial portion of the population infected, the shorter the time it takes the epidemic to run its full course.
Here, I have described the calculations necessary to identify the equilibrium solution of the differential game for social distancing behaviors during an epidemic. The benefits associated with the equilibrium solution can be interpreted as the best outcome of a simple social-distancing policy. We find that the benefits of social distancing are constrained by fundamental properties of epidemic dynamics and the efficiency with which distancing can be accomplished. The efficiency results are most easily summarized in terms of the maximum efficiency , which is the percent reduction in contact rate per percent of infection cost invested per disease generation. As a rule-of-thumb, is an upper bound on the number of transmission generations individuals can isolate before the costs of social distancing outweigh the costs of infection. Social distancing is not practical if this efficiency is small compared to the number of generations in the fastest epidemics (). While social distancing can yield large reductions in transmission rate over short periods of time, optimal social-distancing strategies yield only moderate reductions in the cost of the epidemic.
Our calculations have determined the equilibrium strategies from the perspective of individuals. Alternatively, we could ask what the optimal social distancing practices are from the perspective of minimizing the total cost of the epidemic to the community. Determination of the optimal community strategy leads to a nonlinear optimal control problem that can be studied using standard procedures [23]. Yet, practical bounds on the performance of the optimal community strategy can be obtained without further calculation. The optimal community strategy will cost less than the game-theoretic solution per capita, but must cost more than , as that is the minimum number of people who must become sick to reduce the effective reproduction number below the epidemic threshold. Preliminary calculations indicate that optimal community strategies and game equilibrium strategies converge as grows, and significant differences are only observable for a narrow window of basic reproduction numbers near .
The results presented require a number of caveats. I have, for instance, only considered one particular form for the relative risk function. Most of the analysis has been undertaken in the absence of discounting (), under the assumption that the epidemic will be fast compared to planning horizons. Discounting would diminish importance of long term risks compared to the instant costs of social distancing, and thus should diminish the benefits of social distancing. The benefits of social distancing will also be diminished by incorporation of positive terminal costs of vaccination (). Realistically, mass vaccination cannot be accomplished all-at-once, as we assume. It's much more likely that vaccination will be rolled out continuously as it becomes available. This could be incorporated into our analysis, for instance, by including a time-dependent forcing. Other approaches include extending the model to incorporate vaccination results of Morton and Wickwire [28], or to allow an open market for vaccine purchase [18].
The simple epidemic model is particularly weak in its prediction of the growths of epidemics because it assumes the population is randomly mixed at all times. We know, however, that the contact patterns among individuals are highly structured, with regular temporal, spatial, and social correlations. One consequence of heterogeneous contact structure is that epidemics proceed more slowly than the simple epidemic model naively predicts. Thus, the simple epidemic model is often considered as a worst-case-scenario, when compared with more complex network models [29], [30] and agent-based models [31]–[33]. In the context of social distancing, it is not immediately clear how weaker mixing hypotheses will affect our results. Weakened mixing will prolong an epidemic, increasing the window over which social distancing is needed. But under weakened mixing, individuals may be able to use local information to refine their strategies in ways analogous to the ideas of Funk et al. [9] and Perisic and Bauch [34]. In general, the analysis of aggregate games with stochastic population dynamics require a significant technical leaps, and are the subjects of active research.
One of the fundamental assumptions in our analysis is that there are no cost-neutral behavior changes that can reduce contact rates. In fact, life-experience provides good evidence that many conventional aspects of human behavior are conditional on cultural norms, and that different cultures may adopt alternative conventions. The introduction of a new infectious disease may alter the motivational pressures so that behavioral norms that were previously equivalent are no longer, and that one norm is now preferred to the others. In such cases, there are likely to be switching costs that retard the rapid adoption of the better behaviors that conflict with cultural norms. The rate of behavior change, then, would be limited by the rate of adoption of compensatory changes in cultural norms that reduce the cost of social distancing.
Another deep issue is that behavior changes have externalities beyond influencing disease incidence, but we have not accounted for these externalities. People's daily activities contribute not just to their own well-being but also to the maintenance of our economy and infrastructure. Social distancing behaviors may have serious negative consequences for economic productivity, which might feed back into slowing the distribution of vaccines and increasing daily cost-of-living expenses.
We can extend our analysis to include economic feedbacks by incorporating capital dynamics explicitly. Individuals may accumulate capital resources like food, water, fuel, and prophylactic medicine prior to an epidemic, but these resources will gradually be depleted and might be difficult to replace if social distancing interferes with the economy flow of goods and services. Further capital costs at the community and state scales may augment epidemic valuations. These factors appear to have been instrumental in the recent US debate of school-closure policies. One feature of a model with explicit capital dynamics is the possibility of large economic shocks. This and related topics will be explored in future work.
These calculations raise two important mathematical conjectures which I have not attempted to address. The first is that the social distancing game possesses a unique subgame-perfect Nash equilibrium. There is reasonable numerical evidence of this in cases where the relative risk function is strictly convex, and stronger unpublished arguments of this in cases of piecewise linear . I believe this will also be the case for non-convex but monotone relative risks under some allowances of mixed-strategies. A second conjecture, not yet addressed formally, is that increases in the efficiency of social distancing always lead to greater use of social distancing, all other factors being equal. This seems like common sense, but the precise dependence of Figure 1 on the efficiency has yet to be determined mathematically.
As with all game-theoretic models, human behavior is unlikely to completely agree with our equilibria for many reasons, including incomplete information about the epidemic and vaccine and strong prior beliefs that impede rational responses. On the other hand, our approach is applicable to a large set of related models. We can analyze many more realistic representations of pathogen life-cycles. For instance, arbitrary infection-period distributions and infection rates can be approximated using a linear chain of states or delay-equations [24]. Structured populations with metapopulation-style mixing patterns may also be analyzed. I hope to apply the methods to a wider variety of community-environment interactions in the future.
|
10.1371/journal.pbio.2003864 | Endoplasmic reticulum-plasma membrane contact sites integrate sterol and phospholipid regulation | Tether proteins attach the endoplasmic reticulum (ER) to other cellular membranes, thereby creating contact sites that are proposed to form platforms for regulating lipid homeostasis and facilitating non-vesicular lipid exchange. Sterols are synthesized in the ER and transported by non-vesicular mechanisms to the plasma membrane (PM), where they represent almost half of all PM lipids and contribute critically to the barrier function of the PM. To determine whether contact sites are important for both sterol exchange between the ER and PM and intermembrane regulation of lipid metabolism, we generated Δ-super-tether (Δ-s-tether) yeast cells that lack six previously identified tethering proteins (yeast extended synatotagmin [E-Syt], vesicle-associated membrane protein [VAMP]-associated protein [VAP], and TMEM16-anoctamin homologues) as well as the presumptive tether Ice2. Despite the lack of ER-PM contacts in these cells, ER-PM sterol exchange is robust, indicating that the sterol transport machinery is either absent from or not uniquely located at contact sites. Unexpectedly, we found that the transport of exogenously supplied sterol to the ER occurs more slowly in Δ-s-tether cells than in wild-type (WT) cells. We pinpointed this defect to changes in sterol organization and transbilayer movement within the PM bilayer caused by phospholipid dysregulation, evinced by changes in the abundance and organization of PM lipids. Indeed, deletion of either OSH4, which encodes a sterol/phosphatidylinositol-4-phosphate (PI4P) exchange protein, or SAC1, which encodes a PI4P phosphatase, caused synthetic lethality in Δ-s-tether cells due to disruptions in redundant PI4P and phospholipid regulatory pathways. The growth defect of Δ-s-tether cells was rescued with an artificial "ER-PM staple," a tether assembled from unrelated non-yeast protein domains, indicating that endogenous tether proteins have nonspecific bridging functions. Finally, we discovered that sterols play a role in regulating ER-PM contact site formation. In sterol-depleted cells, levels of the yeast E-Syt tether Tcb3 were induced and ER-PM contact increased dramatically. These results support a model in which ER-PM contact sites provide a nexus for coordinating the complex interrelationship between sterols, sphingolipids, and phospholipids that maintain PM composition and integrity.
| Almost half of the inner surface area of the yeast plasma membrane (PM) is covered with closely associated cortical endoplasmic reticulum (ER). In yeast and human cells, it has been proposed that ER-anchored tether proteins staple the ER to the PM, creating membrane contact sites at which lipid transport between the ER and PM and membrane lipid synthesis are coordinately regulated, but the potential mechanisms are unclear. Here, we test this idea by creating yeast cells that lack all ER-PM tethers. We find that whereas the bidirectional transport of sterols between the ER and PM is unaffected in these cells, sterols within the PM are disorganized due to disruptions in phospholipid biosynthesis that alter PM lipid composition. In particular, we show that phosphatidylinositol-4-phosphate, a phospholipid needed for intracellular signaling and membrane trafficking, accumulates within the PM. Some of these defects can be rescued by reinstating membrane contacts via expression of an artificial tether. However, correction is also achieved without the creation of contacts by supplementing the growth medium with a precursor of membrane phospholipids. Based on these results, we propose that ER-PM contacts do not play a major role as physical conduits for lipid exchange but rather serve as regulatory interfaces to integrate lipid synthesis pathways.
| Most lipids are synthesized in the endoplasmic reticulum (ER) and distributed to other membranes by non-vesicular mechanisms. These mechanisms act in conjunction with lipid metabolic networks to maintain the unique lipid profile of the plasma membrane (PM) and subcellular organelles, and enable rapid membrane lipid remodeling in response to signals and stresses [1–3]. An attractive hypothesis is that non-vesicular lipid transport and lipid biosynthetic and regulatory pathways intersect at ER-PM membrane contact sites (MCSs), where protein tethers retain the ER and PM within about 15–60 nm of each other [4–9]. In this view, ER-PM MCSs would serve as a nexus, coordinating requirements in the PM for lipids with their production in the ER [3, 9]. How this coordination is accomplished is not well understood. Here, we report on the interplay between sterol and phospholipid homeostasis at ER-PM MCSs.
Cholesterol—and its yeast counterpart ergosterol—are synthesized in the ER and transported by non-vesicular mechanisms to the PM [10, 11], where they are found at high concentrations corresponding to about 40 mole percent of PM lipids, i.e., one out of every two to three lipids in the PM is a sterol. The spontaneous exchange of sterols between membranes is slow in vitro and undetectable in vivo, primarily because sterol desorption from the membrane is energetically expensive [12–14]. To move sterols efficiently between the ER, PM, and other membranes, cells make use of sterol transport proteins (STPs), whose proposed function is mainly to reduce the energy barrier for sterol desorption, thereby extracting sterols into a binding pocket within the protein for transit through the cytoplasm [12]. STPs may operate freely in the cytoplasm or at MCSs. Soluble and membrane-bound STPs might work in parallel to provide redundant mechanisms for sterol exchange. As transport is predicted to be rate-limited by the desorption step rather than diffusion of the STP–sterol complex through the cytoplasm [12], the proximity of the ER to the PM at an MCS may not determine the sterol transport rate unless STPs are restricted to these sites. Because a number of sterol biosynthetic enzymes are enriched in PM-associated ER membrane fractions [4], it is attractive to consider that the biosynthetic and transport machineries may colocalize to ER-PM MCSs, effectively channeling sterol between compartments [9] to facilitate sterol homeostasis.
The identity of STPs is controversial and the role of MCSs in sterol transport is unexplored. STP candidates in yeast include members of two protein families: soluble Osh proteins (related to mammalian oxysterol-binding protein [OSBP] [15]) and membrane-bound lipid transfer proteins anchored at MCSs (Lam) (members of the StARkin superfamily of steroidogenic acute regulatory [StAR] protein–related lipid transfer [StART] proteins [16, 17]). Osh4 binds sterols and phosphatidylinositol-4-phosphate (PI4P) [18] and by toggling between its sterol and PI4P bound states, it has been shown to transport sterol against a concentration gradient between vesicle populations in vitro [19]. While this activity may account for certain aspects of sterol homeostasis, Osh4 is not required for retrograde sterol transport [20], nor is it essential for the high rate of sterol exchange between the ER and PM, as evinced by robust sterol transport in cells where all seven OSH genes are inactivated (oshΔ) [21]. The seven Osh proteins share overlapping essential activities [22], but because Osh6 and several other Osh proteins do not bind sterols [23–25], sterol transport is not a function shared by the entire family. Lam proteins each have one or two sterol-binding StARkin domains. The purified domains have been shown to catalyze sterol exchange between vesicles in vitro [26–28]. Lam1–Lam4 are integral ER membrane proteins located at the cell cortex, where they might function as sterol transporters, similar to the mammalian StARkin STARD3, which is anchored to endosomal membranes and has been suggested to facilitate endosome–ER cholesterol transfer [29]. Although elimination of Lam proteins does not inhibit the bidirectional transport of newly synthesized ergosterol between the ER and PM, sterol organization at the PM is altered [30]. If Osh and Lam proteins catalyze ER-PM sterol transport, then they must do so redundantly with each other and/or with additional STPs yet to be identified. Our results address this point.
In addition to their proposed role in sterol homeostasis, ER-PM MCSs are known to be involved in phospholipid biosynthesis and turnover. Phosphatidylcholine (PC) is synthesized via Cho2 and Opi3-mediated methylation of phosphatidylethanolamine (PE). Opi3 is an ER-localized membrane protein in yeast that has been proposed to act in trans at ER-PM MCSs to convert PM-localized PE to PC [31]. In the absence of Opi3 function (either through lack of the enzyme or disruption of ER-PM MCSs), cells rely on the Kennedy pathway, through which PC is synthesized from choline taken up from the growth medium. It has been proposed that phosphoinositide turnover also occurs at ER-PM MCSs, where the ER-localized PI4P phosphatase Sac1 may act in trans to turn over PI4P synthesized in the PM [5]. These examples highlight the possibility that ER-PM MCSs may contribute to a wide range of reactions that underlie cellular phospholipid homeostasis.
In yeast, about 45% of the PM retains a closely associated cortical ER (cER) membrane [5, 32], and this association requires a number of tethering proteins that staple the ER and PM together [5, 31, 33–35]. Six ER-PM tethering proteins are currently known (Fig 1A): the three tricalbins (Tcb1–3), which are yeast homologues of the extended synaptotagmin (E-Syt) family of membrane tethers; Ist2, a member of the TMEM16-anoctamin family of ion channels and phospholipid scramblases; and the yeast vesicle-associated membrane protein (VAMP)-associated protein (VAP) homologues Scs2 and Scs22. Several of these tethers appear to be Ca2+ regulated in mammalian cells, and from their embedded location within the ER membrane, a number of them make contact with the PM through associations with phosphoinositides and/or other phospholipids [6, 8, 36–40]. By eliminating all six of these tethering proteins, Manford and colleagues [5] created Δtether yeast cells, in which large sections of the PM are devoid of cortically associated ER membrane. However, as these authors noted, additional unknown tethers must still exist in Δtether cells, given that small regions of cER were still associated with the PM [5]. Consistent with this conclusion, the localization of Lam1–Lam4 close to the PM near presumptive MCSs is unaffected in Δtether cells [30]. These results suggest that elimination of the six tether proteins is not sufficient to remove all ER-PM contacts and that additional proteins/mechanisms must exist to account for the remaining ER-PM association.
In order to eliminate residual cER in Δtether cells, we focused on Ice2, an integral ER membrane protein (Fig 1A) with established roles in cER inheritance [33, 44], ER-PM contact [31, 33], phospholipid synthesis from stored neutral lipid [31, 45], and ER quality control [46]. Ice2 was first shown to facilitate cER redistribution and inheritance along the PM from mother cells into daughter buds [44]. In cells lacking both ICE2 and SCS2, cER association at the PM is disrupted more than for each single mutant [31, 33]. The defect in ER-PM membrane association in scs2Δice2Δ cells is linked to dysfunctional PC synthesis, likely because the Opi3 methyltransferase is no longer able to act on its PM-localized lipid substrate in trans at contact sites [31]. When cells enter stationary phase, Ice2 has another function, in which it generates a bridge between the ER and lipid droplets [45]. This membrane attachment has been proposed to play a role in channeling droplet-generated diacylglycerol (DAG) to the ER for phospholipid synthesis when cells resume growth [45]. Curiously, the ER-associated degradation (ERAD) substrate carboxypeptidase Y* (CPY*) is stabilized in ice2Δ cells compared with wild-type (WT) cells, pointing to a direct or indirect role for Ice2 in ER-associated degradation [46]. We reasoned that because of its various ER functions, specifically including the generation of ER-PM contacts during mitosis, Ice2 might account for the residual cER in Δtether cells.
If ER-PM contact is necessary for non-vesicular sterol transfer, the rate of ER-PM sterol exchange and/or PM sterol organization would be inhibited by the elimination of all ER-PM MCSs. Likewise, if MCSs serve as regulatory interfaces to coordinate pathways for phospholipid metabolism in the ER and PM, then removing ER-PM MCSs would be predicted to alter cellular phospholipid profiles. We now report that disruption of ICE2 in Δtether cells sharply reduces ER-PM associations to the predicted frequency of randomly finding untethered ER in the vicinity of the PM. The availability of these Δ-super-tether (Δ-s-tether) cells now permits direct tests of hypotheses concerning how ER-PM MCSs impact non-vesicular sterol exchange and inter-membrane lipid regulation.
We now report that the bidirectional movement of sterols between the ER and PM is unaffected in Δ-s-tether cells, indicating clearly that the sterol transfer machinery in yeast is either absent from or not uniquely localized to ER-PM MCSs. Nonetheless, sterol pools within the PM bilayer of Δ-s-tether cells are dramatically altered, and the rate of transbilayer sterol movement within the PM is slowed. We discovered that these defects were associated with changes in the organization and composition of PM lipids and could be largely reversed by supplementing cells with choline or by expressing a nonspecific artificial ER-PM tether. Phospholipid dysregulation in the PM was revealed by changes in the levels of sphingolipids and other PM lipids, as well as by the accumulation of PI4P at the PM of mother Δ-s-tether cells. Interestingly, Δ-s-tether cells were inviable when they also lacked Osh4 or Sac1. After testing the associated roles of Osh6 and the ER-membrane association of Sac1, we conclude that Osh4 and ER-PM MCSs are redundant regulators of PI4P and phospholipid homeostasis. Finally, we found that ER-PM MCS formation is responsive to cellular sterol levels, whereby the tether protein Tcb3 is induced in sterol-depleted cells, resulting in a dramatic increase in membrane association. These results suggest that ER-PM contact sites are dynamic interfaces that adjust and respond to lipid metabolism to maintain PM composition and organization.
Despite the dramatic reduction in ER-PM contact sites caused by the elimination of six tether proteins, the extent of cER in Δtether cells [5] is both significant and heterogeneous, with >35% of the cells possessing fluorescently labeled ER in the vicinity of the PM (Fig 1B and 1C) and individual cells displaying as much as 20% of the average cER found in WT cells (Fig 1D and 1E and S1A Fig). Residual cER in Δtether cells is also evinced by the cortical localization of green fluorescent protein (GFP)-tagged Ysp2/Lam2/Ltc4 (hereafter called Lam2; S2 Fig) and other Lam proteins [30]. These observations suggest that there are additional mechanisms for generating ER-PM association [8]. Because the gene encoding the ER membrane protein Ice2 (Fig 1A) has a negative genetic interaction with SCS2, and Ice2 plays roles in maintaining cER structure and mediating the inheritance of cER from mother cells into daughter buds [33, 44], we hypothesized that Ice2 may contribute to ER-PM association and that its presence in Δtether cells could account for the residual cER seen in these cells.
To explore this possibility, we used confocal fluorescence microscopy to determine the subcellular distribution of Ice2-GFP in Δtether cells (S3 Fig). Ice2-GFP was observed throughout the ER, including cytoplasmic strands radiating from nuclear ER towards the cell periphery (S3A Fig). Notably, Ice2-GFP fluorescence was observed at the cell cortex, visualized by co-expressing the PM marker red fluorescent protein (RFP)-Ras2 (S3A Fig). Optical sections focused at the cell surface showed a concentration of Ice2-GFP fluorescence at remaining ER cortical spots that were visualized using the fluorescent pan-ER marker RFP-ER (dsRED-SCS2220–244) (S3B Fig). This localization pattern resembles that of Scs2 and differs from that of the Tcbs and Ist2 that are restricted to ER-PM MCS spots [5, 35, 42, 47]. These data suggest that Ice2 is correctly localized to contribute to ER-PM tethering. To eliminate residual cER in Δtether cells, we therefore deleted ICE2 in tandem with the other Δtether mutations. We predicted that the resulting Δ-s-tether cells would be largely devoid of ER-PM contact sites and this was indeed the case.
We confirmed the near absence of PM-associated cER in Δ-s-tether cells as follows. First, expression of the pan-ER marker RFP-ER revealed that, unlike WT and Δtether cells, in which cortical fluorescence was observed in 100% and >35% of cells, respectively, fewer than 10% of Δ-s-tether cells had fluorescence at the cell cortex (Fig 1B and 1C). Whereas cortical fluorescence in WT cells and many Δtether cells occurred in the form of linear strands running parallel to the cell perimeter in equatorial views (Fig 1B, arrowheads), the occasional fluorescence seen at the cortex of a small fraction (<10%) of Δ-s-tether cells was in the form of punctae, possibly corresponding to the ends of ER tubules or coincidental positioning of the ER near the PM in the focal plane chosen for imaging (Fig 1B, arrows). Second, whereas GFP-Lam2 is localized exclusively in cortical punctae in WT and Δtether cells (about 15 cortical punctae per cell, on average, for both strains), cortical expression of Lam2 is considerably reduced in Δ-s-tether cells (about four cortical punctae per cell, on average; S2A and S2B Fig), even though the expression level of the protein is unaffected (S2C Fig). Third, cER association along the PM in Δ-s-tether cells was all but absent, as quantified by measuring the cER/PM length ratio in equatorial views of individual cells obtained by transmission electron microscopy (Fig 1D). The average cER/PM ratio was 0.48 and 0.04 in WT and Δtether cells, respectively [5], but only 0.017 in Δ-s-tether cells (Fig 1E). The decrease in the cER/PM ratio in Δ-s-tether versus Δtether cells was statistically significant (Fig 1E, right panel), representing not only an approximately 60% lower average value but also a considerable tightening of the distribution of cER/PM values (Fig 1E and S1A Fig). Finally, we generated 3D models of WT and Δ-s-tether cells by reconstructing images obtained with a focused ion beam–scanning electron microscope (FIB-SEM). These models (Fig 1F) illustrate that the extensive cER coverage of the PM in WT cells is clearly absent in Δ-s-tether cells, where a spaghetti-like accumulation of cytoplasmic tubular ER is observed instead. We estimate that the low amount of cER in Δ-s-tether cells (Fig 1E and S1B Fig) can be accounted for by the random chance of finding untethered ER at the cortex (Materials and methods).
Δ-s-tether cells grow normally on rich media but poorly on minimal media (Fig 2B and 2E), suggesting that the lack of ER-PM contact sites disrupts cell metabolism. If this were indeed the case, then an artificial ER-PM tethering protein might allow the cells to grow normally. Several of the natural ER-PM tethers, e.g., Tcb1–3 and Ist2 (Fig 1A), have a modular architecture, and this design principle was used to assemble an artificial tether ("ER-PM staple") from unrelated non-yeast proteins. As building blocks for the ER-PM staple, we used (i) two ER-anchoring Trans-membrane domains from herpes virus mK3 E3 ubiquitin ligase, (ii) extended helices from mammalian mitofusin 2 to span the gap between the PM and ER, and (iii) the C-terminal polybasic region from mammalian Rit1 (RitC) that targets the PM (Fig 2A). We fused GFP to the N-terminus in order to visualize the ER-PM staple in cells. Expression of the ER-PM staple from the yeast actin promoter largely rescued the growth defect of Δ-s-tether cells cultured on solid medium (Fig 2B), indicating that the artificial staple is a functional substitute for the endogenous tether proteins. Fluorescence microscopy revealed that the ER-PM staple localizes to cER in both WT and Δ-s-tether cells, consistent with the idea that it generates ER-PM contact sites, albeit fewer than endogenous tethers (Fig 2C and 2D). The overall distribution of the ER-PM staples was similar in WT and Δ-s-tether cells, although the staples in Δ-s-tether cells aggregated in larger spots with greater fluorescence. The finding that a wholly heterologous construct can replace endogenous tether proteins in rescuing the poor growth of Δ-s-tether cells indicates importantly that the tethers (Fig 1A) perform a nonspecific bridging function relevant to cell growth that is exclusive of any tether-specific activities. A further conclusion from this result is that the proposed lipid transfer function of the synaptotagmin-like mitochondrial-lipid-binding protein (SMP) domains of the Tcbs (Fig 1A) [48] is not required for cell growth, consistent with observations in HeLa cells and mice lacking E-Syts [49, 50].
Why would the absence of ER-PM tethers cause cells to grow slowly, with growth rescue being achieved by an artificial tether? Tavassoli and colleagues [31] suggested that the ER-anchored phospholipid methyltransferase Opi3 acts at ER-PM contact sites in trans to generate a pool of PC at the PM that is necessary for growth. If Opi3 is unable to act on the PM, as would be expected for cells lacking ER-PM contact sites, then a choline supplement must be provided to generate the necessary PC via the Kennedy pathway [31]. Indeed, we found that Δ-s-tether cells achieve normal growth when the medium is supplemented with choline (Fig 2E). Importantly, the extent of cER was not detectably different in choline-grown Δ-s-tether cells (Fig 2F). We conclude that choline supplementation bypasses the requirement for ER-PM contact sites to support cell growth.
Consistent with these findings, the deletion of either OPI3 or CHO2 in Δ-s-tether cells severely exacerbated the growth defect of the cells unless choline was provided (S4A Fig). However, in the absence of choline, Opi3 overexpression effectively suppressed the choline-dependent growth defect of Δ-s-tether cells (S4B Fig), potentially by providing an alternative route to generate PC pools at the PM. Unlike choline supplementation to the growth medium, the addition of ethanolamine or inositol, which promote PE and PI synthesis, respectively, did not rescue Δ-s-tether growth defects (S5 Fig).
The ability of choline to rescue the poor growth of Δ-s-tether cells, and the functional requirements of Δ-s-tether cells for CHO2 and OPI3, suggested that PC synthesis/levels might be dysregulated in these cells. Surprisingly, whole cell lipidomics (Fig 2G) revealed that PC levels were only about 20% lower in Δ-s-tether cells compared with WT cells, but the relative amounts of a number of other lipids, notably PE, phosphatidylserine (PS), and the yeast sphingolipids inositol-phosphoceramide (IPC) and mannosylinositol phosphoceramide (MIPC), were considerably reduced. Reduced levels of these lipids were also found in Δtether cells that grow almost as well as WT cells, but the lipid compositional effects were generally more pronounced in Δ-s-tether cells. For example, we found the mole percentage of IPC content in Δtether cells to be about 67% of that in WT cells, whereas in Δ-s-tether cells, the level of this lipid fell to about 40% of that in WT cells. Increases in some lipids were also measured, most notably DAG, which was 1.3-fold higher in Δ-s-tether compared with WT cells. These results suggest a possible threshold effect, in which the lipid compositional changes in Δ-s-tether cells have a severe impact on growth, whereas the somewhat lesser changes in Δtether cells do not.
Comparison of the lipid composition of Δ-s-tether cells cultured with or without choline supplementation revealed changes that could be predicted based on the deployment of the Kennedy pathway because of the availability of choline (S6A Fig). Thus, PC and PS levels increased in the choline-supplemented cells, bringing the levels of these lipids closer to those in WT, whereas levels of mono- and dimethyl-PE (mPE and mmPE, respectively) fell. Choline supplementation of Δ-s-tether cells also resulted in an increase in MIPC levels, although other sphingolipids and their precursors were only slightly affected. Unlike lipid compositional changes seen upon choline addition, rescue of Δ-s-tether growth defects by the artificial tether indicated a different mechanism. The lipidomic profile of Δ-s-tether cells expressing the artificial tether was more consistent with a restoration of normal phospholipid synthesis through the cytidine diphosphate diacylglycerol (CDP-DAG) pathway (S6B Fig). The artificial tether increased PS, PI, and mmPE levels, although levels of PC were not appreciably changed from Δ-s-tether cells cultured without choline. The artificial tether also affected storage lipids: esterified ergosterol and triacylglycerol (TG) showed especially large increases as a proportion compared to WT. To our surprise, expression of the artificial tether in Δ-s-tether cells did not restore levels of sphingolipids or their immediate precursors. We conclude that the molecular basis of growth rescue in Δ-s-tether cells by choline and the artificial tether is multifactorial and is likely finely tuned to the precise pools and relative abundance of several lipids, depending on the mode of suppression.
To determine whether ER-PM contact sites play a role in sterol exchange between the two membranes, we compared the rate of retrograde transport of dehydroergosterol (DHE) from the PM to the ER in WT, Δtether, and Δ-s-tether cells using a previously described assay (Fig 3A) [21, 51]. DHE is a fluorescent sterol that is widely used as a reporter of intracellular sterol transport and distribution [52]; it is particularly appropriate as a sterol reporter in yeast cells, as it is closely related to ergosterol and as effective as ergosterol in supporting the growth of hem1Δ cells that cannot synthesize sterols [21]. To load DHE into the PM, the cells are incubated under hypoxic conditions to overcome "aerobic sterol exclusion," which represses endogenous sterol synthesis in favor of exogenous sterol import. When the DHE-loaded cells are transferred to aerobic conditions, ergosterol synthesis resumes and DHE is displaced from the PM. On reaching the ER, DHE becomes esterified by ER-localized sterol acyltransferases. The extent of esterification—detected by the appearance of lipid droplets containing fluorescent DHE or direct measurement of DHE esters by high-performance liquid chromatography (HPLC) analysis of lipid extracts from the cells [21]—provides a measure of retrograde transport.
At the start of the chase period, DHE fluorescence was observed as a “ring stain” in WT, Δtether, and Δ-s-tether cells (Fig 3B), indicating insertion of the fluorescent sterol into the PM [21, 53]. After a 2 h incubation (“chase”) under aerobic conditions, fluorescence was concentrated in lipid droplets in WT and Δtether cells, but the same punctate fluorescence was not observed in Δ-s-tether cells (Fig 3B). To quantify retrograde transport, the amount of imported DHE that was converted into DHE esters was measured at different times following the aerobic chase (Fig 3C). DHE esterification proceeds linearly after a lag period of about 1 h, during which the cells adapt to aerobic conditions, allowing resumption of ergosterol synthesis [21]. Compared to WT or Δtether cells, we observed an approximately 4-fold decrease in the rate of transport-coupled esterification of DHE in Δ-s-tether cells (Fig 3C and 3D). This reduction in esterification rate was not seen in the progenitor strains Δtether or ice2Δ and could be restored to WT levels by expressing the ER-PM staple, or by growing the cells in choline (Fig 3D, S7A and S7B Fig). The latter result (i) suggests that sterol transport between the PM and ER does not depend on ER-PM MCSs, as these structures are equally absent in Δ-s-tether cells grown with or without choline (Fig 2F), and (ii) argues against a recent proposal [54] that the sterol acyl transferases Are1 and Are2 act in trans at ER-PM MCSs, directly receiving sterols from the ATP-binding cassette (ABC) transporters Aus1 and Pdr11, thereby eliminating the need for STP-mediated sterol transport between the PM and ER.
Transport-coupled esterification of DHE is a complex process that can be separated into a series of discrete mechanistic steps (Fig 3A): (1) insertion of DHE into the PM, requiring the ABC transporters Aus1 and Pdr11; (2) equilibration of DHE amongst PM sterol pools, e.g., pools located in the outer and inner leaflets; (3) non-vesicular transport of DHE from the cytoplasmic face of the PM to the ER (3a), a process that requires resumption of ergosterol synthesis (3b) as the cells recover from hypoxia, and transport of ergosterol to the PM (3c); and, finally, (4) esterification of DHE at the ER by the acetyl-CoA acyltransferase (ACAT) enzymes Are1 and Are2. Defects in one or more of these steps could account for the slowdown in DHE esterification seen in Δ-s-tether cells. We verified that DHE loading (step 1) (Fig 3E) and ACAT activity (step 4) (Fig 3F) were similar in WT and Δ-s-tether cells grown in the absence of choline, and the same was true when the cells were grown in the presence of choline (S7 Fig [panels C and D]). However, the level of endogenous ergosterol in Δ-s-tether cells at the start of the aerobic chase was higher than in WT cells on a per cell basis, although it reached the same value at the end of the chase, indicating that ergosterol resynthesis (step 3b) occurs normally (Fig 3G). No difference between ergosterol content and resynthesis was seen when the cells were grown in the presence of choline (S7E Fig). As ergosterol synthesis is largely abolished under hypoxic conditions, the ergosterol content of each cell diminishes with each cell division and is replaced in our protocol by DHE. Because Δ-s-tether cells grow slowly in the absence of choline, the ergosterol “wash-out” is less complete for these cells than for WT cells. The presence of a significant amount of residual ergosterol in Δ-s-tether cells at the start of the aerobic chase could conceivably reduce the rate at which newly synthesized ergosterol is able to displace DHE from the PM, resulting in an apparently slower DHE esterification rate and obscuring information on whether sterol transport between the PM and ER is indeed affected. Thus, our results suggest that the slow rate of esterification observed in Δ-s-tether cells could be due to a defect in steps 2 (DHE equilibration within the PM) and/or 3 (sterol [DHE and ergosterol] exchange between the PM and the ER) (Fig 3A).
To test directly whether sterol exchange between the ER and PM is affected in Δ-s-tether cells, one of the possibilities suggested by our results on retrograde sterol transport (Fig 3C and 3D), we used a pulse-chase protocol (Fig 4A) to compare the rate at which newly synthesized ergosterol is transported from the ER to the PM [14, 21]. The assay was performed as described previously, using [3H]methyl-methionine to pulse-radiolabel ergosterol in the ER [21]. Aliquots of cells taken at different chase time points were homogenized, and the PM was separated from the ER and other internal membranes by sucrose gradient centrifugation. For each time point, the specific radioactivity of [3H]ergosterol (SR = scintillation counts [cpm] ÷ absorbance at 280 nm) was determined for the unfractionated cell homogenate and specific fractions after resolving the corresponding lipid extracts by HPLC, and the relative specific radioactivity for each fraction (RSRfraction = SRfraction ÷ SRcell) was calculated.
Identical subcellular fractionation profiles were obtained with WT and Δ-s-tether homogenates (Fig 4B), displaying clear separation of the PM from internal membranes, as judged by immunoblotting using antibodies against organelle-specific proteins. The quality of the fractionation was exactly as reported in a previous study, in which a wide spectrum of antibodies was used to confirm the separation of the PM from other membranes [21]. The majority of ergosterol was recovered in the PM fraction from WT cells, as expected, and this was also the case for Δ-s-tether cells, indicating that the subcellular distribution of ergosterol is not affected by the absence of ER-PM contact sites (Fig 4B, bottom panel).
We analyzed fractions 7 (PM) and 2 (ER-enriched; we designated this fraction ER* to indicate that it contains other intracellular membranes [Fig 4B]). The results are shown in Fig 4C. For both WT and Δ-s-tether cells, RSR for ER* was high (>2.0) on completion of the labeling pulse because [3H]ergosterol is synthesized in the ER before declining over the chase period to reach a value of 1.0. Conversely, RSR for the PM started at a low level (<0.5; the nonzero value indicates that [3H]ergosterol is transported to the PM even as it is being synthesized during the pulse-labeling period) and increased to 1.0 by the end of the chase. The final RSR values of 1.0 for both fractions indicate equilibration of the ergosterol pulse between the ER and PM, as previously reported [14, 21, 55]. Mono-exponential fits of the data indicate that [3H]ergosterol is exchanged between the ER and PM with a half time of about 10 min for both WT and Δ-s-tether cells. Thus, the exchange of newly synthesized ergosterol between the ER and PM is normal in Δ-s-tether cells.
We considered the possibility that conventional vesicular transport might deliver sterols to the PM to compensate for the possible failure of non-vesicular modes of transport in Δ-s-tether cells. To test this possibility, we constructed a Δ-s-tether sec18-1ts strain that eliminates both exocytosis and ER-PM contact at elevated temperatures. Sec18 is required for exocytosis and most modes of vesicular trafficking [56, 57], and the sec18-1ts conditional mutation blocks vesicular transport of secretory proteins and lipids to the PM [56–58]. Whether on its own or in the context of the Δ-s-tether mutations, the sec18-1ts allele does not allow cells to grow at 37 °C, and Δ-s-tether sec18-1 cells do not even grow at 30 °C (S8 Fig). However, the combined growth defects of the Δ-s-tether mutations and sec18-1 at 30 °C are additive, as would be predicted for unrelated pathways, and do not correspond to a synergistic interaction, as would have been observed between mutations disrupting convergent pathways. After culturing at 23 °C, strains were incubated for 20 min at 37 °C and pulse-labeled with [3H]methyl-methionine followed by a 15 min chase. The calculated RSRs of [3H]ergosterol in PM fractions showed no significant differences between WT, sec18-1, Δ-s-tether, and Δ-s-tether sec18-1 cells, indicating that ergosterol exchange between the ER and PM is unaffected in all of these strains (Fig 4D). These results indicate that secretory vesicles do not provide a compensatory sterol transport mechanism in Δ-s-tether cells.
We conclude that the exchange of newly synthesized ergosterol between the ER and PM does not require ER-PM contact sites. This result has two clear implications. First, the sterol transfer machinery in yeast is either absent from or not uniquely localized to ER-PM MCSs. This suggests that contact site–localized proteins such as Lam1–Lam4 are not essential for sterol exchange between the PM and ER; these proteins may act redundantly with soluble STPs or play other roles in intracellular sterol homeostasis [30]. Second, yeast cells do not possess soluble STPs capable of lowering the energy barrier for sterol desorption by >10 kBT, which would make intracellular sterol transport diffusion-limited rather than desorption-limited [12]. Thus, non-vesicular sterol transport in yeast is likely mediated by cytoplasmic STPs that lower the energy barrier for desorption by a more typical 2–3 kBT [12] and that are present in a sufficient number per cell to account for the measured sterol exchange rate [12].
Osh4 is one subset of Osh proteins capable of binding sterols, and it is present in high levels in yeast at >30,000 copies per cell [59]. Although elimination of Osh4 had no effect on sterol transport, as measured via assays of sterol import [20] or ER-PM sterol exchange (S9 Fig), we tested whether Osh4 might nevertheless provide a compensatory sterol transport mechanism to allow normal ER-PM sterol exchange in Δ-s-tether cells, where the absence of ER-PM contact sites would prevent any putative membrane-bound STPs from reaching their target membrane. Consistent with a possible redundancy between Osh4 and ER-PM contact sites in sterol transport, we discovered that osh4Δ Δtether cells grew poorly and osh4Δ Δ-s-tether cells were inviable (Fig 5A). In contrast, deletion of the putative sterol transporter encoded by LAM2 had no impact on Δ-s-tether cells, whether cultured with or without added choline (S10 Fig). This result is consistent with the fact that the Δ-s-tether mutations compromise Lam2 function by eliminating its proximity to the cell cortex (S2 Fig), and therefore no further effect would be anticipated on eliminating expression of the protein itself. Expression of Scs2 from a plasmid rescued osh4Δ Δ-s-tether cell lethality, but choline supplementation did not (Fig 5A). This result suggests that an ER-PM tether is required to restore viability to these strains; tether-independent, choline-induced phospholipid synthesis is either irrelevant to osh4Δ Δ-s-tether synthetic lethality, or choline supplementation is simply insufficient to overcome the severity of the phospholipid defect.
To test if osh4Δ Δ-s-tether synthetic lethality results from defects in intracellular sterol transport, we generated a conditionally viable strain that combines Δ-s-tether mutations with a temperature-sensitive osh4-1 allele [60]. At 36 °C, osh4-1 osh4Δ Δ-s-tether cells do not grow, and so we measured DHE transport from the PM to the ER 1 h after switching to the inactivating temperature. When OSH4 is inactivated in Δ-s-tether cells in this manner, DHE transfer and esterification were found to be the same as in Δ-s-tether cells (S11 Fig). We conclude that the elimination of OSH4 has no further impact on sterol transport from the PM to the ER in Δ-s-tether cells.
Each of the seven OSH genes can provide the essential requirement for the entire family of OSH genes [22]. Even though they are defined as “OSBP homologues,” not all Osh proteins are able to bind sterols, but all likely bind PI4P [61, 62]. Thus, Osh6 binds PI4P and PS in a mutually exclusive fashion but cannot bind sterols [23, 24]. As another way to determine if the osh4Δ Δ-s-tether synthetic lethality relates to the sterol- versus PI4P-binding activities of Osh4, we therefore tested if Osh6 could functionally replace Osh4 in this context. As shown in Fig 5B, expression of Osh6 from a multicopy plasmid rescued the growth defect of osh4Δ Δ-s-tether cells. Plasmid-based expression of Osh6 was important for growth rescue, as the chromosomally expressed protein, present at fewer than 2,000 copies per cell [59], was not able to support growth. Lipidomics analysis of osh4Δ Δ-s-tether cells rescued with multicopy OSH6 did not reveal an obvious mode of suppression by changes in lipid metabolism (S6C Fig). Compared to Δ-s-tether cells, levels of most sphingolipid precursors and phospholipids (including PS) were unchanged in the OSH6-rescued cells or showed minor reductions. The minor reduction in free ergosterol measured in Δ-s-tether cells was restored to WT levels in OSH6-rescued osh4Δ Δ-s-tether cells, and ergosterol ester levels doubled over WT. These results are consistent with a model in which Osh4 and ER-PM tethers function redundantly and independently, but in an important function revolving around PI4P, with indirect effects on sterol metabolism.
To explore this model further, we used the PI4P marker GFP-PHOsh2 to compare the distribution of PI4P in WT and Δ-s-tether cells. It had been previously reported that PI4P was dysregulated in Δtether cells [5] and we anticipated that this phenotype might be exacerbated in Δ-s-tether cells. WT cells showed PI4P concentrated at the PM only in buds and also localized to the Golgi apparatus (Fig 5C and 5D); this distribution was disrupted in Δ-s-tether cells, in which PI4P was evenly distributed throughout the PM in both mother cells and buds (Fig 5C and 5D). The intensity of PI4P staining in the PM of Δ-s-tether mother cells was greater than that seen for Δtether cells (Fig 5C and 5D). To test if the artificial staple could correct the PI4P accumulation/depolarization phenotype, GFP-PHOsh2 and the ER-PM staple were both expressed in Δ-s-tether cells (S12 Fig). Although at a gross level, the artificial tether did not restore normal PI4P polarization, PI4P was absent at the immediate cortical sites where the staple interacted with the PM, suggesting a potentially local corrective effect. The addition of choline to Δ-s-tether cells had no impact on GFP-PHOsh2 depolarization (100% of Δ-s-tether cells cultured with or without choline had equal GFP-PHOsh2 fluorescence in mother and bud PM, compared to 4.7% of WT cells grown with no added choline and 3.6% of WT cells with choline; n > 104 cells), indicating that ER-PM MCS regulation of PI4P in the PM is distinct from the role of MCSs in PC metabolism. Taken together, our results suggest that the synthetic lethality of the Δ-s-tether mutations with osh4Δ is associated with dysregulation of PI4P homeostasis.
Osh4 and several other Osh proteins have been shown to be upstream regulators of the PI4P phosphatase Sac1, inducing its activity and thereby affecting PI4P levels in several cellular membranes, including the PM [63]. Sac1 is an ER-membrane protein that interacts with most of the ER-PM tethers deleted in the Δ-s-tether strain [5], placing it in a position to act across ER-PM contact sites to dephosphorylate PM-localized PI4P. If Osh4 acts through Sac1 in the same pathway, then sac1Δ might also be synthetically lethal in Δ-s-tether cells. This was indeed the case (Fig 5E), consistent with the model that Osh4 and Sac1 function in a PI4P regulatory pathway operating alongside ER-PM tethers. One possibility is that Sac1 itself provides limited tethering, a function that might be induced by the absence of the other tethers in Δ-s-tether cells. However, expressing the soluble enzymatic domain of Sac1 (Sac11–522) without its ER membrane-binding domain suppressed sac1Δ Δ-s-tether lethality (S13A Fig). This result indicated that Sac1 does not act as a tether and that Sac1 dependence on MCSs can be partially bypassed if Sac1 is released from the membrane, so that it can access PI4P in the PM. Because Osh4 acts in vitro as a soluble PI4P transport protein [64], it might function in cells to extract and transport PI4P from the PM to Sac1 in the ER. We tested if the requirement for Osh4-mediated PI4P transport could be circumvented if Sac1 was freed from the ER membrane to diffuse to the PM to dephosphorylate PI4P. However, soluble Sac11–522 expressed from a multicopy plasmid did not rescue osh4Δ Δ-s-tether lethality, indicating that the requirement for Osh4 cannot be bypassed by liberating Sac1 from the ER (S13B Fig). Thus, in the context of ER-PM MCSs, Osh4 might play an important role in Sac1 regulation, but it clearly has other independent functions as well.
We have shown that the absence of ER-PM contact sites does not affect sterol exchange between the ER and PM (Fig 4C) and that this lack of effect is not due to compensatory sterol transport by secretory vesicles (Fig 4D) or by the single most abundant sterol-binding protein in yeast, Osh4 (Fig 5B). We can therefore pinpoint the cause of slow transport-coupled esterification of exogenously supplied DHE in Δ-s-tether cells to step 2 in the scheme depicted in Fig 3A, i.e., the exchange of sterol between sterol pools within the PM lipid bilayer. To investigate this point, we tested the growth of Δ-s-tether cells in the presence of three drugs that report on the lipid organization of the PM: nystatin, duramycin, and edelfosine.
Nystatin is an ergosterol-binding polyene antimycotic compound. Nystatin resistance is observed in viable sterol biosynthesis mutants and some mutants, such as osh4Δ [51], that disrupt sterol organization within the PM. Conversely, many mutants with altered lipid composition and/or PM organization exhibit nystatin sensitivity [51, 65]. On nystatin-containing medium, Δ-s-tether cells exhibited an exacerbated growth sensitivity compared to Δtether cells, and both strains were more sensitive than WT or nystatin-resistant osh4Δ cells (Fig 6A).
Duramycin is a lantibiotic that disrupts cell growth by directly binding PE in the outer leaflet of the PM. As PE is principally located in the cytoplasmic leaflet of the PM, duramycin sensitivity indicates changes in PE bilayer asymmetry, as seen in the phospholipid-flippase mutant lem3Δ. Growth of WT, Δtether, and Δ-s-tether cells was not significantly affected by duramycin (Fig 6B), indicating that transbilayer phospholipid asymmetry is unaffected. We next tested edelfosine, a cytotoxic lysophosphatidylcholine analogue whose activity in yeast is modulated by PM phospholipid flippase activity, and by sterol and sphingolipid pathways. A flippase defect confers edelfosine resistance [66], whereas changes in the lipid composition and physical properties of the PM confer edelfosine sensitivity [67]. Δ-s-tether cells displayed acute cytotoxicity to edelfosine compared to WT or even Δtether cells (Fig 6B), consistent with changes in PM properties.
Based on the sensitivity of Δ-s-tether cells to edelfosine (Fig 6B) as well as the significant reductions in their sphingolipid levels revealed by lipidomics analyses (Fig 2G), we considered the possibility that the cells would exhibit a growth phenotype in response to the sphingolipid synthesis inhibitor myriocin (Fig 6C). Indeed, previous work had shown that elimination of the three Tcbs alone causes myriocin sensitivity [35]. Unexpectedly, both Δtether and Δ-s-tether cells were myriocin resistant (Fig 6C). These results suggest that myriocin toxicity in Δ-s-tether cells is mitigated by compensatory alterations either in membrane composition or in the sphingolipid biosynthesis apparatus. Taken together, the results of our drug screening experiments indicate that changes in PC and sphingolipid organization in Δ-s-tether cells might indirectly modulate sterol pools within the PM.
The perturbation in PM lipid organization revealed by drug tests (Fig 6A–6C) was not evident in measurements of the ergosterol “status” of the cell (S14 Fig). Thus, when comparing WT and Δ-s-tether cells, we found no significant difference in the fraction of total cellular ergosterol that was recovered in detergent-insoluble membranes (DIMs) (S14F Fig), a crude readout of the extent to which ergosterol associates with phospholipids and sphingolipids containing saturated acyl chains [14, 21]. Likewise, there were no significant differences in the total ergosterol content of the cells (S14A Fig), the ergosterol/phospholipid ratio (S14B Fig), or the fraction of cellular ergosterol located in the PM (Fig 4B). Because these bulk measurements are unlikely to be responsive to nuanced changes in lipid composition and organization, we chose a more sensitive technique to probe ergosterol organization at the PM.
Methyl-β-cyclodextrin (MβCD) extracts only a very small fraction, <0.5%, of total cellular ergosterol from the outer leaflet of the PM of WT cells under our standard conditions [14, 21], indicative of the unusual physical properties of the yeast PM [14, 21, 68–70]. When PM lipid organization is perturbed, then the amount of MβCD-extractable sterol can increase dramatically, as seen as in oshΔ osh4-1ts cells and sphingolipid-deficient lcb1-100ts cells at the nonpermissive temperature [14, 21]. We compared the MβCD-extractability of ergosterol in WT cells versus the tether mutants (Fig 6D). As reported previously, the proportion of ergosterol extracted from WT cells by MβCD is about 0.25% of total cellular ergosterol [14, 21]; a similarly low level of extraction (<1%) was obtained with Δtether cells (Fig 6E). However, in Δ-s-tether cells, the MβCD-accessible ergosterol pool in the PM was >5%, about 20-fold greater than for WT cells, consistent with a major change in the PM lipid bilayer that enabled greater extraction of ergosterol. This effect was largely reset by expression of the ER-PM staple and completely restored to WT levels by supplementing the growth medium with choline (Fig 6E). The ability of both the ER-PM staple and choline to restore PM lipid organization, as revealed by MβCD-extractability of ergosterol, parallels their ability to correct the slowdown in retrograde transport of DHE (Fig 3D). Thus, these results are consistent with the idea that the reduced rate of transport-coupled esterification of DHE is due to perturbations of the PM lipid bilayer that delay the access of exogenously supplied DHE to cytoplasmic STPs (Fig 3D, step 2). The ability of choline to provide the same corrective effect as the ER-PM staple without inducing membrane contacts indicates that the role of tethers in this context is to support normal phospholipid and/or sphingolipid homeostasis, and thereby membrane organization.
Although the exchange of ergosterol between the ER and PM as a whole was unchanged in Δ-s-tether cells (Fig 4C and 4D), we investigated if movement of ergosterol within the PM bilayer might be affected. Non-vesicular transport of newly synthesized [3H]ergosterol deposits ergosterol molecules in the cytoplasmic leaflet of the PM. At a minimum, these molecules must exchange with the outer leaflet pool of ergosterol before they fully equilibrate with PM ergosterol pools and become accessible to MβCD (Fig 6F). We tested if the exchange of ergosterol within the PM was affected in Δ-s-tether cells by measuring the rate at which newly synthesized ergosterol becomes accessible to MβCD extraction. We used [3H]methyl-methionine to pulse-label ergosterol in the ER and then chased the cells for 30 min. The samples were subjected to MβCD extraction and, in parallel, samples were taken for subcellular fractionation to isolate the PM (as in Fig 4B). Both the MβCD extract and the PM fraction were processed with organic solvents to extract ergosterol for HPLC analysis and measurement of RSR. The RSR for the PM fraction after a 30 min chase was about 0.8 for both WT and Δ-s-tether cells (Fig 6G, dashed line), as expected (Fig 4C). However, the RSR for MβCD-extracted ergosterol in WT cells was about 0.65 (Fig 6G, WT), indicating a slight delay in the transport of ergosterol within the PM to the MβCD-accessible pool in the outer leaflet, consistent with our previous report [21]. This delay was considerably greater in Δ-s-tether cells, where the MβCD-extracted ergosterol had an RSR of only about 0.15 after a 30 min chase (Fig 6G, Δ-s-tether). Expression of the ER-PM staple reduced the delay significantly, such that the RSR increased to about 0.3 in Δ-s-tether cells chased for 30 min (Fig 6G, Δ-s-tether + staple). We conclude that (i) the transfer of ergosterol from its site of arrival at the cytoplasmic leaflet of the PM to the outer leaflet pool, from which it can be extracted by MβCD, is slower than the rate at which ergosterol exchanges between the ER and PM as a whole, as reported previously [21], and (ii) the intra-PM movement of ergosterol, from the inner to the outer leaflet, is dramatically slower in Δ-s-tether cells compared with WT cells. Taken together with the fact that the abundance of characteristic PM lipids, e.g., IPC, MIPC, and PS, in Δ-s-tether cells differs significantly from WT cells (Fig 2G), it seems likely that the changes in ergosterol organization in the PM and the rate of exchange between ergosterol pools in the PM are an indirect consequence of changes in PM phospholipid and sphingolipid composition.
We have shown that membrane contact between the ER and PM impacts the abundance of PM lipids (sphingolipids, PE, PS [Fig 2G], and PI4P [Fig 5C]), PM lipid organization (Fig 6A, 6B, 6C and 6E), and the intra-PM movement of ergosterol (Fig 6G). In turn, it is known that PM lipids play a role in the establishment of contact sites (Fig 1A); thus, phosphoinositides and PS in the PM provide anchors for ER-localized Tcb1–Tcb3, Ist2, and Scs2 [6, 8, 36–40]. As sterols represent a large fraction of PM lipids and are critical determinants of PM organization [10, 11], we analyzed the potential role of sterols in establishing contact sites between the ER and PM.
To test the dependence of MCS formation on ergosterol, we depleted yeast cells of sterols and visualized cER-PM association by both transmission electron microscopy and Tcb3-GFP and RFP-ER distribution by fluorescence microscopy. Squalene synthase (Erg9) represents the first sterol-specific enzymatic step in the production of all sterols, and inhibition of Erg9 specifically blocks sterol synthesis without directly affecting other isoprenoids [71]. In erg9Δ PMET3-ERG9 cells, methionine addition to the growth medium represses Erg9 expression and de novo sterol synthesis stops. To our surprise, electron microscopy showed that sterol depletion in erg9Δ PMET3-ERG9 cells resulted in a dramatic expansion of cER (Fig 7A), such that the inner face of the PM was nearly completely covered with associated ER membrane (Fig 7B). This finding was confirmed by confocal fluorescence microscopy in live Tcb3-GFP–expressing cells. In sterol-replete WT cells, Tcb3-GFP fluorescence exhibited a characteristic discontinuous stitched pattern around the cortex (Fig 7C) [35]. In about 90% of sterol-depleted erg9Δ PMET3-ERG9 cells, however, cortical fluorescence was essentially contiguous (Fig 7C). Although sterol-depleted cells accumulate as unbudded cells in the G1-phase of the cell-cycle, G1-arrested cdc42-101 cells did not induce any change in Tcb3-GFP distribution, indicating that increased ER-PM contact is not due to G1 arrest per se (S15 Fig). These results indicate that ER-PM membrane association is induced when cellular sterol synthesis is blocked.
In addition to its altered distribution along the PM, Tcb3-GFP fluorescence was generally greater in sterol-depleted cells relative to WT, suggesting an induction of Tcb3 protein levels in response to sterol reduction. This point was verified by analyzing cell extracts prepared from both WT and erg9Δ PMET3-ERG9 sterol-depleted cells expressing Tcb3-GFP and determining relative levels of Tcb3-GFP by SDS-PAGE/immunoblotting using anti-GFP antibodies. When normalized to levels of the actin (Act1) internal control, Tcb3-GFP protein levels were seen to be induced about 6-fold in sterol-depleted cells, compared to similarly treated WT cells (Fig 7D). In genome-wide analyses of gene expression by DNA microarray, sterol depletion had no impact on transcript levels of any of the tether genes; relative to WT cells, methionine repression of de novo sterol synthesis in erg9Δ PMET3-ERG9 cells showed transcriptional changes between 0.93 and 1.05 ± 0.03 (mean ± SD; independent duplicate trials) for each of the seven tether protein genes. These results indicated that Tcb3 protein levels are posttranscriptionally regulated.
Because of the long half-life of cellular sterols, an extended period is required after ERG9 repression for complete sterol depletion. To determine how quickly sterol-depleted cells recover their normal distribution of ER-PM association, ER-RFP and Tcb3-GFP redistribution was measured in response to exogenously added cholesterol. Under standard culture conditions, yeast does not import sterols from the medium as discussed above (Fig 3), but the deletion of HEM1 permits cholesterol uptake [72]. A hem1Δ erg9Δ PMET3-ERG9 strain could grow after sterol depletion, but only when exogenous cholesterol (or δ-aminolevulinic acid [δ-ALA], the product of the Hem1 enzyme) was supplemented to the growth medium (S16 Fig). In hem1Δ cells, ER-RFP and Tcb3-GFP distributions were the same with or without cholesterol supplementation (Fig 7E). In sterol-depleted hem1Δ erg9Δ PMET3-ERG9 cells, return to the normal discontinuous stitched fluorescence of cortical Tcb3-GFP commenced 1 h after cholesterol addition, and the characteristic WT pattern was observed after about 4 h (Fig 7E and 7F). In these cells, recovery of normal ER-RFP morphology lagged behind the restoration of the normal Tcb3-GFP distribution (Fig 7F), consistent with the idea that tethering complexes dictate changes in cER association. These results indicate that tethering between the ER and PM responds to cellular sterol pools.
MCSs are widely hypothesized to facilitate non-vesicular lipid exchange. We tested this hypothesis in the context of sterol exchange between the ER and PM in yeast by creating Δ-s-tether cells that lack ER-PM membrane contacts. We now report that these contact sites are not required for ER-PM sterol exchange but rather function as regulators of PM lipid homeostasis, controlling the organization and dynamics of sterols within the PM and acting redundantly with the OSBP homologue Osh4 in an essential pathway related to PI4P homeostasis. We also report our unexpected discovery that in the absence of sterol biosynthesis, ER-PM contact sites proliferate, such that the entirety of the PM is associated with ER because of increased expression of tethers. These results invite a revision of current thinking about the role of contact sites as hubs for lipid transport and reveal a reciprocal relationship between the formation and function of contact sites on the one hand and lipid homeostasis on the other.
To create Δ-s-tether cells, we eliminated ICE2 in the previously described Δtether strain. The role of Ice2 in distributing ER along the PM between mother and daughter cells during mitosis is well established [33, 44], hinting that it may play a direct role in tethering ER to the PM. Ice2 is a polytopic ER membrane protein with a single prominent cytoplasmic loop that has been implicated in associating the ER with lipid droplets during the stationary phase of growth, and potentially channeling DAG to the phospholipid biosynthetic machinery in the ER as cells resume growth [45]. In analogy to its proposed tethering role in stationary phase cells, we speculate that Ice2 may play a role in bridging the ER and PM in rapidly dividing cells. Indeed, fluorescence microscopy reveals that Ice2 is located at the cell cortex in Δtether cells (S3 Fig). As a potential tether protein, the cytoplasmic loop of Ice2 may interact directly in trans with the cytosolic face of the PM or, similar to the Scs2 tether [5, 73], Ice2 might form a bridge across the ER-PM interface via an interaction with another protein. If the latter scenario is correct, then the mechanism of tethering by both Ice2 and Scs2 would differ from that of the autonomous membrane attachments conferred by Ist2 and the E-Syt homologues Tcb1–Tcb3 (Fig 1A). Nevertheless, eliminating Ice2 in the context of Δtether cells results in quantifiable reductions in ER-PM association beyond those previously reported for Δtether cells (Fig 1E and S1 Fig), leading to clear functional outcomes. For example, synthetic lethality of Δ-s-tether with osh4Δ or sac1Δ was not manifested in the progenitor Δtether strain and only occurred with the additional deletion of ICE2. Likewise, slowing of transport-coupled esterification of DHE (Fig 3C) and increased extractability of ergosterol by MβCD (Fig 6E) were observed only after deletion of ICE2 in Δtether cells. Taken together, these findings show that Ice2 is an important contributor to ER-PM tethering and associated functions.
We found that bidirectional sterol exchange between the ER and PM occurs at the same rate in Δ-s-tether and WT cells, indicating that ER-PM contact sites do not contribute quantitatively to the mechanism of sterol movement between these two membranes. If any elements of the sterol transport machinery are localized to ER-PM MCSs, then their function must be subsumed by other sterol transport mechanisms in Δ-s-tether cells. However, we show that in and of themselves, neither secretory vesicles (Fig 4D) nor the cytoplasmic sterol-binding protein Osh4 (Fig 5A and 5B) or the ER-anchored Lam2 protein (S10 Fig) provide this putative compensatory mechanism in Δ-s-tether cells. Our results also make it clear that yeast cells do not possess STPs with the ability to lower the energy barrier for sterol desorption to the point at which transport becomes a diffusion-limited rather than desorption-limited process [12]. Thus, non-vesicular sterol transport in yeast is likely mediated by unremarkable cytoplasmic STPs (i.e., STPs that are able to lower the energy barrier for sterol desorption by only 2–3 kBT) present in a sufficient number per cell to account for the measured sterol exchange rate [12]. The identification of these STPs is a focus of future work.
Even though ER-PM sterol exchange was unaffected by the lack of ER-PM contact sites, trafficking of exogenously supplied DHE to the ER was unexpectedly slow in Δ-s-tether cells compared with WT cells (Fig 3C). Careful analysis of the various mechanistic steps of the transport process (Fig 3A) revealed that the slowdown could be linked to a dramatic lowering of the rate at which sterols equilibrate between the inner and outer leaflet of the PM in Δ-s-tether cells (Fig 6F and 6G). While this slowdown should not affect the equilibration of DHE with PM sterol pools during the extended hypoxic loading period used for this assay (Fig 3A, steps 1 and 2), it would affect the rate at which newly synthesized ergosterol displaces DHE from the PM during the aerobic chase, thereby resulting in a slower rate of transport-coupled DHE esterification. We previously reported that the appearance of newly synthesized ergosterol in the MβCD-extractable ergosterol pool in the outer leaflet of the PM lags behind its arrival at the PM in WT cells (Fig 6D and 6E), suggesting that equilibration of sterol across the yeast PM is considerably slower than that seen for cholesterol flip-flop in synthetic, liquid crystalline membranes and in red blood cells [74–76]. This may be a consequence of the unusual properties of the yeast PM, exemplified by the slow lateral diffusion of both lipids and proteins [69, 70] and the organization of PM proteins into a mosaic of domains [77]. In the case of Δ-s-tether cells, the rate of transbilayer sterol equilibration was about 5-fold slower than for WT cells (Fig 6G), and this was reflected in changes in PM bilayer organization, as manifested in the nystatin and edelfosine sensitivity of these cells and the greater accessibility of sterols to MβCD extraction (Fig 6A, 6B and 6E). Quantification of cellular lipids revealed reductions in PE, PS, and the sphingolipids IPC and MIPC (Fig 2G). As these lipids generally reflect PM composition [78], we propose that relative changes in lipid levels are the underlying cause of the disturbance in the PM bilayer, resulting in a change in ergosterol dynamics.
How do ER-PM contact sites affect PM lipid organization and intra-PM ergosterol dynamics? Because the growth defect of Δ-s-tether cells was rescued by choline supplementation (Fig 2E), which increases flux through phospholipid biosynthetic pathways [79] without inducing contact site formation (Fig 2F), the primary defect in cells lacking ER-PM contact sites appears to involve phospholipid regulation. Remarkably, both choline and the artificial staple corrected most defects inherent to Δ-s-tether cells (i.e., slow growth [Fig 2B], slow transport-coupled DHE esterification [Fig 3D], high MβCD-extractability of ergosterol [Fig 6E], and slow rate of ergosterol exchange between PM pools [Fig 6G]). These results indicate that the function of the endogenous tethers, even those such as Tcb1, Tcb2, and Tcb3, with lipid-transporting SMP domains [48] might be largely structural in this context, i.e., the tethers provide a means of mechanical attachment of the ER to the PM, thereby enabling ER-localized proteins, such as the PC-synthesizing phospholipid methyltransferase Opi3, to act in trans.
An exception to this general conclusion is that neither choline nor the artificial staple was able to re-establish normal PI4P polarization in Δ-s-tether cells (Fig 5C and 5D and S12 Fig). PI4P dephosphorylation at the PM is proposed to be due to the ER-localized Sac1 phosphatase acting in trans at ER-PM contact sites [80], although the protein clearly also acts in cis [81]. The inability of the artificial staple to facilitate PM access for Sac1 might stem from either (i) an insufficient number or improper positioning of membrane contacts established by the staple and/or (ii) the inability of the staple to provide a specific requirement for Sac1 activation, which is otherwise provided by endogenous tethers (almost all tethers were identified by virtue of specific interactions with Sac1 [5], whereas the artificial tether would lack this ability). Our data suggest that PM PI4P is reduced in the immediate vicinity of contact sites generated by the artificial staple (S12B Fig), indicating that Sac1 might access PI4P locally at those points. In contrast, endogenous tethers and their ancillary factors might further expand cER-associated regions of the PM that are accessible to Sac1.
In yeast, all phospholipids are synthesized from phosphatidic acid (PA) via the CDP-DAG pathway; PE and PC are also synthesized by the Kennedy pathway using DAG and salvaged or exogenously supplied ethanolamine and choline [79, 82] (S17 Fig). PA levels and the DAG:PA ratio are critical in determining the amount of CDP-DAG available for phospholipid synthesis (S17 Fig). The mole percentage of PA in Δ-s-tether cells is 20% lower than that in WT cells, and the DAG:PA ratio is 2-fold greater (Fig 2G), indicating dysregulation of phospholipid synthesis in the absence of ER-PM contact sites. Consistent with this, levels of PS and PI-derived sphingolipids are much lower in Δ-s-tether than in WT cells (Fig 2G), although PI levels are unaffected (see below). As previously shown [31], ice2Δ scs2Δ cells have a diminished ability to convert PS to PC via Opi3-mediated phospholipid methylation at ER-PM contact sites, necessitating choline supplementation for normal growth. In Δ-s-tether cells, choline supplementation would not only bypass the need for Opi3 but also compensate for the lower overall rate of phospholipid synthesis resulting from decreased PA levels by generating PC via the Kennedy pathway for membrane growth. Lipidomic analysis of Δ-s-tether cells cultured with choline did in fact show restoration of PC to levels comparable to WT (S6A Fig). These results are also consistent with the choline-reversible synthetic growth defects observed when either OPI3 or CHO2 is deleted in Δ-s-tether cells (S4A Fig). Thus, ER-PM contact sites may function as regulatory interfaces that coordinate the CDP-DAG and Kennedy pathways to balance convergent mechanisms for phospholipid synthesis.
Δ-s-tether cells have a normal mole percentage of PI and increased PI4P, yet their content of inositol sphingolipids is about 60% lower than in WT cells (Fig 2G), indicating dysregulation of phosphoinositide homeostasis because of loss of ER-PM contacts. Total cellular PI is generated predominantly by the CDP-DAG pathway and Sac1-mediated dephosphorylation of PI4P (itself generated from PI) in separate cellular locations, with both routes providing the biosynthetic precursor for complex inositol sphingolipids (S17 Fig). Brice and colleagues [83] reported that disruption of SAC1 alone reduces PI levels dramatically and IPC and MIPC levels by >70%. The Δ-s-tether mutations would reduce the Sac1-mediated route for PI production by distancing the enzyme from its substrate in the PM (Fig 5C). However, sac1Δ and Δ-s-tether mutations are lethal when combined, likely due to limiting PI levels, indicating that Δ-s-tether mutations must also inhibit PI production from the CDP-DAG pathway. Normal PI levels in Δ-s-tether cells may therefore be a consequence of preserving this lipid at the expense of inositol sphingolipid production. It should be noted that Sac1 is also a component of the SPOTS (SPT, Orm1/2, Tsc3, and Sac1) complex that regulates early steps in sphingolipid synthesis [84, 85]. However, levels of ceramide, an early precursor in sphingolipid synthesis, were essentially normal in Δ-s-tether cells (Fig 2G). It therefore seems unlikely that the SPOTS complex plays a direct role in ER-PM contact site regulation of inositol sphingolipids.
While investigating whether normal ER-PM sterol transport in Δ-s-tether cells could be due to compensatory activity of soluble or membrane-bound STPs, we discovered that the deletion of OSH4 was synthetically lethal with Δ-s-tether mutations (Fig 5A). Lethality was not due to a sterol-related process because Osh6, which does not bind sterols, could rescue osh4Δ Δ-s-tether growth defects (Fig 5B). Consistent with previous proposals that Osh proteins represent important regulators of PI4P [15, 80], the deletion of SAC1 in Δ-s-tether cells also resulted in synthetic lethality (Fig 5E). Based on these findings, we propose that Osh4 (and Sac1) functions in a parallel pathway alongside ER-PM contact sites for PI4P regulation. However, expression of the soluble enzymatic domain of Sac1 did not suppress osh4Δ Δ-s-tether lethality (S13 Fig), suggesting that the downstream regulation of Sac1 is not the only role Osh4 plays at ER-PM MCSs. The availability of Δ-s-tether cells now allows further interrogation of the mechanism by which ER-PM MCSs function as interfaces for regulating PI4P signaling and phospholipid metabolism.
Our results are consistent with the hypothesis that ER-PM contact sites constitute a regulatory nexus to balance sterol and phospholipid concentrations to maintain PM structure. In the absence of contact sites, the ratio of sterols to specific phospholipids and sphingolipids is uncoupled, which negatively impacts PM organization, intra-PM sterol dynamics, and PI4P levels. When sterols become limiting, posttranscriptional induction of Tcb3 increases the extent of ER-PM association, potentially facilitating compensatory changes to phospholipid synthesis to re-establish bilayer stability. Although the mechanisms that control Tcb3 levels in sterol-replete or sterol-depleted cells are not known, it is interesting to speculate that the induction of a tether protein may represent a new homeostatic mechanism for regulating PM composition and structure.
Yeast strains and plasmids are listed in the Supporting information (S1 and S2 Tables, respectively). Unless otherwise stated, yeast cultures were grown in synthetic complete or YPD rich media at 30 °C. All temperature-sensitive alleles were cultured at permissive growth temperatures (30 °C unless otherwise stated) and shifted to the restrictive temperature of 37 °C, as specified. DNA cloning and bacterial and yeast transformations were carried out using standard techniques [86, 87].
For the choline and ethanolamine supplementation growth assays, yeast strains were cultured in synthetic minimal media for 48 h, then streaked onto solid synthetic complete media containing 1 mM choline chloride or 1 mM ethanolamine (Sigma-Aldrich Chemicals, St. Louis, MO). Growth in response to inositol supplementation was tested on synthetic minimal media containing 75 μM myo-inositol (Sigma-Aldrich Chemicals). For the nystatin sensitivity plate assay, 10-fold serial dilutions of yeast cultures were spotted onto solid synthetic media containing 2.5 μM nystatin (Sigma-Aldrich Chemicals). Cell growth was also tested on solid rich media containing 60 μM edelfosine (Cayman Chemical, Ann Arbor, MI), 5 μM duramycin (Sigma-Aldrich Chemicals), or 0.5 μg/mL myriocin (Sigma-Aldrich Chemicals). To select against URA3-marked plasmids (e.g., pCB1183), yeast cultures were grown on rich growth media and then streaked onto a synthetic solid growth medium containing 1 g/L 5-fluoroorotic acid (Gold Biotechnology, St. Louis, MO). For assays of cholesterol uptake by cells lacking HEM1, yeast cultures were spotted onto a solid synthetic medium lacking methionine, containing 25 μg/mL cholesterol in 1% (vol/vol) Tween 80–ethanol (1:1 [vol/vol]) and 50 μg/mL δ-ALA (Sigma-Aldrich Chemicals). For sterol depletion assays, erg9Δ PMET3-ERG9 cells were grown at 30 °C for 10 h in synthetic media lacking methionine and grown to mid-log phase before adding 100 mg/L methionine.
DNA cloning and bacterial and yeast transformations were carried out using standard techniques [86, 87]. The functional artificial tether fusion plasmids pCB1185 and pCB1188 were derived from pRS416-PYSP1-eGFP-Myc-HMH-RitC, a kind gift from Tim Levine (UCL Institute of Ophthalmology). To construct pCB1185, coding sequences from pRS416-PYSP1-eGFP-Myc-HMH-RitC were amplified using CACTCGAGTTATGGAGCAAAAGCTCATTTCTGAAGAG and CAGGTACCCTATACTGAATCCTTTTTCTTACGGAAT primers, and the product was digested with XhoI/KpnI for subcloning in frame with GFP under the control of an ACT1 promoter in a YCplac111 vector. To construct pCB1188, coding sequences from pRS416-PYSP1-eGFP-Myc-HMH-RitC were amplified using the primers: CATCCGGACTTATGGAGCAAAAGCTCATTTCTGAAGAG and CATCTAGACTATACTGAATCCTTTTTCTTACGGAATGG. The amplified product was digested with KpnI/XbaI and subcloned in frame with coding sequences for mCherry under the control of an ACT1 promoter in a YCplac111 vector.
All genomic manipulations were performed by integration of PCR amplified product as previously described [88]. All natMX4 and hphMX4 deletions were generated by homologous recombination into the yeast genome of targeted P4339 and pAG32 amplified products; transformants were selected for on YPD media containing 100 mg/L nourseothricin (Gold Biotechnology) and 400 mg/L hygromycin B (Toku-E, Bellingham, WA), respectively. For growth of hem1Δ::natMX4 cells, selective growth media contained 50 μg/mL δ-ALA. The sec18-1:URA3 temperature-sensitive allele strains were isolated at 23 °C after genomic recombination of the sec18-1:URA3 gene cassette amplified from CBY2853 genomic DNA. All transformants were confirmed by genomic PCR or genetic complementation assays.
Yeast cells were grown to mid-log phase and prepared (fixation, dehydration, infiltration/embedding) as previously described [89]. Minor changes were made to the infiltration schedule as follows: ethanol:resin (2:1) was incubated overnight while ethanol:resin (1:1) was incubated for 5 h. For calculations of cER abundance in electron micrographs, the ratio between PM and the length of PM associated with cER were determined using ImageJ (www.imagej.nih.gov/ij/index.html); cER was assigned as previously described [90].
The resin block was microtomed to expose a clean face and then attached to a metal SEM stub with carbon tape. The sides were coated with silver paint to increase conductivity. The block was then sputter coated with a thin coat of Au/Pd and inserted into the FIB-SEM. Areas of interest were identified by viewing with the electron beam at 25 keV. Serial block-face imaging: the area of interested was coated with 1 μm-thick Pt in the microscope using the Pt deposition needle. The sample was tilted to 52 degrees and an approximately 30 μm trench was cut in front of the area. The FEI Slice and View G2 program was used for data collection, with the following parameters: imaging at 2 keV; 50 pA current; 30 μs dwell time; horizontal field width, 17.74 μm; tilt angle, 60 degrees (cross-sectional viewing angle, −30 degrees); working distance, 2.5 mm; and TLD detector set to −245 V suction tube voltage for backscatter imaging. The slice thickness was set at 20 nm, so the final voxel size was 8.66 nm in X, 10 nm in Y, and 20 nm in Z. Image processing: raw images were aligned using the xfalign tool in IMOD [91]. Images were then corrected for density gradients using ImageJ software [92]. The aligned and corrected tiff images were imported into Amira for density-guided segmentation (FEI Software, Hillsboro, OR) and display.
Confocal fluorescence microscopy was performed as previously described [93]. For all experiments, yeast cells were grown to mid-log phase before visualization. GFP-Staple (pCB1185) and mCherry-Staple (pCB1188) fusion proteins were imaged using 150 and 750 ms exposures, respectively. RFP-ER (pCB1024 and pCB1277) and RFP-RAS2 (pCB1204) were imaged using a 750 ms exposure on the confocal. GFP-2xPHOSH2 (pTL511) was imaged by confocal microscopy using a 250 ms exposure. Tcb3p-GFP and Ice2p-GFP were imaged by confocal microscopy using 350 ms and 1.5 s exposures, respectively.
Widefield fluorescence microscopy was performed as previously described [94]. GFP-2xPHOSH2 (pTL511) imaged by widefield epifluorescence was acquired using a 200 ms exposure, 30% arc lamp intensity, and analog gain set to full. Bleed-through between fluorescence channels was undetectable under the conditions used for image acquisition. All contrast enhancement was kept constant for each series of images.
Transport of exogenously supplied DHE from the PM to the ER was determined as previously described [21, 51]. Briefly, DHE was loaded into the PM of cells under hypoxic conditions, and its transport to the ER upon subsequent aerobic chase was monitored by fluorescence microscopy and quantified by lipid extraction and HPLC to determine the extent of conversion to DHE-ester. To quantify the initial fluorescence of DHE-loaded cells, individual cells were outlined using ImageJ; then, the corresponding area and integrated density were measured to determine the corrected total cell fluorescence (CTCF), as previously described [95]; CTCF = (integrated density − (area of selected cell × mean background fluorescence)). At least 40 cells were counted (from 4 individual fields) for each strain. The ACAT activity of the cells was assayed with microsomes using a modification of a published procedure [96], as described [21, 51].
Biosynthetic sterol transport was measured using a pulse-chase labeling procedure as previously described [21, 51]. Briefly, cells were labeled with [3H]methyl-methionine for 4 min to generate a pulse of [3H]ergosterol in the ER and subsequently chased with unlabeled methionine. Transport was assessed after subcellular fractionation to isolate the PM or after MβCD extraction to sample ergosterol in the outer leaflet of the PM. Ergosterol in cells, subcellular fractions, and MβCD extracts were solubilized using organic solvents and quantified by HPLC.
For lipidomics analysis, cells were grown to about OD600 0.8 and lipids were extracted with chloroform:methanol (2:1). Yeast lipid extracts were prepared using a standard chloroform-methanol mixture, spiked with appropriate internal standards, and analyzed using a 6490 Triple Quadrupole LC/MS system (Agilent Technologies, Santa Clara, CA) [97]. Glycerophospholipids and sphingolipids were separated with normal-phase HPLC as described before [97], with a few changes. An Agilent Zorbax Rx-Sil column (inner diameter 2.1 × 100 mm) was used under the following conditions: mobile phase A (chloroform:methanol:1 M ammonium hydroxide, 89.9:10:0.1, v/v) and mobile phase B (chloroform:methanol:water:ammonium hydroxide, 55:39.9:5:0.1, v/v); 95% A for 2 min, linear gradient to 30% A over 18 min and held for 3 min, and linear gradient to 95% A over 2 min and held for 6 min. Sterols and glycerolipids were separated with reverse-phase HPLC using an isocratic mobile phase as before [97], except with an Agilent Zorbax Eclipse XDB-C18 column (4.6 × 100 mm).
Quantification of lipid species was accomplished using multiple reaction monitoring (MRM) transitions [97, 98] in conjunction with the referencing of appropriate internal standards: PA 17:0/14:1, PC 17:0/20:4, PE 17:0/14:1, PG 17:0/20:4, PI 17:0/20:4, PS 17:0/14:1, LPC 17:0, LPE 14:0, Cer d18:1/17:0, D7-cholesterol, cholesteryl ester (CE) 17:0, 4ME 16:0 diether DG, D5-TG 16:0/18:0/16:0 (Avanti Polar Lipids, Alabaster, AL). Quality and batch controls [99] were included to assess instrument stability and reproducibility and allow for correction of drift and other systematic noise, e.g., biases correlated with analysis order and/or sample preparation. Values are represented as mole fraction with respect to total lipid (mole percentage) [97]. All lipid species and subclasses were analyzed with one-way ANOVA followed by a post hoc Bonferroni test.
For analysis of Tcb3 protein expression, 10 OD600 units of Tcb3p-GFP–expressing cells post sterol depletion were prepared as described by Ohashi and colleagues [100]. Pellets were resuspended in SDS sample buffer and boiled for 5 min before SDS-PAGE. Protein transfer to nitrocellulose membranes and immunoblot conditions were as previously described [101]. To detect Tcb3p-GFP, immunoblots were incubated with a 1:1,000 anti-GFP antibody (ThermoFisher Scientific Inc., Waltham, MA) followed with 1:10,000 anti-rabbit-HRP secondary antibody (Bio-Rad Laboratories, Mississauga, ON). Actin was detected using 1:1,000 anti-actin antibody (Cedarlane, Burlington, ON) followed with 1:10,000 anti-mouse-HRP secondary antibody (ThermoFisher Scientific Inc.).
For analysis of Ysp2 protein expression, 10 OD600 units of GFP-Ysp2–expressing cells were prepared and proteins extracted as above. To detect GFP-Ysp2, immunoblots were incubated with 1:2,000 anti-GFP antibody (Sigma-Aldrich Chemicals) followed by 1:10,000 anti-rabbit-HRP secondary antibody (Promega, Madison, WI). GAPDH was detected using 1:10,000 anti-GAPDH antibody (ThermoFisher Scientific Inc.) followed with 1:10,000 anti-mouse-HRP secondary antibody (Promega).
We derive a rough estimate of the chance of finding cER at the cell cortex in yeast cells that lack the ability to tether the ER to the PM as follows. Assuming that a yeast cell has a volume of 65 μm3 (radius = 2.5 μm) [102, 103], we estimate the volume of a cortical shell defined by the reported distance (30 nm) at which the ER is retained at the PM by tethers as 1.9 μm3. As about 45% of the PM is associated with ER in WT cells (Fig 1E and references [5, 32]), the volume of the cortical shell that is occupied by ER is 0.9 μm3. If this amount of ER were to become untethered, then it could be found anywhere in the total volume of the cell. Approximately 65% of the total cell volume is available for this purpose, i.e., 42 μm3, as the rest is occupied by the nucleus and organelles [104]. Thus, the random chance of finding the dispersed complement of cER anywhere in the cell, including the cell cortex, is 0.9/42 = 0.02, or about 2%.
|
10.1371/journal.pcbi.1006292 | Gain control with A-type potassium current: IA as a switch between divisive and subtractive inhibition | Neurons process and convey information by transforming barrages of synaptic inputs into spiking activity. Synaptic inhibition typically suppresses the output firing activity of a neuron, and is commonly classified as having a subtractive or divisive effect on a neuron’s output firing activity. Subtractive inhibition can narrow the range of inputs that evoke spiking activity by eliminating responses to non-preferred inputs. Divisive inhibition is a form of gain control: it modifies firing rates while preserving the range of inputs that evoke firing activity. Since these two “modes” of inhibition have distinct impacts on neural coding, it is important to understand the biophysical mechanisms that distinguish these response profiles. In this study, we use simulations and mathematical analysis of a neuron model to find the specific conditions (parameter sets) for which inhibitory inputs have subtractive or divisive effects. Significantly, we identify a novel role for the A-type Potassium current (IA). In our model, this fast-activating, slowly-inactivating outward current acts as a switch between subtractive and divisive inhibition. In particular, if IA is strong (large maximal conductance) and fast (activates on a time-scale similar to spike initiation), then inhibition has a subtractive effect on neural firing. In contrast, if IA is weak or insufficiently fast-activating, then inhibition has a divisive effect on neural firing. We explain these findings using dynamical systems methods (plane analysis and fast-slow dissection) to define how a spike threshold condition depends on synaptic inputs and IA. Our findings suggest that neurons can “self-regulate” the gain control effects of inhibition via combinations of synaptic plasticity and/or modulation of the conductance and kinetics of A-type Potassium channels. This novel role for IA would add flexibility to neurons and networks, and may relate to recent observations of divisive inhibitory effects on neurons in the nucleus of the solitary tract.
| Neurons process information by generating spikes in response to two types of synaptic inputs. Excitatory inputs increase spike rates and inhibitory inputs decrease spike rates (typically). The interaction between these two input types and the transformation of these inputs into spike outputs is not, however, a simple matter of addition and subtraction. Inhibitory inputs can suppress outputs in a variety of ways. For instance, in some cases, inhibition adjusts the rate of spiking activity while preserving the range of inputs that evoke spiking activity; an important computational principle known as gain control. We use simulations and mathematical analysis of a neuron model to identify properties of a neuron that determine how inhibitory inputs affect spiking activity. Specifically, we demonstrate how the gain control effects of inhibition depend on the A-type Potassium current. This novel role for the A-type Potassium current provides a way for neurons to flexibly regulate how they process synaptic inputs and transmit signals to other areas of the brain.
| The activity of a neuron is driven by barrages of synaptic inputs. Synaptic inputs are classified as either excitatory (those that promote spike generation) and inhibitory (those that impede spike generation). The interplay between these two “opposing” inputs influences how neurons process and transmit information in the brain.
To characterize the nature of inhibition, researchers often distinguish between inhibition that has a subtractive effect on neural firing, versus inhibition that has a divisive effect [1]. Inhibition is said to be subtractive if it reduces the firing activity of a neuron by (roughly) a constant amount, regardless of the strength or amount of synaptic excitation. Inhibition is said to be divisive if it reduces the firing activity of a neuron by an amount that is (roughly) proportional to the neuron’s firing rate. We illustrate this distinction in Fig 1, by showing output firing rate of a neuron as a function of the rate of its excitatory inputs (not actual data).
The differences between these modes of inhibition has important consequences for neural coding. Subtractive inhibition suppresses responses to “non-preferred” stimuli that evoke infrequent responses in the absence of inhibition. This can be useful for promoting the representation of “preferred” inputs. In contrast, divisive inhibition is a mechanism for neural gain control: it reduces the firing rate of a neuron while preserving the overall range of inputs to which the neuron is responsive [2]. Understanding the physiological mechanisms that determine how and why inhibition acts in these two modes is key for understanding how neurons and networks function. Past studies have identified numerous possibilities for mechanisms underlying these two modes of inhibition, including the stochastic (noisy) nature of synaptic inputs [3], the balance between excitatory and inhibitory inputs [4], shunting inhibition [5, 6], synaptic depression [7], and circuit structure [2, 8], and see [1] for additional review.
In this study, we identify a novel role for A-type voltage-gated potassium current in determining whether inhibition acts in a subtractive or divisive manner. Voltage-gated K+ channels play an important role in regulating neuronal excitability [9, 10]. Here, we focus on the class of K+ channels that produce an A-type current [11]. These outward currents are mediated by a variety of membrane-bound channels [12–14], found primarily on dendrites [15] but with a somatic location in some cells [16, e.g.]. A-type currents vary greatly in their voltage dependence and kinetics. Although a limited number of channels are typically open (active) at the resting membrane potential, producing a “window current” [17, 18], additional hyperpolarization further “primes” [19] or de-inactivates the membrane [20], making more channels available for activation by a depolarizing stimulus. Thus, the magnitude of A-type currents are particularly sensitive to inhibitory inputs. Inactivation kinetics vary greatly, ranging from less than 20 ms to as much as 600 ms, even within populations of neurons sharing a single potassium channel subfamily [14, e.g.].
Through mathematical analysis and simulations, we explore the combined effects of synaptic inputs and voltage-gated ion currents on spiking dynamics of a neuron model. We find that if the A-type current is sufficiently large and activates rapidly, then it combines with inhibitory inputs to suppress firing activity in a subtractive manner. If, instead, the A-current is sufficiently weak or activates slowly (relative to spike initiation dynamics), then inhibition has a divisive effect on firing rates. Our work identifies a route through which adaptive or dynamic changes to the intrinsic dynamics of neurons (for example, through modification of ion currents [21]) can modulate the effects of inhibition. This capability for individual neurons to switch between different inhibition “regimes” could provide useful flexibility to neural systems.
We simulate and analyze two models of neural dynamics. The first is a one-compartment model that approximates a neuron as a single, isopotential unit (a “point neuron” model). The second is a multi-compartment model that includes a region of voltage-gated currents attached to a spatially-extended region of passive membrane (“soma” and “dendrite” regions, respectively). We describe these models below.
The dynamics of membrane potential, V, in the one-compartment neuron model are
C V ′ = - I L - I K - I A - I N a - I S y n , E - I S y n , I (1)
where the membrane capacitance is C = 1 μF/cm2. Here, and throughout, we use V′ to indicate the time-derivative of V, i.e. dV/dt. The ionic currents (leak, potassium, A-type potassium, and sodium) are given by the equations
IL=gL(V−VL),IK=gKn4(V−VK),IA=gAa3b(V−VK),INa=gNam3h(V−VNa). (2)
We use the following fixed parameter values for maximal conductances: gL = 1 mS/cm2, gK = 45 mS/cm2, and gNa = 37 mS/cm2. We use a range of values for the maximal conductance of the A-current (gA) to observe transitions between subtractive and divisive effects of inhibition. The reversal potentials are VL = −70 mV, VK = −80 mV, and VNa = 55 mV.
We make several simplifications, similar to those first suggested in [22], to the gating variables in the model. We identify sodium activation as a fast process and assume it evolves instantaneously to its voltage-dependent steady-state value. That is, we let m = m∞(V) = 1/(1 + e−(V+30)/15). In addition, we observe an approximately linear relationship between sodium inactivation and potassium inactivation and thus set h = 1 − n.
The remaining gating variables are n, a, and b. Their dynamics are described by equations of the form
X ′ = ϕ X X ∞ ( V ) - X τ X ( V ) , X = n , a , b . (3)
The voltage-dependent steady-state functions are of the form X ∞ ( V ) = 1 / ( 1 + e ( X - θ X ) / σ X ). For the potassium activation variable, n, we assume that ϕn = 0.75, θn = −32 and σn = −8. The time-scale for the n variable is voltage-dependent: τn(V) = 1 + 100/(1 + e(V+80)/26). Similar to the model presented in [23], we assume that ϕa = 1, θa = −50 and σa = 20 for A-type potassium activation, and ϕb = 1, θb = −70 and σb = −6 for A-type potassium inactivation. The time-scales for the A-type current are constants: τa = 2 ms and τb = 150 ms.
Inputs to the model include synaptic excitation (ISyn,E) and inhibition (ISyn,I). Excitatory current is ISyn,E = gSyn,E sE(V − VE) and inhibitory current is given by an analogous equation. The maximal excitatory and inhibitory conductances (gSyn,E and gSyn,I) are parameters that we vary in simulations. The reversal potentials are VE = 0 mV for excitation and VI = −85 mV for inhibition. The gating variables, sE and sI, reset to one at the time of a synaptic event and decay with an exponential time-course. That is, the excitatory gating variable is defined as
s E ( t ) = { 1 if t = t E e - β E ( t - t E ) if t > t E (4)
where tE is the time of the most recent excitatory event and the decay time constant is βE = 0.2 ms−1. A similar equation holds for the inhibitory gating variable sI, but with a time constant βI = 0.18 ms−1. Excitatory event times are randomly distributed according to a homogeneous Poisson process with rate rE. Inhibitory event times are periodic with rate rI. We vary the values of the rate parameters (rE, rI) in our investigations. Our choice of these input patterns simplifies some of our mathematical analysis. In addition, our choice of periodic inhibitory events was motivated by the design of in vitro experiments, presented in [24], in which inhibitory interneurons were activated periodically using optogenetic techniques. In additional simulations included as S1 and S2 Figs we allowed the timing of inhibitory inputs to be random with event times drawn from a homogeneous Poisson process. We did not observe substantially different results in simulations that used these non-periodic inhibitory inputs.
In some simulations we augment the one-compartment (point neuron) model by attaching additional compartments that represent a dendritic process. We assume that the dendrite consists of nine equally-sized compartments. Moreover, the neuron receives inhibitory input at its soma (the first compartment) and excitatory input at a dendritic compartment.
Voltage in the first compartment (soma) is denoted V1 and is given by Eq 1 with synaptic excitation removed and with a new term representing the flow of current between compartments (axial current):
C V 1 ′ = - I L , 1 - I K - I A - I N a - I A x , 1 - I S y n , I . (5)
The remaining dendritic compartments do not include potassium, A-type potassium, or sodium currents, and thus Vj for 2 ≤ j ≤ 10 follows the linear dynamics of a passive cable:
C V j ′ = { - I L , j - I A x , j - I S y n , E at location of excitatory inputs - I L , j - I A x , j at other locations . (6)
Leak conductance in the dendrite compartments is gL = 0.1 mS/cm2 (one-tenth the value in the first compartment). Axial current is
I A x , j = { g A x ( V 1 - V 2 ) for j = 1 g A x ( - V j - 1 + 2 V j - V j + 1 ) for 2 ≤ j ≤ 9 g A x ( V 10 - V 9 ) for j = 10 (7)
where gAx = 10 mS/cm2.
Input currents are defined in a manner identical to inputs in the one-compartment model. Excitatory and inhibitory gating variables follow Eq 4. Excitatory synaptic event times are drawn from a homogeneous Poisson process with rate rE and inhibitory synaptic event times are periodic with rate rI. These constants, as well as synaptic input strengths (gSyn,E, gSyn,I) and the compartment targeted by the excitatory inputs, are parameters we vary in our investigations.
We simulated the point-neuron and multi-compartment neuron models using software written in the C computer programming language. We integrated differential equations using the fourth order implicit Runge-Kutta method available in the GNU scientific library. We also simulated the one-compartment model and a reduced model version of the one-compartment using XPPAUT, and performed bifurcation analysis of these models using the AUTO feature of XPPAUT [25]. Simulation code is available for download at https://github.com/jhgoldwyn/Gain-Control-With-IA.
We first study the relationship between excitatory input rate (rE) and firing output rate (rout) of the one-compartment model. In Fig 2A, we plot examples of this input/output relationship for simulations without inhibition (empty circles, gSyn,I = 0) and with inhibition (filled circles, gSyn,I = 1). The A-channel conductance in these simulations is gA = 20 mS/cm2. For these parameters, we observe that inhibition reduces the model neuron’s output firing rate, but the neuron continues to fire in response to arbitrarily low input rates.
An additional way to view the effect of inhibition is to plot output firing rates in the presence of inhibition as a function of output firing rates in the absence of inhibition, as we have done in Fig 2C. There is a roughly linear relationship between these output firing rates, which we describe by fitting these data with a threshold-linear function of the form
y = [ m ( x - x 0 ) ] + (8)
where the symbol [⋅]+ indicates we set y = 0 if the argument m(x − x0) is negative. We obtain the slope parameter m and the x-intercept parameter x0 by applying a curve-fitting procedure (using the fminsearch command in MATLAB) to the portion of data for which the output firing rate in the presence of inhibition is less than five spikes per second. In this example, inhibition affects the value of the slope parameter m, but the value of x0 is nearly zero. We identify responses with these characteristics as cases in which the effect of inhibition is divisive.
In Fig 2B, we increase the A-channel conductance to gA = 40 mS/cm2. We observe that inhibition has a different effect on the input/output curve in these simulations. In the presence of inhibition (filled circles), there is now a non-zero value of the input rate below which the neuron model does not spike (rout = 0 for rE ⪅ 30). Moreover, when we view the relationship between output firing rates with and without inhibition in Fig 2C, we observe a rightward shift of the threshold-linear function fit to these data (positive-valued x-intercept). We identify responses with these characteristics as cases in which the effect of inhibition is subtractive.
Although we refer throughout to the effect of inhibition on responses as being either divisive or susbtractive, this is a simplification of a more complicated and subtle reality. In fact, responses can show characteristics of both divisive and subtractive inhibition. In particular, the input/output curves can be right-shifted (evidence of subtractive inhibition) and have slopes that are decreased relative to slopes for gA = 0 (evidence of divisive inhibition). We provide evidence of such “mixed” responses in S3 Fig. To be clear: we refer to scenarios in which the input/output curves have only a change of slope as divisive, and scenarios in which the input/output curve is right-shifted as subtractive. In other words, the subtractive case will also include “mixed” responses.
We identify two parameters in the one-compartment model that are key factors in determining whether inhibition has a divisive or subtractive effect on firing rate responses: the A-channel conductance (gA) and the excitatory synaptic conductance (gSyn,E). In Fig 3A we show a set of threshold-linear functions computed using gA = 20, 30 and 40, and synaptic excitation strength fixed at gSyn,E = 0.5. The transition from divisive to subtractive inhibition is evident in the rightward shift of these threshold-linear functions with increasing values of gA. This transition occurs, for this parameter set, for gA ≈ 33, a point we investigate in more detail below, with simulations and phase plane analysis.
In Fig 3B, we show a set of threshold-linear functions with gA = 30 fixed, but now varying the value of gSyn,E from 0.4 to 0.7. The stronger excitatory inputs (gSyn,E = 0.5, 0.7) cause inhibition to have a divisive effect, while the weaker excitatory input (gSyn,E = 0.4) causes inhibition to have a subtractive effect. In these simulations, we do not vary the parameters associated with inhibition. They are gSyn,I = 1 and rI = 50 Hz.
From these simulations, we conclude that the effect of inhibition on firing rates in the one-compartment model can switch from divisive to subtractive for sufficiently strong A-current conductance or sufficiently weak excitatory synaptic conductance. In the parameter plane of gA and gSyn,E, then, there is a boundary that separates parameter sets that produce divisive inhibition from parameter sets that produce subtractive inhibition. We map this boundary by performing simulations throughout the (gA, gSyn,E) parameter space. For each simulation, we fit threshold-linear functions to characterize the relationship between output firing rates in the presence and absence of inhibition. We then find the smallest value of gA for which the x-intercept of the threshold-linear function is right-shifted by more than two spikes per second and label this as boundary between subtractive and divisive inhibition.
In Fig 4, we show the results of this parameter exploration. We performed these simulations and classification procedure for several values of inhibition conductance strength (varying values of gSyn,I, in Fig 4A), and for several values of inhibition rate (varying values of rI, in Fig 4B). The lines in each panel separate parameter regions for which inhibition is divisive (lower right corners in each panel) from parameter regions in which inhibition is subtractive. This confirms our earlier observation that the effect of inhibition is subtractive if A-channel conductance is sufficiently strong or excitatory inputs are sufficiently weak.
These simulations also demonstrate that inhibition parameters modify (weakly) the location of the boundary between divisive and subtractive inhibition in the (gA, gSyn,E) parameter plane. Stronger inhibition (either through larger gSyn,I or larger rI values) decreases the portion of the (gA, gSyn,E) parameter plane in which inhibition has a divisive effect on firing rate responses.
We use mathematical analysis to derive the parameter regions in which the model exhibits either a divisive or subtractive response to inhibition. We begin by considering a reduced model in which activation of the A-current is instantaneous; that is, a = a∞(V). Later, we discuss how the model’s response to inhibition may change if this assumption does not hold.
A key assumption in the formulation and analysis of the reduced model is that the A-current activates sufficiently fast so that the dynamical variable a can be set to its voltage-dependent steady state value; that is, we set a = a∞(V). One effect of this change from a evolving dynamically with τa = 2 ms (the full model), to a evolving instantaneously as a∞(V) (the reduced model), is that excitation must be much stronger in the reduced model to observe subtractive inhibition. Typical values of gSyn,E in the full model are around 0.5 (see Fig 4), and typical values of gSyn,E in the reduced model are around 3. This suggests that the speed of A-current activation (not just the strength of the A-current) plays a role in switching the effect of inhibition from divisive to subtractive.
For inhibition to have a subtractive effect, responses to infrequent excitatory inputs must be suppressed. In the reduced model, this occurs when gA is sufficiently strong because the A-type channel activates “instantaneously” and can prevent spike initiation. In the one-compartment model with “non-instantaneous” a variable, large gA could switch the effect of inhibition to subtractive, but only if excitatory input strength was also sufficiently small (recall Fig 4). The importance of small gSyn,E is demonstrated in Fig 11. We show time-courses of voltage in the one-compartment model for gSyn,E = 0.2, 0.5, and 1, and with gA = 0 and gI = 0. In all cases, the input evokes an output spike. Notice, however, that as input strength weakens, there is a marked delay in the time before spike initiation. For the weakest input used (gSyn,E = 0.2), there is a delay of roughly 2 ms before the rapid upstroke of V at the onset of the action potential. During this “pause”, voltage is slowly ramping up and, simultaneously, recruiting additional A-current as the a variable activates. The amount of IA available to suppress spike initiation depends, therefore, on A-channel maximal conductance (gA) and also the time-constant of IA activation (τA).
From this observation, we draw the following conclusion: “non-instantaneous” IA can act to switch the effect of inhibition from divisive to subtractive, but only if it activates rapidly enough relative to the dynamics of spike initiation. To illustrate our point, we simulated the model with three values of A-channel activation time constant (τA = 0.5, 1, 2), using gSyn,E = 0.5, gSyn,I = 1, and rI = 50 Hz. As shown in Fig 11B, inhibition is divisive for slower activation kinetics (τA = 1, 2) and subtractive for faster activation kinetics (τA = 0.5). In Fig 11C, we map the boundary between divisive and subtractive inhibition in the (gSyn,E, gA) parameter plane. There is a strong effect of τA. Faster activation kinetics (smaller τA values) shift the critical point at which inhibition switches from divisive to subtractive to lower values values of gA.
Our prior observation, that delaying spike initiation allows inhibition to have a subtractive effect for “non-instantaneous” A-channel activation, led us to investigate other cellular mechanisms that could have a similar effect. To this end, we considered a multi-compartment neuron model that describes a soma and passive dendrite. Inhibition and voltage-gated currents are restricted to the soma compartment, and excitation targets a location somewhere on the dendrite. Passive cable theory tells us that the amplitudes of excitatory post-synaptic potentials attenuate and their rising slopes become less steep as signals spread along the cable [28]. By varying the location of excitatory synaptic inputs to the dendrite in the multi-compartment model, we can, therefore, adjust the shape of excitatory post-synaptic potentials as they arrive in the soma.
Examples of action potentials, recorded in the soma compartment, in response to inputs at different locations along the dendrite are shown in Fig 12A. Synaptic conductance strength is constant (gSyn,E = 2 in these simulations). Inputs that arrive proximal to the soma are large and fast-rising relative to responses to more distal inputs, and thus evoke action potentials with shorter latencies. The parameter cptin identifies the compartment that receives synaptic excitation. It takes values from 1 (proximal) to 9 (distal).
We included synaptic inhibition in the model targeting the soma, and used simulations to characterize the effect of inhibition as either subtractive or divisive. In Fig 12B we observe a transition from divisive to subtractive inhibition as we move the location on the dendrite at which synaptic excitation targets the cell. This matches our expectation: synaptic excitation placed at more distant locations will generate weaker and slower rising inputs in the soma. This will, in turn, lead to spikes that initiate slowly and that give time for A-channel conductance to activate and prevent spike generation. In additional simulations included as S4 Fig we targeted inhibitory inputs to the dendrite and did not observe substantially different results.
We explored the (gA, gSyn,E) parameter plane and identified the boundary separating regions in which inhibition has a divisive effect on firing rates and regions in which inhibition has a subtractive effect (following the procedure used previously for Fig 4). We find that there is a dramatic effect of input location, as shown in Fig 12C. The region of the (gA, gSyn,E) parameter plane over which inhibition has a divisive effect is smaller when inputs are more distant from the soma.
Neurons process and convey information in the brain by converting barrages of synaptic inputs into spiking outputs. This transfer from inputs to outputs is a highly complex process due to the inherently noisy and nonlinear nature of synaptic and neural processes. Using a combination of computer simulation and mathematical analysis of biophysically-based neuron models, we have probed the relation between synaptic inputs and spiking outputs. We found that the A-type potassium current (a fast-activating, negative feedback current) can act to switch the effect of inhibition on output firing from divisive to subtractive. This provides a clear demonstration of how the internal dynamics of a neuron can control the functional impact of inhibition.
Using simulations and phase plane analysis, we systematically investigated conditions under which inhibition acts on firing rate outputs in a divisive or subtractive manner. We first identified critical values of IA conductance (gA) at which the effect of inhibition switched from divisive (for lower gA values) to subtractive (for higher gA values) (Fig 4). In the reduced model, we approximated this critical value of gA using bifurcation analysis. By tracking the left-knee of the V-nullcline (Fig 5), we identified this critical value of gA as a bifurcation point at which the neuron model ceased to be excitable in response to synaptic inputs (Fig 8). Key in this analysis was the separation of time scales between fast variables (V) and slow variables (n, b). In fact, the inactivation variable b was sufficiently slow that it could be treated as a constant with a value that depended on the input rate (Fig 7). This simplification enabled further analysis. By viewing the spiking output of the model as a Poisson process modified by a refractory period and inhibition-dependent firing threshold, we approximated firing rates at the point of spiking onset (Fig 9) as well as for arbitrary input rates (Fig 10).
A-type potassium current is a source of dynamic, voltage-gated negative feedback that is fast activating. We leveraged this property to obtain analytical results by assuming that the gating variable for IA activation, a, evolved instantaneously to its voltage-dependent equilibrium value (see also [23]). We also performed simulations without this assumption and discovered a delicate interaction between the speeds of IA activation and spike initiation. In particular, subtractive inhibition required that IA is sufficiently strong and that it activates sufficiently rapidly to prevent spike initiation (Fig 11). For our standard value of IA activation (τA = 2 ms), we found that, in conditions of slow spike initiation, IA could “ramp up” during slowly-developing spikes and suppress spike initiation. Weak excitatory inputs, or excitatory inputs targeting more distal regions in a model that included a spatially-extended dendritic process (Fig 12) produced spikes that were slow to initiate, and were therefore scenarios in which inhibition was subtractive for low to modest levels of gA.
Divisive inhibition is a mechanism of neural gain control and has been the subject of numerous studies; see [1] for review. We found that, at lower levels of gA, inhibition can have a divisive effect on the input/ouput properties of a spiking neuron responding to a mixture of random excitatory inputs and periodic inhibitory inputs. The amount of gA altered the slope (gain) of the output firing rate, and thereby tunes the gain control in this system. This result is consistent with the results of a recent in vitro study of neurons in the rostral nucleus of the solitary tract [24]. In that experiment, Chen and colleagues controlled inhibition using optogenetic techniques and constructed threshold linear functions to express the relation between firing responses with and without inhibition (analogous to our Fig 2C, and similar figures). They observed that slopes of threshold-linear function were more shallow for neurons with IA, as compared to neurons in the same nucleus that did not have IA. Thus, the presence of IA enhanced the divisive effect of inhibition. Previous modeling work has identified similar gain control effects by IA [29].
We observed, additionally, that at higher levels of gA, the A-type current can switch the effect of inhibition from divisive to subtractive. This demonstrates a novel example of how the internal dynamics of a neuron interact with synaptic inhibition to change the neuron’s computational properties (input/output relation). Previous studies have explored the multi-faceted ways in which IA current can alter neural dynamics. Connor and Stevens established IA current as a mechanism to prolong interspike intervals of repetitively-firing neurons to arbitrary lengths (“type I” firing dynamics) [11]. Other identified functions of IA include prolonging first spike latency [30], producing burst firing patterns and preventing anodal break (rebound) firing [23], filtering synaptic inputs in favor of slow time-scale NMDA receptor-mediated inputs [31], and affecting the correlation in spiking among neurons responding to common inputs [32]. Our contribution adds to the rich repertoire of IA function.
We have identified routes to subtractive inhibition that depends only on mechanisms that could be readily adjusted by processes of plasticity and neuromodulation. In particular, we have shown that strong and fast IA can lead to subtractive inhibition. The strength and kinetics of the A-type Potassium channels can be modulated in a variety of ways [33–36]. For example, in neurons involved in gastro-intestinal function, A-type potassium channels were modified both by diet [37–39] and gastric disorders [40].
We also found that weak excitatory inputs or more distally-located excitatory inputs led to subtractive inhibition by slowing the onset of action potentials. Synaptic plasticity and modulation of the electrical properties of dendrites can adjust the strength and propagation of excitatory inputs [41], and plasticity of spike initiation zones could change the dynamics of spike initiation [42, 43]. These changes happen at the level of the output neuron. They do not require “global” modulatory effects to change background network activity, circuit structure, or the balance of excitation and inhibition. We conclude, then, that IA can add flexibility to neural systems by allowing neurons to “self-regulate” whether inhibition acts in a subtractive or divisive manner.
|
10.1371/journal.ppat.1000971 | Genome-Wide Mutagenesis Reveals That ORF7 Is a Novel VZV Skin-Tropic Factor | The Varicella Zoster Virus (VZV) is a ubiquitous human alpha-herpesvirus that is the causative agent of chicken pox and shingles. Although an attenuated VZV vaccine (v-Oka) has been widely used in children in the United States, chicken pox outbreaks are still seen, and the shingles vaccine only reduces the risk of shingles by 50%. Therefore, VZV still remains an important public health concern. Knowledge of VZV replication and pathogenesis remains limited due to its highly cell-associated nature in cultured cells, the difficulty of generating recombinant viruses, and VZV's almost exclusive tropism for human cells and tissues. In order to circumvent these hurdles, we cloned the entire VZV (p-Oka) genome into a bacterial artificial chromosome that included a dual-reporter system (GFP and luciferase reporter genes). We used PCR-based mutagenesis and the homologous recombination system in the E. coli to individually delete each of the genome's 70 unique ORFs. The collection of viral mutants obtained was systematically examined both in MeWo cells and in cultured human fetal skin organ samples. We use our genome-wide deletion library to provide novel functional annotations to 51% of the VZV proteome. We found 44 out of 70 VZV ORFs to be essential for viral replication. Among the 26 non-essential ORF deletion mutants, eight have discernable growth defects in MeWo. Interestingly, four ORFs were found to be required for viral replication in skin organ cultures, but not in MeWo cells, suggesting their potential roles as skin tropism factors. One of the genes (ORF7) has never been described as a skin tropic factor. The global profiling of the VZV genome gives further insights into the replication and pathogenesis of this virus, which can lead to improved prevention and therapy of chicken pox and shingles.
| The Varicella Zoster Virus (VZV) is the causative agent of chicken pox and shingles. The long-term efficacy of the current chickenpox vaccine is yet to be determined, and the current shingles vaccine fails to provide protective immunity for a substantial number of individuals. Shingles can also lead to post-herpetic neuralgia (PHN), a debilitating condition associated with an intractable pain that can linger for life. Therefore, VZV remains an important public health concern. We use growth-rate analysis of our genome-wide deletion library to determine the essentiality of all known VZV genes, including novel annotations for 51% of the VZV proteome. We also discovered a novel skin-tropic factor encoded by ORF7. Overall, our identification of genes essential for VZV replication and pathogenesis will serve as the basis for multiple in-depth genetic studies of VZV, which can lead to improved prevention and therapy of chicken pox and shingles. For example, essential genes may be appealing drug targets and genes whose deletion causes a substantial growth defect may be prospective candidates for novel live attenuated vaccines.
| Human varicella-zoster virus (VZV) is a widespread human alpha-herpesvirus, and the majority of the US population has been previously exposed [1]. VZV is the causative agent of chicken pox and shingles, the latter of which is associated with a significant incidence of post-herpetic neuralgia [2], [3]. A universal chicken pox vaccine (v-Oka strain) was first introduced to the United States in 1995, and this immunization program has dramatically reduced chicken pox incidence [4]–[10]. However, outbreaks of chicken pox are still seen [11]–[13], and shingles remains an important concern because the current shingles vaccine only reduces the risk of infection by about 50% [14]. Therefore, VZV is still an important pathogen and remains a public health concern in the U.S. [7], [15]. A better understanding of the biology and pathogenesis of VZV is essential to improve the medical prevention and the treatment of VZV infections.
VZV is the smallest member of the human herpesvirus family, with a linear double-stranded DNA genome (125 kb) that encodes 70 unique ORFs. As a result of the recent development of a VZV cosmid system and of the severe combined immunodeficient mouse model with xenografts of human tissue (SCID-hu), many viral ORFs have been investigated in both biochemical and functional studies, shedding light upon several VZV gene functions [16]–[18]. However, the majority of VZV's 70 unique ORFs have not been studied, and their roles in viral replication and cell-/tissue-specific pathogenesis remain unclear. This is partly due to the absence of an efficient genetic tool to quickly isolate a large number of mutants and a true animal model to screen for in vivo virulence factors on a large scale [2]. Though the functions of many ORFs can only be predicted based on their homologies to other herpesviruses, such as herpes simplex virus 1, our direct manipulation of VZV's ORFs has enabled us to provide functional annotations for the entire VZV genome.
The knowledge of VZV replication and pathogenesis is limited, in part because of its highly cell-associated nature in cultured cells and the difficulty of generating recombinant viruses. In order to circumvent some of these problems, we cloned the VZV (p-Oka strain) genome as a bacterial artificial chromosome (BAC) carrying both green fluorescent protein (GFP) and luciferase reporter genes [19]. We then systematically deleted every open reading frame in the VZV genome. An overview of our method for genome-wide mutagenesis is shown in Figure 1. With a highly efficient homologous recombination system and the dual-reporter system, the recombinant viruses were isolated and analyzed.
Human fetal skin organ culture (SOC) has been previously established to mimic VZV skin infection, which allows for the study of VZV replication and pathogenesis [20]. We further combined SOC with the luciferase assay-based viral detection method to facilitate screening of skin tropism determinants. Although many investigators utilize SCID-hu models (grafts of human tissue in severe combined immunodeficient mice) to study VZV pathogenesis in vivo [21], SOC is a more suitable and cost efficient approach for genome-wide screening the VZV mutant phenotypes. Nevertheless, any interesting findings can be further verified by further in-depth SCID-hu model studies.
The luciferase VZV BAC (VZVLuc) was used to individually delete and/or mutate each of the 70 unique ORFs by employing the E. coli DY380 strain recombination system [22]. As a result, a library of whole-ORF deletion mutants was created. Each mutant DNA obtained from E. coli was transfected into human melanoma (MeWo) cells, and the results provide direct evidence of that 44 ORFs are essential for viral replication in cultured MeWo cells and 26 are non-essential. Moreover, among the non-essential gene group, 8 ORF deletion mutants showed significant growth defects compared to the wild-type strain (p-value <6.07×10−21; see “Statistical Analysis of Mutant Growth Kinetics” section in Materials and Methods), while 18 ORFs were dispensable. All 26 non-essential ORF deletion mutant VZV variants obtained have been tested in SOC. Interestingly, four ORFs were found to be required for optimal viral replication in cultured skin tissue samples, but not in MeWo cells, suggesting their potential roles as skin tropism factors. The results obtained from this study are in agreement with most of those regarding these particular ORFs that have been published to date, and we have provided explanations of all possible discrepancies in the literature. Overall, we provide 51% novel functional annotations to the VZV proteome (36 ORFs).
All VZV ORF deletion mutants were constructed from BAC mutants with a luciferase reporter (VZVLuc) using a PCR-based approach [19], [22] (also see Supplementary Text S1). Construction of ORF rescued BAC mutants was carried out by adapting a two-step homologous recombination approach in E. coli [19], [22] (also see Supplementary Text S1). The generation of a rescue virus is important in order to prove that the deleted fragment was responsible for any growth defect observed in analyses of the mutants. The rescue virus should be able to fully restore the wild-type phenotypes. Because of the large number of ORFs, we chose a small subset of VZV open reading frames to rescure and we have shown these rescue mutants behave as the wild –type strain. A detailed description of these protocols is provided in [19], [22] and an overview is shown in Figure 1. Previous studies in our laboratory have shown that the BAC mutant has an identical growth curve to the wild-type virus [19] and that addition of the luciferase reporter to the BAC virus does not change its growth properties [22].
All of VZV's 70 unique ORFs were deleted and analyzed based on a bioluminescence detection method, as described previously [19]. For 14 ORFs that overlap with adjacent ORFs (ORF8, ORF9A, ORF25, ORF26, ORF27, ORF28, ORF46, ORF47, ORF48, ORF49, ORF50, ORF54, ORF59 and ORF60), respective partial ORF deletions have been constructed and analyzed. A detailed description of these partial ORFs is included in Supplementary Table S2. The results suggest that 44 ORFs are essential for viral replication in cultured MeWo cells (Table 1 and Figure 2). We have confirmed that ORF4 and ORF5 are essential by making genetic rescue viruses. For the essential group, we provide novel functional annotations for 31 of 44 ORFs. All of these VZV essential genes have HSV-1 homologies (Table 1), and the majority of them are conserved among other herpesviruses. These ORFs encode important viral structural proteins, enzymes involved in DNA replication, and transcriptional regulatory proteins.
Among VZV's 44 essential ORFs, the majority encodes proteins with vital functions throughout the viral life cycle. Most VZV proteins that regulate transcription (ORF4, ORF62/71, ORF63/70, and ORF 61) were found to be essential in this study. ORF4 and ORF62/71 are incorporated into the viral tegument, and both encode immediate-early (IE) proteins with transcriptional regulatory activity [23]–[26]. ORF4 and ORF62/71 have been extensively studied, and their essential natures have been suggested previously [27], [28]. Both ORF63/70 and ORF61 encode phosphoproteins primarily localized to the nuclei of infected cells [25]. Although it has been suggested that ORF 63/70 is not essential for viral replication in vitro [29], we could not generate a viable virus from a 63/70 double deletion; this result is in agreement with several other studies [30], [31].
Most of the VZV ORFs that encode glycoproteins are essential. Glycoprotein K (gK) (encoded by ORF5) [32], gB (ORF31), gH (ORF37), gM (ORF50) [33], gL (ORF60) [34], [35], and gE (ORF68) [32], [36] are required for viral replication, and many of them had previously been investigated and reported. Only glycoprotein C (ORF14) [37], [38] and gI (ORF67) [36], [39], [40] deletion mutants produced viable viral progenies, and both of these mutants appeared to suffer a severe growth defect. The results regarding the essentiality of VZV glycoprotein genes in this study are in agreement with the published data.
Essential VZV genes have significantly different enrichment for functional categories than do non-essential genes (Figure 3A). In order to make this calculation, we first listed every gene in a functional category, such as “DNA replication” for a DNA polymerase gene. Then, we compared the proportion of essential (and then of non-essential) genes in each functional category to the background rate expected by chance (e.g. the proportion of genes in that functional category for the entire VZV genome). This calculation was performed using a hypergeometric test. For example, essential genes are significantly enriched for DNA replication (Bonferroni corrected p-value <10−4) and for DNA packaging (Bonferroni corrected p-value <10−4); ORF28 encodes the catalytic subunit of VZV DNA polymerase and ORF16 encodes the subunit of the viral DNA polymerase processivity factor [2]. DNA binding proteins include proteins encoded by ORF6 (primase), ORF29, ORF33 (capsid protein), ORF41 (capsid protein), ORF51 (helicase), ORF52 (component of helicase/primase complex), and ORF55 (component of helicase/primase complex) [41], [42]. Not surprisingly, almost all of the ORFs that encode DNA packaging proteins—including ORF25, ORF26, ORF30, ORF34, ORF42/45, ORF43, ORF54, and nucleocapsid proteins including ORF21, ORF33.5, and ORF40—also fall into the essential gene category. In contrast, non-essential genes were significantly enriched for other (Bonferroni corrected p-value <10−3) and unknown (Bonferroni corrected p-value <0.01) functional categories (Figure 3B).
In this study, we found that 26 ORFs are non-essential genes and 6 of these lack HSV-1 homologies (ORF0, ORF1, ORF2, ORF13, ORF32, and ORF57) (Table 1). According to the growth kinetics (in cultured MeWo cells), 8 ORF mutants had significant growth defects (p-value <6.07×10−21), and the peak signals of the viral detection assay were at least 5-fold less than were those of the wild-type parental strain (Figure 4A). Two of these VZV ORF deletion growth phenotypes, ORF18 and ORF32, have not been previously reported, and two others (ORF23 and ORF35 deletions) have been confirmed to have growth defects in vitro, which is in accordance with previously published data [43], [44]. ORF0 deletion's growth defect has been confirmed by making its genetic revertant [19]. ORF18 and ORF19 respectively encode the small and large subunits of ribonucleotide reductase, and both of them diminished viral growth when deleted in this study. The result on ORF19 is in accordance with previous publications [45]. ORF32 encodes a phosphoprotein that is post-translationally modified by ORF47 protein kinase [46]. Among these 8 viral mutants showing severe growth defects, atypical morphology of virally infected cells was frequently observed, including reduced plaque sizes and altered syncytia formation.
The remaining 18 VZV ORFs had wild type growth curves for viral replication in cultured MeWo cells. In vitro growth curve analysis showed that these ORF deletion mutants had the same growth kinetics as their wild-type parental strain, VZVLuc (Table 1). Previous studies have reported that 15 of these genes (ORF1, ORF2, ORF3, ORF8, ORF10, ORF11, ORF12, ORF13, ORF14, ORF47, ORF57, ORF59, ORF58, ORF64/69, and ORF65) are non-essential [17], [31], [38], [46]–[54]. In this study, three of these ORF mutants (ORF7, ORF15, and ORF36) have been shown to be dispensable for in vitro viral replication for the first time.
The human fetal skin organ culture (SOC) model has been proven to be a simple and convenient alternative to the SCID-hu mouse model in the study of VZV pathogenesis [20], especially in the case of an initial genome-wide screening for skin tropism determinants. Although 26 VZV ORFs were found to be non-essential for viral replication in cultured MeWo cells, it was possible that some of these viral genes encode proteins critical for optimal viral infection in skin tissue. To test this hypothesis, all 26 non-essential ORF deletion viruses were further analyzed in cultured skin-tissue samples.
Every deletion mutant that showed severe growth defects in cultured MeWo cells also demonstrated significantly slow growth kinetics in human skin samples when compared to wild-type VZV (p-value <9.05×10−19). Only two of these genes (ORF35 and ORF67) have been previously reported to be required for viral growth in SCID-hu skin mouse models [40], [44]. Therefore, we have been able to provide novel functional annotations for the other 6 deletion mutants with severe growth defects.
Many non-essential genes appear to cluster together, particularly between ORF0 to ORF15 (Figure 2). More than 70% of ORFs (12 out of 17) in this region are non-essential, compared to 37% of the entire genome. Four out of six VZV ORFs without HSV-1 homologues are also located in this region, so this region may be more evolutionarily variable compared to other highly conserved regions
Among the 18 VZV ORFs dispensable for viral replication in cultured MeWo cells, 14 were also dispensable for viral replication in skin tissue (Table 1). Among the above 14 deletion mutants, we have been able to provide novel ex vivo functional annotations for all but one of these 18 genes (ORF64/69) [31].
Interestingly, among non-essential VZV ORFs, four ORFs (ORF7, ORF10, ORF14, and ORF47) appear to have selective impacts on viral replication in skin tissue. The growth of each virus in SOC was compared with its growth in cultured MeWo cells. These ORF deletion mutants grew like the wild-type strain in vitro (Figure 4B). In contrast, they showed significant growth defects in skin organ cultures (p-value <9.51×10−19; Figure 4C). For instance, the ORF10 deletion mutant had a growth defect in SOC. The bioluminescence signal kept increasing during the entire 7-day experiment period; the total photon count values consistently remained approximately 10-fold less than those of the wild-type strain. The ORF7 deletion mutant virus quickly reached its growth peak around 3 days after inoculation, and then bioluminescence steadily declined. To prove that the VZV ORF7 and ORF10 growth defects observed in cultured skin-tissue samples were due to the functions of the deleted genes rather than to undesirable mutations in other regions of the genome, rescue viruses ORF7R and ORF10R were generated. The growth curve analysis indicated that ORF7R and ORF10R viruses grew in MeWo cells indistinguishably from wild-type VZV, as expected (Figure 4B). In skin organ cultures, they were also able to fully recover the growth defects of the corresponding deletion mutant viruses and grew as well as the wild-type strain (Figure 4C). In contrast, ORF47 deletion virus had a more severe growth defect, approximately 80–100 fold (2 log) less than wild-type VZV. Our results suggest these three ORFs are important for viral replication in human skin tissue but not in cultured MeWo cells.
Three skin-tropic virulence factors (ORF10, ORF14, and ORF47) have been previously identified. VZV ORF10 encodes a tegument protein that enhances transactivation of VZV genes, and it was shown to be dispensable for VZV replication in vitro [47]. Recent studies showed that ORF10 protein is required for efficient VZV virion assembly and is a specific determinant of VZV virulence in SCID-hu skin xenografts but not in human T cells in vivo [55], [56]. ORF14 (gC) has been reported to have reduced infectivity in an SCID-hu skin model [38]. VZV ORF47 encodes a serine/threonine protein kinase and was shown to be dispensable for viral replication in cultured MeWo cells [48]. It has been designated as a virulence factor for both skin tissue and T cells in SCID-hu models [38]. The findings of these three skin-tropic ORFs not only confirmed previous studies but also further verified the similarity between SCID-skin and SOC systems.
In the current study, ORF7 has been identified as a novel skin-tropic virulence factor. In order to confirm the accuracy of our results, we also produced a premature stop-codon mutant (ORF7S) by mutating the 5th codon from TGT to the TGA stop codon (see Table S1). Just like ORF7D, ORF7S showed wild-type growth in MeWo (Figure 4B) but had a growth defect in SOC (Figure 4C).
In this study, a global functional analysis of the entire VZV genome was performed that emphasized the identification of viral ORFs important for viral replication both in cultured MeWo cells and human fetal skin organs. We took full advantage of the highly efficient luciferase VZV BAC system and obtained a library of single ORF deletion mutants. Advanced live culture bioluminescence imaging technology allowed us to systematically test a large number of mutant viruses for comparing viral growth kinetics in different systems.
VZV has a 125-kb DNA genome encoding 70 unique open reading frames (Table 1, Figure 2). In this study, all of the predicted 70 ORFs were individually deleted. Our results directly showed that 44 ORFs encode essential genes and 26 ORFs encode non-essential genes. Among the non-essential group, 8 ORF deletion mutants suffered severe growth defects in MeWo cells. Fourteen ORFs were shown to be dispensable for viral replication, both in MeWo cells and in SOC. We also found 4 tissue tropic factors (ORF7, ORF10, ORF14, and ORF47) that showed a growth defect in SOC but normal growth in MeWo. Three of these tissue-tropic factors (ORF10, ORF14, and ORF47) have been previously identified, but ORF7 has never been previously studied.
In the current study, we have reported ORF7 as a novel VZV skin-tropic factor. ORF7 encodes a 29-kDa tegument protein, and its function remains unknown. The homolog of the VZV ORF7 protein in the herpes simplex virus is the UL51 protein. Recent studies showed that deletion of HSV-1 UL51 causes reduced size plaque formation and low infectivity [56]. Similarly, the function of the UL51 gene product of the pseudorabies virus (PrV) has been investigated by generating a deletion mutant, and the result suggested that the UL51 protein is involved in viral egress, but is not essential for viral replication [57]. Our result suggests that VZV ORF7 might serve as a skin-specific virulence factor. However, the role of ORF7 in pathogenesis needs further investigation.
Despite the large differences between herpesvirus genomes (ranging from 125 kb to >230 kb), all the herpes viruses studied thus far have a similar number of essential genes. For example, HSV-1 encodes 37 essential genes and 48 non-essential genes [58]; human cytomegalovirus (HCMV), which is one of the largest human DNA viruses, encodes 45 essential genes and 117 non-essential genes [59]. Our data suggest that VZV, which contains the smallest genome, encodes 44 essential genes and 26 non-essential genes. A comparison between the essentiality of HSV-1 and HCMV homologues to essential VZV genes is provided in Supplementary Table S3. Of the 44 essential genes, 26 have essential homologues in HSV, and all essential gene homologues conserved in CMV (18 of 44 essential VZV genes) are essential. Therefore, we believe that several of these essential genes perform core functions for all of these herpesviruses.
Unlike the other functional profiling studies performed on HCMV [59], our results did not reveal any VZV-encoded factors that repress viral replication in cultured MeWo cells or in human fetal skin tissue. If such VZV temperance genes existed, enhanced growth kinetics should have been observed by making the corresponding ORF deletion mutants.
There is also an apparent size difference between essential and non-essential ORFs. Essential ORFs are significantly larger in size compared to non-essential ones (μ = 1250 bp vs. μ = 970 bp, respectively, p = 6×10−4 by t-test). The 10 largest VZV ORFs are all essential, while 8 out of 11 VZV ORFs less than 600 bp are non-essential.
All of our results are in agreement with previous VZV functional annotations, except for those on ORFs 9A, 17, 61 and 66, for which we could not generate viral deletion mutant progenies with sufficient titers for growth studies. For example, previous studies indicated that was ORF9A not essential viral growth in cell culture (due to insertion of a premature stop codon) yet they also showed that failure to express either of these genes resulted in growth defects [60]. Therefore, we believe that our findings are at least in partial agreement with previous studies because this previous study utilized a premature stop codon (thus allowing expression of a partial protein), whereas we completely removed ORF9A from the VZV genome.
Although some studies have shown ORF17 to be dispensable for viral replication [61], other studies have shown the gene to be essential for growth under certain conditions [62]. Therefore, we believe this discrepancy can probably be explained by subtle differences in experimental design (such as the temperature of the growth culture, as described in [62], and we believe that our analysis for ORF17 deletion best reflects conditions in vivo.
ORF61 has also been suggested to be a non-essential gene for viral replication in vitro in a previous study [63], [64]. However, we could not retrieve enough infectious viral progeny from the ORF61 deletion clone, even after repeated transfection and extensive incubation. Large deletion mutants of ORF61 [63] and promoter bashing experiments [64] have shown ORF61 to be important for viral replication (albeit non-essential) due to a considerable growth defect shown in the deletion. However, no complete deletion virus has ever been created, so it is possible that the large deletions may have only been sufficient to cause a growth defect, whereas our complete deletion results in a complete loss of VZV replication.
ORF66 has been previously cited as dispensable for viral replication, but we have found it to be essential [65]–[67]. In previous studies, a premature stop codon mutant of ORF66 resulted in a decrease in viral titer, but not in a complete loss of viral replication [65], [66]. Premature stop codons were inserted such that more than 50% of the original coding sequence remained and was able to be expressed, so we believe this discrepancy can be explained by the possible attenuated activity of the partial protein (which did have a substantial growth defect), while our ORF66 deletion removed the entire sequence. For the cosmid-based studies [67], [68], a premature stop codon mutant (with a 21-amino acid partial protein expressed) had to be used to assess the impact of ORF66 on viral replication. However, the authors [68] were also unable to produce infectious virus with a complete ORF66 deletion mutant (which is identical to our results).
In this study, we have presented novel functional annotations for 36 VZV genes. Due to the global nature of our study and the lack of well-defined upstream and downstream regulatory regions for most VZV genes, some of our annotations may have to be redefined by more detailed studies (genes most likely to be affected by adjacent genes are specifically noted in Table 1). Moreover, the current profiling study has provided the first global view of VZV genomic functions in viral replication, which is likely to serve as the basis for further investigative studies on many VZV genes.
Human melanoma (MeWo) cells were grown in DMEM supplemented with 10% fetal calf serum, 100U of penicillin-streptomycin/ml, and 2.5ug of amphotericin B/ml, as previously described, and used to propagate VZV in vitro [18], [69]. VZVLuc containing the entire p-Oka VZV genome was constructed as previously described [19]. Recombinant VZVLuc virus was derived by transfection methods [19], [22] (also see Supplementary Text S1). All primer sequences are listed in Supplementary Table S1. Primer sequences were designed based upon the Dumas VZV strain (Accession Number: NC_001348).
VZVLuc DNAs were transfected into MeWo cells using the FuGene 6 transfection kit (Roche, Indianapolis, IN) [19], [22] (also see Supplementary Text S1). Recombinant viruses were titrated by infectious focus assay. MeWo cells were seeded in 6-well tissue culture plates and inoculated with serial dilutions of VZV-infected MeWo cell suspensions. Plaques were counted by fluorescence microscopy at 3 days after inoculation. All transfections were performed a minimum of 3 times. Since VZV is highly cell-associated under tissue culture conditions, mutant VZV-infected MeWo cells were harvested, titrated and stored in liquid nitrogen. Wild-type infections served as positive controls and mock infections served as negative controls.
In vitro growth curve analyses were carried out by live-cell bioluminescence detection assay. MeWo cells were infected with 100 PFU of infected MeWo cell suspensions on 6-well tissue culture plates. Every 24 h, the cell culture medium was replaced with medium containing 150 ug/ml D-luciferin (Xenogen, Alameda, CA). After incubation at 37°C for 10 min, the bioluminescent signals were quantified and recorded using an IVIS Imaging System (Xenogen), following the manufacturer's instructions. After each measurement, the luciferin-containing medium was replaced with fresh cell culture medium. Measurements were taken daily from the same plate for 7 days. Bioluminescence signal data from each sample were quantified by manually demarcating regions of interest and analyzed using LivingImage analysis software (Xenogen). It has been demonstrated previously that both the infectious center assay and the luciferase assay correlate well [19], [22].
Human fetal skin-tissue samples (∼20 weeks gestational age) were acquired from Advance Biosciences Resources (Alameda, CA). Skin organ-culture techniques were as previously described [20]. Ex vivo growth curve analyses were carried out by live-tissue bioluminescence assay. Infected MeWo cells were titrated and then re-suspended in skin organ culture media (SOCM). After 24 h of incubation, each skin-tissue section was injected five times with 10 ul of the virus-infected cell suspension (total inoculation was 5×103 PFU per tissue) by a 1-ml syringe fitted with a 27-gauge needle attached to a volumetric stepper (Tridak, Brookfield, CT). After inoculation, the sections were placed individually on 500 um mesh NetWell inserts (Corning, Corning, NY) that rested above 1ml of SOCM in each well of 12-well plates and followed by a 24 h incubation in a tissue culture incubator, 37°C, 5%CO2. Each 24 h, SOCM was replaced with media containing 150ug/ml of D-luciferin. Following 10 min incubation at 37°C, the bioluminescence being emitted from individual cultured skin-tissue samples was recorded using the IVIS Imaging System. After the measurements, each sample (still on a NetWell insert) was transferred onto new 12-well plates with fresh SOCM. The measuring process was repeated every 24 h for 7 days. Bioluminescence signals from manually defined regions of interest were quantified and analyzed. All experiments were performed in triplicate. Wild-type infections served as positive controls and mock infections served as negative controls.
Wild type and mutant growth curves (7 time points, 3 replicates each) were compared using the “timecourse” Bioconductor package [70], [71]. The difference in growth rate for wild type and mutant growth curves was estimated by the mb.long function was used to estimate a Hotelling T2 test statistic using the mb.long function. P-values for the T2 test statistic were calculated using an F-distribution. The T2 test statistic did an excellent job of quantifying the difference in growth curves, but a very strict p-value cutoff was required in order to define statistically significant growth defects (implying that the test statistic may be too sensitive). Therefore, we used a Mann-Whitney U test in order to determine which individual time points significantly differed between wild type and deletion mutant strains. All strains with reported growth defects have at least 6 significantly reduced time points (p<0.05).
|
10.1371/journal.pbio.1002435 | How Many Parameters Does It Take to Describe Disease Tolerance? | The study of infectious disease has been aided by model organisms, which have helped to elucidate molecular mechanisms and contributed to the development of new treatments; however, the lack of a conceptual framework for unifying findings across models, combined with host variability, has impeded progress and translation. Here, we fill this gap with a simple graphical and mathematical framework to study disease tolerance, the dose response curve relating health to microbe load; this approach helped uncover parameters that were previously overlooked. Using a model experimental system in which we challenged Drosophila melanogaster with the pathogen Listeria monocytogenes, we tested this framework, finding that microbe growth, the immune response, and disease tolerance were all well represented by sigmoid models. As we altered the system by varying host or pathogen genetics, disease tolerance varied, as we would expect if it was indeed governed by parameters controlling the sensitivity of the system (the number of bacteria required to trigger a response) and maximal effect size according to a logistic equation. Though either the pathogen or host immune response or both together could theoretically be the proximal cause of pathology that killed the flies, we found that the pathogen, but not the immune response, drove damage in this model. With this new understanding of the circuitry controlling disease tolerance, we can now propose better ways of choosing, combining, and developing treatments.
| It is an intuitive assumption that the severity of symptoms suffered during an infection must be linked to pathogen loads. However, the dose–response relationship explaining how health varies with respect to pathogen load is non-linear and can be described as a “disease tolerance curve;” this relationship can vary in response to the genetic properties of the host or pathogen as well as environmental conditions. We studied what changes in the shape of this curve can teach us about the underlying circuitry of the immune response. Using a model system in which we infected fruit flies with the bacterial pathogen Listeria monocytogenes, we observed an S-shaped disease tolerance curve. This type of curve can be described by three or four parameters in a standard manner, which allowed us to develop a simple mathematical model to explain how the curve is expected to change shape as the immune response changes. After observing the variation in curve shape due to host and pathogen genetic variation, we conclude that the damage caused by Listeria infection does not result from an over-exuberant immune response but rather is caused more directly by the pathogen.
| The clinical goal of treating infectious diseases is to reduce the levels of sickness experienced by infected hosts. One approach to studying this problem is to quantitate illness by correlating the dose response of pathology to pathogen load; graphs like these are called “disease tolerance curves” [1–6]. We argue that the shape of the dose response curves and the underlying mathematical functions producing these curves can teach us how to alter the health-by-microbe relationship. For example, if disease tolerance curves are linear, we need only discover the molecular mechanisms that control the curves’ slopes. If these curves have a more complex shape—for example, a sigmoid shape—then we will need to measure more variables.
Since there are not universally agreed upon definitions for many of the words we use to describe the immune response and health, we start by defining the following terms: vigor, resistance, resilience, and tolerance. “Vigor” is the health of an uninfected individual, and “resistance” describes the host’s ability to limit microbe load [1–6]. We use “resilience” to describe the ability of infected hosts to return to their original healthy condition, much in the same way the word is used to describe the way a perturbed ecological system returns to its origin [7,8]. For example, a resilient host may suffer from an infection but would easily bounce back to health, whereas infection in a non-resilient host might lead to permanent disability or death. We apply the term resilience to an infected individual because we can follow the path that individual takes from health through sickness and back. “Disease tolerance” is the dose response curve summarizing the pathogen loads that are required to produce a range of host responses. Disease tolerance differs from resilience in that tolerance is an emergent property of populations, while resilience is a property of an individual; disease tolerance measures how resilience changes across a population as pathogen load is varied. Disease tolerance, by definition, cannot be measured in a single individual [9]. The model described in this paper shows how these distinctions between resilience of individuals and tolerance of populations disappear once we understand infection dynamics to the point that we can predict how a population would behave knowing the properties of an individual.
To study disease tolerance we used a simple infection system in which we injected a pathogen (Listeria monocytogenes) into a fruit fly (Drosophila melanogaster) [10–14]. These L. monocytogenes-infected flies suffer from a variety of ailments, including alterations in their circadian rhythm, cold-coma recovery, climbing ability, feeding behavior, metabolism, and death [12,15–17]. To determine the disease tolerance curve, we plot the survival rates of the infected flies against microbe loads measured 2 days post infection (DPI). We used this system because it allowed us to measure large numbers of samples and to take advantage of the genetic tools and the understanding of innate immunity available for this model organism.
Our immediate goal is to define the mechanisms controlling tolerance. We start with a simple mathematical model representing feedbacks between microbe growth, the immune response, damage, and health, modeling each using logistic equations. We then measured the ground-truth of these models using a D. melanogaster/L. monocytogenes infection that can lead to lethal outcomes. We experimentally measured microbe growth, the immune response, and disease tolerance and found that each of them was well described by sigmoid curves, suggesting that we need at least three variables to describe each. In the case of tolerance, this meant that we had to measure host vigor, the number of bacteria required to injure the host, and the maximum death rate achieved by the fly. In the case of antimicrobial peptide (AMP) transcript production, we measured basal levels, the number of bacteria required to induce transcripts to 50% of their maximum value, and the maximum transcript level. We examined how alteration of the values of the parameters in the model changed the output of the model to distinguish between changes that we might see if microbes or the immune response were the principal mediators of damage. Upon testing a variety of D. melanogaster natural variants and mutants along with some L. monocytogenes mutants, we concluded that, in this system, the bacteria and not the immune response is responsible for the damage caused by the infection. When we monitored how disease tolerance changed as we varied host or pathogen properties, we found the tolerance curve changed shape in the manner predicted by the model; for example, the number of bacteria required to damage health changed, while the vigor and maximum death rate remained constant. We demonstrate that this dual approach of measuring and modeling full-length disease tolerance curves can reveal previously unrecognized parameters controlling disease tolerance.
We started by making a graphical model to explain four parameters: microbe load, the immune response, microbe-induced damage, and health (Fig 1). We modeled the immune response so that it could both limit microbe growth and kill microbes. We previously found that Mycobacterium marinum infected flies waste upon infection and predicted that the day they died they would have exhausted their glycogen and fat stores [18,19]. We also found that L. monocytogenes infected flies waste during infection [15]; to explain death, we used these empirical data to imagine that there was a store of “health” that could be depleted by the infection. In our model, both the immune response and the microbes can cause damage, which depletes health and increases the death rate. Immunopathology directly affects health, while microbes secrete damage effectors that impact health.
We built a simple mathematical model based on the graphical model largely using sigmoid functions to define each parameter (see Experimental Procedures) to help define our assumptions about the mechanisms governing disease tolerance. This was an obvious choice for microbe growth and the immune response, as these equations are well described in classical mathematical models of infections [20]. We used the same mathematical function to describe microbe-induced damage, since we anticipate that the shape of the curve is due to some enzymatic rate-limiting step that will likely follow sigmoid Michaelis-Menten kinetics [21,22].
To determine the ground-truth of this model, we measured the behavior of microbes, the immune response, and health in a model system in which we inject L. monocytogenes into the body cavity of D. melanogaster. We chose this model infection because it is simple to obtain the measurements we need to define resistance and disease tolerance and because we have a collection of host mutants that we know affect resistance and tolerance to L. monocytogenes [10–12]. We can readily measure microbe loads by homogenizing whole flies at any point in the infection to plate them to determine colony-forming units (CFU). We can determine the health of the infected flies by measuring the median time to death following infection.
When L. monocytogenes is injected into flies that are incubated at 29°C, the resulting bacterial growth is best described as logistic (Fig 2A, logistic adjusted r2 = 0.594, adjusted linear and adjusted exponential r2 < 0.1) with a maximum growth rate of 0.1656 log10 (CFU)/h. L. monocytogenes loads reach a plateau at 24 h post infection.
We wanted to define the relationship between AMP levels to a given microbe load to understand how many parameters we need to use to measure the immune response. Because bacteria grow rapidly to reach their plateau, and wounding flies to introduce bacteria causes an immune response on its own, we reasoned it would be difficult to measure the immune responses at intermediate microbe loads by simply injecting flies with different amounts of microbes. Instead, we took flies that had high levels of bacteria and slowly cured them of the infection so that we could measure the relationship between AMP transcripts and microbe load essentially at steady state. The fly immune response cannot clear L. monocytogenes on its own; therefore, to clear the bacteria we transferred infected flies to fly media containing 1 mg/ml of ampicillin 2 DPI. We observed that L. monocytogenes clearance followed exponential decay dynamics (Fig 2B, adjusted r2 = 0.7140). L. monocytogenes infection in the fly decreases survival, but ampicillin-treated flies lived just long as uninfected flies (Fig 2C).
Having developed a method to slowly decrease microbe loads, we performed the following experiment: We injected one set of flies with phosphate buffered saline (PBS), fed them with ampicillin, and followed them regularly to serve as wounding, aging, and antibiotic treatment controls. We took a second set of flies and injected L. monocytogenes. This group was split in two; one portion was fed ampicillin to clear the microbes, and the other was not. We collected flies for plating to determine microbe loads, and for RNA preparation to run a microarray at regular intervals (S1 Fig).
We analyzed the expression of the known and presumed AMPs with respect to microbe loads and found that AMPs are expressed according to a sigmoid relationship with respect to microbe load. We plotted the expression of AMPs to CFU levels in the fly and found the relationship between AMP transcription and microbe load is best fit by a four-parameter sigmoid (Fig 3A). Accordingly, each AMP transcript has a baseline (bottom plateau), maximal concentration (top plateau), effective bacterial concentration producing a half maximal effective concentration (EC50), and slope. We observe that the AMPs expressed during infection differ in their EC50 (Fig 3B) over a range of about 100-fold (20–2,000 CFUs). AMP expression does not continue to increase as microbe number increases past a certain point, and each AMP transcript reaches a characteristic plateau. Each AMP transcript can thus be characterized by its maximum expression in relation to its baseline, and this relationship varies over an 8-fold range between AMPs. We observed no significant relationship between EC50 and maximum expression, suggesting that EC50 and maximum expression can vary independently of each other. We found an exponential relationship between Hill slope and EC50 (Fig 3D), suggesting that AMPs requiring large numbers of bacteria for induction turn on in a switch-like manner, while those turned on at low microbe loads are turned on gradually. This gives us insight into the tuning of the fly’s immune response; overall, AMPs will be turned on gradually at microbe loads below 100 bacteria per fly but will rapidly increase as loads pass 1,000 bacteria per fly.
Though AMPs had a rather simple sigmoid relationship with microbe load, this was not the case for many other genes modulated during infection. Some of these other gene expression patterns showed hysteresis, by which we mean that gene expression depends upon past conditions and not only the immediate conditions. For example, gene expression for these hysteretic transcripts had one pattern as the fly moved from health to sickness and another as the fly returned from sickness to health. This hysteresis can be useful because it helps us define different states of the infection. To visualize the “disease space” traversed by infected flies and to identify the different states of the infection process, we used topological data analysis (TDA). It is common practice when clustering data to plot the data using some sort of tree structure, as is done with hierarchical clustering or spade analyses [23,24]. This makes good sense when dealing with data for a developing system in which the animal or cell enters a new state, but it is problematic when dealing with a system that returns to its original state; resilient systems should not be fit by a tree and are better described by loops. Topological data analysis is sensitive to the topology of the data and will not arbitrarily linearize a loop and then force it to fit a tree [25,26]. Instead, the analysis simply clusters related data points, represented as nodes on a network graph, and the shape of that graph reveals the connections between the time points (for recent examples see [27–29]). In the case of a resilient system, such as flies recovering from an infection, we find this graph forms a loop.
Fig 4 shows the TDA graph we built to describe our data, and 4A describes the three basic treatment groups in the dataset (uninfected, infected, and antibiotic-treated). In Fig 4B, the phase curve is colored by CFU (blue indicates low and red indicate high microbe loads). The green arrow indicates progression from health to infection and back, moving clockwise around the figure. Uninfected flies are at the top of the figure, and the figure progresses through acutely wounded flies, sick flies, and on to recovering flies, which link back to the original uninfected flies. Infection was initiated by the injection of 100 CFU, and L. monocytogenes levels increase following this. The path followed by moribund flies deviates from recovering flies to form a spur on the loop (marked as section iv). Flies that received ampicillin decrease in CFUs, and flies that have low CFUs loop back to overlap with uninfected controls.
To determine which genes show altered expression in different parts of the disease map, we performed fuzzy c-means clustering on this dataset to identify genes with similar expression patterns (S2 Fig, S2 Table). We then performed gene ontology analysis on each cluster to identify characteristic physiological changes within the cluster (S3 Table) and then asked about the expression patterns of these genes on the TDA graph. Acute stress response genes, such as heat shock protein 26, were activated immediately after injection and then quickly returned to baseline levels (Fig 4B and 4C). AMP expression follows CFU (Fig 4B and 4D). We identified groups of genes that are highly expressed by flies in the spur approaching death (Fig 4B and 4E) and recovering from infection (Fig 4B and 4F). The transcripts for these “morbidity genes” (clusters 1, 2, 11, 12, 22, 35, 43) continue to increase even when microbe levels plateau. Gene ontology analysis of fuzzy c-means clusters containing these morbidity genes did not yield significantly enriched terms. Among the top 50 differentially expressed morbidity genes, Bteb2 and run are predicted to encode proteins with DNA binding activity. CG3117 has predicted protease activity, and diedel is a putative negative regulator of JAK/STAT signaling. By contrast, the recovery genes (clusters 18, 20, 23, 25, 29, 30) are enriched for sugar and fatty acid metabolism genes that are turned up above baseline upon recovery.
To measure a disease tolerance curve, we recorded the response of the host to a broad range of initial pathogen doses. We did this by injecting L. monocytogenes into flies over a range of ten to 100,000 bacteria and allowing the infected flies to die. We injected L. monocytogenes into the hemocoel of flies and monitored bacterial numbers 2 DPI to measure the ability of the fly to resist microbe growth when challenged with a range of infection intensities (Fig 5). We determined the median time to death (MTD) for each inoculum and used this time as a measurement of health (Fig 5B). Plotting microbe load versus MTD produced a curve that was readily fit by a four-parameter logistic sigmoid model (r2 > 0.96) (Fig 5C).
Occasionally, tolerance curves are fit with mathematical functions, but the reason these functions are chosen is that they fit the data and not that they are dissected for further biological insights [5,30–32]. More typically, tolerance is visualized as a linear system, which requires just two parameters, vigor and slope [5,11]. In contrast, our sigmoid model provides four parameters (Fig 5D). In addition to vigor and slope used to describe a line, the sigmoid model adds the parameters of EC50 and maximum effect. The EC50 of the system is the number of microbes present at day 2 that cause a 50% change in MTD. Maximum effect is defined by the sigmoid’s asymptotic tail at high microbe loads. This defines a maximum death rate, suggesting that there is a previously unrecognized rate-limiting step for death.
In the model shown in Fig 1A, depletion of health could be induced either by bacterial damage effectors or indirectly by self-harm caused by the resistance response. We modeled the two possibilities by observing how the shape of the curve changed as we altered the model such that either the resistance response or bacteria were the sole cause of damage. We concentrated on changes in the rate that the immune response was turned on (τ), and the inflection points for the relationships between microbe density and the rates of immune induction and microbe-induced damage induction (σI and σM) (Fig 6A–6D, S4 Table). In the case in which resistance mechanisms drove damage (Fig 6A and 6C) and bacterial damage was set to zero, changes in τ or the inflection point for microbe-induced immunity caused shifts in both the EC50 and microbe loads. The effect is more extreme when the inoculum is far below the microbial carrying capacity, as this gives the immune response an opportunity to control microbe loads. This results in a reciprocal mechanistic link between resistance and tolerance in which one is always high when the other is low. In the model in which bacteria drive damage, a loss of resistance results in high microbe loads and health pegged at maximum severity (Fig 6B). Shifts in the EC50 of the bacterial damage-driven system, in which resistance-induced damage is set to zero, are caused by changes in the inflection point for bacterial-induced damage, as might be expected for differentially virulent strains of microbes or hosts that are better at neutralizing bacterial toxins (Fig 6D). In this second case, there is no mechanistic reciprocal link between resistance and tolerance, and the two vary independently of each other.
To test whether the shape of the disease tolerance curve in this host–pathogen pairing was driven by immunological damage or microbial damage, and to determine how the parameters of a sigmoid disease-tolerance curve varied, we measured the tolerance curves for a collection of Drosophila mutants that we previously showed differ in their resistance and tolerance defenses to L. monocytogenes [10–12]. We also tested natural variant flies from the Drosophila Genetics Reference Panel that were preselected because they showed extreme changes in their response to L. monocytogenes infections [33,34]. In addition, we tested L. monocytogenes mutants that altered pathogenesis in the fly [11]. These data supported the model in which bacteria caused damage in the system and did not support the immunological damage model.
A mutation in the D. melanogaster gene CG2247 was found previously to reduce the fly’s ability to both survive an infection and control L. monocytogenes growth [11]; here, we found that at 2 DPI, even CG2247 mutant flies injected with just ten bacteria had the same high microbe loads as flies injected with 100,000 bacteria, supporting the idea that these mutants alter resistance. This is visible in the tolerance curves, in which all of the points from infected mutant flies clustered on the bottom asymptote of the parental tolerance curve (Fig 6E–6H, S5 Table). An additional example of a fly strain with poor resistance is shown in the supplementary data (S3 Fig and S5 Table). Since flies suffering a loss of resistance showed no change in the maximum death rate of the infection, these results support the model in which bacteria are responsible for damage.
RNAseq analysis of CG2247 mutant flies suggests a molecular mechanism for the observed phenotype. We observed that in wild-type flies, the transcripts encoding the enzymes required to generate a reactive oxygen response against L. monocytogenes dropped over the course of the infection but recovered upon antibiotic treatment (S6 Fig) [11,35]. In contrast, these transcripts dropped to 10-fold lower levels in CG2247 mutants and did not recover to their original levels. We observed a reduction in the characteristic dark spots resulting from the action of this immune response (S6 Fig). We conclude that mutations in this gene disrupt the melanization immune response.
The disease-tolerance curve for the natural variant D. melanogaster strain, RAL 359, showed a reduction in the EC50 of the system without a change in resistance [33]. RAL 359 was as capable as our lab control strain w1118 in controlling L. monocytogenes growth; however, low doses of the microbe had a larger effect on health in this strain than in the control. This shifted the EC50 from 922 to 83 colony-forming units per fly (Fig 6I–6L and S5 Table). The vigor and maximum death rate of these two strains was similar. This shift in EC50 was a common phenotype, as shown in S4 and S5 Figs.
Changing the virulence of the L. monocytogenes also altered the EC50 of the system. Mutant L. monocytogenes lacking the virulence factors actin assembly-inducing protein (ActA) or listeriolysin O (LLO) had previously been shown to kill flies more slowly than wild-type L. monocytogenes (Fig 5C versus Fig 6O) [11]. The tolerance curves for flies infected with an ΔactA mutant showed a shift in the EC50 from 922 to 5,753 (Fig 6M–6P and S5 Table). We observed a ceiling for the number of L. monocytogenes that can be maintained in a wild-type infected fly (Figs 2A and 5A) and hypothesize that this is defined by the number of phagocytes in the fly that provide a niche for L. monocytogenes growth [13]. The ΔactA mutant bacteria approached this limit before they induced maximal pathology and, thus, we did not observe a low asymptotic tail. Δhly mutants produced a similar result to ΔactA mutants, only more extreme (S4 Fig).
As a field, we can generate complex network diagrams showing how the parts of the immune system communicate with each other and identify key signaling pathways and nodes [36,37]. Though this lets us predict that the system will fail when we remove a necessary component of the network, evolution and medicine work by subtler means, for example, by changing rate constants in addition to deleting entire pathways. To predict how evolution or medicine will change an immune response, we need to identify mathematical functions that accurately describe the behavior of that process so that we can understand all of the parameters that control the behavior.
L. monocytogenes growth in the fly follows logistic kinetics; the variation of these curves observed in natural fly variants suggests that microbes are controlled by separable parameters defining the growth rate and carrying capacity of the host [34]. It seems reasonable that this is not a property limited to L. monocytogenes growing in the fly and that it will be true for many pathogens, especially considering that this is a near textbook description of microbial growth [20,38,39]. Both of these properties should be under natural selection, but studies of infections in D. melanogaster typically do not measure microbe growth at all and, when they do, they measure only one time point, conflating growth rates with carrying capacities. The fly is studied because it is easily manipulated and has a relatively simple immune response, but despite decades of work, the field has been underestimating the number of dimensions in which natural variation can change the immune response. Thinking about this generally, it raises the question of whether there might be separate host mechanisms that control microbe growth rates versus ceilings and, thus, diseases might result from a defect in either of these mechanisms. To study these different disease mechanisms, we must ensure that our experiments let us observe the phenomenon.
The AMP response in the fly is under the control of two signaling pathways, Toll and imd [36,40]. The classical descriptions of these pathways argue that the pathways show specificity for different elicitors; a problem with the experiments supporting this model is that the experiments did not correlate response to elicitor levels, and most experiments tested just one arbitrarily determined time point [41,42]. We find that AMP induction is described well by a four-parameter sigmoid function. This means that there are parameters describing the basal level, EC50, maximum expression level, and slope, which may all vary independently of each other; that we don’t see a correlation between maximal gene induction and EC50 supports this hypothesis. If the field’s experiments measure gene expression at just one time point post infection and do not correlate this to microbe loads, we cannot learn which of the three induction parameters are changing. Thus, though it is clear that the Toll and imd signaling pathways, as well as sex, age, the environment, and evolution, can change the immune response, we don’t have a good definition of specificity in this system. Now that we understand that there are new variables to explain, we need to explain this from an evolutionary perspective; for example, what selective pressures cause changes in the EC50 of gene expression of an AMP rather than the maximal gene expression or slope? Will a host benefit from a quick induction of antimicrobials at low microbe loads, or should it invest in an enormous response as microbe loads climb higher?
We found that health was correlated with microbe load according to a four-parameter sigmoid function. As described above for microbe growth and immune induction, past experiments looking at health outputs in model infections are problematic because they don’t measure the full shape of the response curve and don’t report where an experiment is performed on the curve. Perhaps it should have been obvious that microbes would have a growth ceiling and AMPs would have a maximum expression level; classic infectious disease modeling texts start with chapters describing these functions [20,43]. However, one thing these books don’t do is link pathogen growth and the immune response to health. When we examine this relationship, we find that we can make simple models that predict the presence of more parameters than we’ve been studying. For example, the implication of a sigmoid response for health suggests that there is a rate-limiting step for death and that death has a maximum rate. This means that it should be possible not only to change the EC50 of the system to reduce the rate at which damage occurs, but we should also be able to increase the health reservoir of a host so that it has high stamina and will be able to suffer damage for longer periods of time.
The underlying problem with the field’s experimental characterization of immune responses is that we focus on when responses are induced in terms of time and not why they are induced in terms of the quantity of inducer. The result is that we systematically overlook and conflate control mechanisms. By correlating effects to effectors, we can better understand microbial pathogenesis.
The disease space analysis of recovering and dying flies reveals that far more is going on in these sick animals than AMP gene expression, and it lets us organize these events with a simple map. A commonly used time point for gene expression studies in Drosophila is 6 h post infection [44,45]; this is a particularly problematic time point because it combines transient expression of heat shock genes with expression of antimicrobials. Flies fated to die have a progressing gene signature in which a set of transcripts is increased without a corresponding increase in microbe load. We anticipate that one will need to measure the course of pathogenic infections to identify these genes, rather than follow microbes that simply elicit an immune response but do not kill wild-type flies (for example Escherichia coli or Micrococcus luteus [40,46]. Just as each pathogen causes a different disease in humans, we expect that there will be a range of pathologies caused by insect pathogens and do not assume that these L. monocytogenes-induced genes are the only morbidity genes in the fly. The nature of this particular death response is unclear, as the identity of the genes does not provide obvious clues. In contrast, recovering flies have a unique gene expression signature that suggests function; for example, enzymes in biosynthesis and energy metabolism are repressed during infection but pop up hysteretically as microbes are removed, suggesting the induction of a recovery response that rebuilds damaged tissues.
The shapes of tolerance curves are important because they will define which medications can be used to improve the health of an infected patient depending upon their position on the sigmoid curve. Here is a concrete example: If we think about the dose response curve to sepsis in humans as a linear response, then any drug that limits damage will help every patient. We come to a different conclusion if we consider a sigmoid relationship. Humans suffer from sepsis when relatively small numbers of microbes enter the blood, and very sick patients can have much higher levels of microbes. This suggests the sickest patients will be found on the asymptote of a sigmoid curve, where they will be suffering maximum pathological effects. Drugs that change the basal level of a response, EC50, or slope will not help these patients; they will only respond to drugs that manipulate maximal response levels. Thus, we need to understand the shape of these tolerance curves and where the patients lie on the curves to select appropriate treatments.
The discussion of tolerance in patients raises a tricky point. We’ve shown it is possible to measure a disease tolerance curve in a model system; this demonstrated the nonlinear nature of tolerance and suggested the existence of new dimensions we should follow when measuring disease outcomes. The problem is that this approach requires us to perform many more experiments than we do currently when analyzing phenotypes, and this approach is likely impossible to apply to most human infections for the following two reasons: First, disease tolerance is plotted using summary characteristics of an infection. For example, one might measure the maximum parasite load and minimum health for an infection [5]. This is accessible experimentally, but it is unethical to gather these data from a patient, as you need to treat the patient when they enter the clinic. In one rare case of HIV, infected patients’ tolerance curves have been recorded because these chronically infected patients are followed for years, but that approach isn’t going to be useful for acute infections [47]. Second, disease tolerance is a measurement of the behavior of a population and not an individual. The information we gather about an individual allows us to place a datum on a health by microbe graph, but we don’t know the shape of the curve that should be fit through that individual [9]; thus, we can’t easily determine which parameters need fixing in the sick patient. What we learn from model systems is that there are underlying rules that can be used to explain disease outcomes, and that these rules are nonlinear. By studying these rules in models, we can identify the molecular mechanisms corresponding to the mathematical parameters and then apply this knowledge to human biology by analogy. If we want to measure changes directly in human immune infections, we need to rely on descriptions of disease space, which are described in the accompanying paper [48].
Flies were maintained on standard dextrose fly media (129.4 g dextrose, 7.4 g agar, 61.2 g corn meal, 32.4 g yeast, and 2.7 g tegosept per liter) at 25°C with 65% humidity and 12 h light/dark cycles. Shortly after eclosion, adult flies were collected into bottles containing dextrose fly media. At least 24 h post eclosion, adult flies were anesthetized with carbon dioxide, and males were sorted into groups of 20 and placed into vials containing standard dextrose fly media. Experiments were performed on flies 5–7 d post eclosion unless otherwise indicated. CG2247 piggybac allele (BL18050), Pcmt piggybac allele (BL18398), kenny piggyback allele (BL11044), CG7408 piggyback allele (BL19305), piggybac allele parental strain w1118 (BL6326), RAL 359 (BL28179), RAL 787 (BL28231), RAL 375 (BL25188), RAL 309 (BL28166), RAL 73 (BL28131), RAL 380 (BL25190), and RAL 821 (BL28243) strains were obtained from the Bloomington stock center.
Bacteria were injected into flies essentially as described previously [10,12,14,49]. Flies were anaesthetized with carbon dioxide. A drawn glass needle carrying L. monocytogenes was used to pierce the cuticle on the ventrolateral side of the abdomen. A picospritzer III was used to inject 50 nl of liquid into the fly. Bacteria were delivered at different concentrations to produce injections of approximately 10, 100, 1,000, 10,000 or 100,000 CFU. Infectious doses were determined for each experiment by plating a subset of flies at time zero. Approximately 200–400 flies were used for each dose in the experiment to measure survival and to count colonies.
All L. monocytogenes stocks: Wild type/mutant parental strain (10403S), Δhly (DP-L2161) [50], and ΔactA (DP-L3078) [51] were stored at -80°C in brain and heart infusion (BHI) broth containing 25% glycerol. To prepare L. monocytogenes for injection, bacteria were streaked onto Luria Bertani (LB) agar plates containing 100 ug/mL streptomycin and incubated at 37°C overnight. Single colonies of L. monocytogenes from the LB agar plate were used to inoculate 4 mL of brain and heart infusion (BHI) broth and incubated overnight at 37°C without shaking. Bacteria were removed from the incubator at log growth phase. Prior to injection, L. monocytogenes cultures were diluted to the desired optical density (OD) 600 in phosphate buffered saline (PBS) and stored on ice.
Single flies were homogenized in PBS using a motorized plastic pestle in 1.5 ml tubes. The supernatants were plated using an Autoplate spiral plater and counted using a Qcount automated counter. At least six samples were counted to determine the median number of bacteria for each inoculum. Bacteria were plated onto LB medium and incubated overnight at 37°C before counting.
200–400 flies were injected and checked daily to measure mortality for each inoculum. Flies were housed in vials containing 20–25 flies each.
Each set of conditions was repeated at least three times; for example, an experiment for a mutant fly line would be set up independently on three different days to gather microbe load and survival data. Pairs of plating and survival data from these multiple experiments were all plotted on the same tolerance curve.
Flies were injected with 50 nl of Listeria (OD600 = 0.001, or approximately 100 CFUs) or left manipulated. Following injection, flies were placed in vials containing dextrose fly media and incubated at 29°C. Samples were separated into three groups: Moribund, recovering, and uninfected control. Moribund flies were infected, and samples were collected on the following DPI: 0.25, 1, 2, 2.25, 3, 4, 5, and 6. Recovering flies were infected and flipped onto dextrose fly media containing 1 mg/ml ampicillin 2 DPI. Recovering samples were collected on the following DPI: 2.25, 3, 4, 5, 6, 7, 9, and 16. Uninfected control flies were not infected but were flipped onto dextrose fly media containing 1 mg/ml ampicillin 2 dpi. Uninfected control samples were collected on the following dpi: 2.25, 3, and 9. At each of the indicated time points, groups of 20 flies were homogenized in TRIzol, and RNA was isolated using a standard TRIzol preparation. Additional flies were used to determine CFUs and monitor survival. Biological triplicates were obtained from three independent experiments. Quality of RNA was determined using a BioAnalyzer 2100. Samples were labeled using the Quick Amp Labeling Kit, One-Color (Agilent), and hybridized to 4x44K (V2) Drosophila Gene Expression Microarray (Agilent) using the Agilent Gene Expression Hybridization kit following the manufacturer’s protocol. Microarrays were scanned using an Agilent Technologies Scanner, and processed signal intensities were determined using Agilent’s Feature Extraction software. RNA quality assessment, labeling, hybridization, and microarray feature extraction were performed at the Stanford Functional Genomics Facility.
The microarray data were analyzed using Genespring v12.1 (Agilent). Microarrays were normalized to the 75th percentile. The median expression level on 0 dpi was set as the baseline. Prior to statistical analysis, low-quality spots were removed based on flag calls. To determine differential gene expression, one-way ANOVA was performed comparing all samples to 0 dpi. P-values were corrected by the Benjamini-Hochberg method. Genes with p-values <0.05 at any time point with a fold-change greater than 2 were categorized as differentially expressed. Differentially expressed genes were then clustered using Mfuzz [52].
Only genes that were differentially expressed in moribund and recovering but not in uninfected control were used in the topological data analysis using the Ayasdi 3.0 software platform (ayasdi.com, Ayasdi Inc., Menlo Park, California). Nodes in the network represent clusters of samples of infected flies, and edges connect nodes that contain samples in common. Nodes are colored by the average value of their samples for the variables listed in the figure legends. TDA was used to map the way hosts loop through the disease space in an unsupervised fashion. Two types of parameters are needed to generate a topological analysis. First is a measurement/notion of similarity, called a metric, which measures the distance between two points in some space (usually between rows in the data). Second are lenses, which are real valued functions on the data points. Lenses are used to create overlapping bins in the dataset. Overlapping families of intervals are used to create overlapping bins in the data. Metrics are used with lenses to construct the Ayasdi 3.0 output. Multiple lenses can be used in each analysis. There are two parameters used in defining the bins. One is resolution, which determines the number of bins; higher resolution means more bins. The second is gain, which determines the degree of overlap of the intervals. Once the bins are constructed, we perform a clustering step on each bin, using single linkage clustering with a fixed heuristic for the choice of the scale parameter. This heuristic is described in [26]. This gives a family of clusters within the data, which may overlap. We built a network with one node for each such cluster and where we connect two nodes, if the corresponding clusters contain a data point in common. We used two types of lenses. The first type was lenses based on dimension reduction algorithms such as multidimensional scaling and nearest neighbor analyses; these helped analyze the data in an unsupervised manner. The second type was lenses based on the data alone, for example, the levels of antimicrobial gene expression or recovery gene expression. The gene expression markers helped us dissect the graphs using knowledge about the biology of the system. To build our map, we analyzed our dataset with samples as rows and genes as columns. We used the Variance Normalized Euclidean metric. We used the PCA coord1 and PCA coord2 lens as well as two data lenses: CG32373 and Fuca (Resolution = 19, Gain = 6, and equalized was used for all lenses).
To evaluate the impact of within-host infection dynamics on host health curves, we created a compartmental model consisting of four ordinary differential equations for microbes (M), immune effectors (I), microbial damage effectors (D), and health (H).
Microbes grow in a sigmoid fashion at rate r as they deplete the host substrate, and immune effectors can either kill microbes outright at rate ε per effector or inhibit microbial growth at rate η. In these simulations, growth rate limitation was used as the main immune response. Two mechanisms contribute to immune effector production. Reflecting the sigmoid relationship between microbe density and AMP levels, immune effectors can increase in proportion to microbe density up to a maximal rate of τ*kM until effector levels saturate immune pathway machinery (kI). The parameter σI reflects the microbe concentration at the inflection point for immune induction. Immune effectors decay at rate μ. Microbes secrete virulence factors and toxins at a maximal rate φ*kD, modulated by a sigmoidal relationship between damage factor production and microbe density to reflect a process like quorum sensing controls on virulence factor production. As with immunity, there is a carrying capacity imposed on damage effectors. This could reflect either negative feedbacks on further effector production at high effector densities or a bottleneck limiting flux through the system. The parameter σD reflects the microbe density that produces a half-maximal induction rate. Damage effectors degrade at rate ρ. Immunopathology and damage effectors deplete health (H) at rates γ and α, respectively, while hosts can recover health according to a sigmoidal relationship with current health (reflecting the difficulty of achieving recovery at low health due to organ failure and other catastrophe) at rate z. “Death” is called when health dips below 10% of maximum (kH). All simulations were conducted within a parameter space outlined in S4 Table and run in Matlab (v. 7.11.0) using the ode45 solver.
The sigmoidal curves were fit using the four-parameter method in Prism. We tested several models (linear, exponential, logistic, and sigmoid) and picked the model that gave the best adjusted r2. We used the adjusted r2 to account for overfitting. When the curve-fitting program suggested a clearly erroneous result, such as an extremely high top or low bottom, we fixed the top or bottom at the average level for the vigor or for the average of the last three points for the bottom.
Flies were injected with 50 nL of 0.01 OD600 (1000 CFUs) of L. monocytogenes resuspended in PBS. Flies were placed in vials containing dextrose fly media and incubated at 29°C. 3 DPI flies were flipped onto dextrose fly media containing 1 mg/mL ampicillin. Groups of 20 flies were homogenized in TRIzol for RNA extraction on the following DPI: 0, 1, 2, 3, 4, 6, 8, 10, 12, 14, 16, and 18. Additional flies were used to determine CFU and monitor survival. We tested w1118, CG2247, Pcmt, and RAL359 strains. As a control, additional sets of w1118 were left uninfected and flipped onto dextrose fly media containing 1 mg/mL ampicillin 3 dpi, and samples for RNAseq were collected at the indicated time points. RNA was isolated using standard TRIzol preparation. Library preparation and sequencing was performed by the Duke Center for Genomic and Computational Biology. Briefly, quality of RNA was assessed by Bioanalyzer 2100. Polyadenylated RNA was enriched from total RNA. Barcoded TruSeq cDNA Libraries were constructed and quality was assessed by Qubit and Agilent Tapestation. 50 bp single end reads were obtained by sequencing the libraries on an Illumina HiSeq 2000 using a full flow cell. Reads were mapped using STAR, and RPKM for each gene was determined by Cufflinks.
Groups of 20 female flies were anesthetized with carbon dioxide and injected with 50 nl of 0.01 OD600 (~1,000 CFUs) of L. monocytogenes resuspended in PBS. After injection, flies were flipped onto dextrose fly media and incubated at 29°C. At 2 dpi, the percent melanized for each group was determined by anesthetizing the flies with carbon dioxide and scoring each abdomen for presence of melanization spots. For each genotype, eight groups of flies were scored.
|
10.1371/journal.pbio.1001771 | Def1 Promotes the Degradation of Pol3 for Polymerase Exchange to Occur During DNA-Damage–Induced Mutagenesis in Saccharomyces cerevisiae | DNA damages hinder the advance of replication forks because of the inability of the replicative polymerases to synthesize across most DNA lesions. Because stalled replication forks are prone to undergo DNA breakage and recombination that can lead to chromosomal rearrangements and cell death, cells possess different mechanisms to ensure the continuity of replication on damaged templates. Specialized, translesion synthesis (TLS) polymerases can take over synthesis at DNA damage sites. TLS polymerases synthesize DNA with a high error rate and are responsible for damage-induced mutagenesis, so their activity must be strictly regulated. However, the mechanism that allows their replacement of the replicative polymerase is unknown. Here, using protein complex purification and yeast genetic tools, we identify Def1 as a key factor for damage-induced mutagenesis in yeast. In in vivo experiments we demonstrate that upon DNA damage, Def1 promotes the ubiquitylation and subsequent proteasomal degradation of Pol3, the catalytic subunit of the replicative polymerase δ, whereas Pol31 and Pol32, the other two subunits of polymerase δ, are not affected. We also show that purified Pol31 and Pol32 can form a complex with the TLS polymerase Rev1. Our results imply that TLS polymerases carry out DNA lesion bypass only after the Def1-assisted removal of Pol3 from the stalled replication fork.
| DNA damages can lead to the stalling of the cellular replication machinery if not repaired on time, inducing DNA strand breaks, recombination that can result in gross chromosomal rearrangements, even cell death. In order to guard against this outcome, cells have evolved several precautionary mechanisms. One of these involves the activity of special DNA polymerases—known as translesion synthesis (TLS) polymerases. In contrast to the replicative polymerases responsible for faithfully duplicating the genome, these can carry out DNA synthesis even on a damaged template. For that to occur, they have to take over synthesis from the replicative polymerase that is stalled at a DNA lesion. Although this mechanism allows DNA synthesis to proceed, TLS polymerases work with a high error rate even on undamaged DNA, leading to alterations of the original sequence that can result in cancer. Consequently, the exchange between replicative and special polymerases has to be highly regulated, and the details of this are largely unknown. Here we identified Def1—a protein involved in the degradation of RNA polymerase II—as a prerequisite for error-prone DNA synthesis in yeast. We showed that after treating the cells with a DNA damaging agent, Def1 promoted the degradation of the catalytic subunit of the replicative DNA polymerase δ, without affecting the other two subunits of the polymerase. Our data suggest that the special polymerases can take over synthesis only after the catalytic subunit of the replicative polymerase is removed from the stalled fork in a regulated manner. We predict that the other two subunits remain at the fork and participate in TLS together with the special polymerases.
| The stalling of the replication machinery that occurs as a consequence of encountering unrepaired DNA damages is a challenging problem for cells. Stalled replication forks can undergo DNA breakage and recombination that can lead to chromosomal rearrangements and cell death. To ensure survival, cells have evolved different mechanisms that can sustain DNA replication on damaged templates. These so-called DNA damage tolerance or DNA damage bypass processes allow replication to continue on damaged DNA without actually removing the damage. DNA damage tolerance is achieved through two main mechanisms: template switching and translesion synthesis (TLS). Template switching is inherently error-free, as replication continues by using the undamaged nascent sister chromatid as a template for the bypass of the lesion [1], whereas during TLS, specialized polymerases take over the nascent primer end from the replicative polymerase and carry out synthesis opposite the DNA lesion in an error-free or error-prone way [2].
Rad6 and Rad18 are key mediators of DNA damage tolerance in the yeast Saccharomyces cerevisiae [3],[4]. They govern at least three different pathways for the replication of ultraviolet light (UV)-damaged DNA: (1) Rad5-dependent error-free DNA damage bypass, (2) Rad30-dependent error-free TLS, and (3) Rev3-dependent error-prone TLS. Rad6 is an ubiquitin conjugase [5], and it forms a complex in the cell with Rad18, a RING finger ATP-ase with single-stranded DNA-binding activity [6]. Upon UV-treatment, the Rad6–Rad18 ubiquitin–conjugase–ligase complex monoubiquitylates proliferating cell nuclear antigen (PCNA), the essential processivity factor for replicative DNA polymerases, at its lysine-164 residue at the stalled replication fork [7]. Monoubiquitylated PCNA activates the Rev3, and the Rad30-dependent subpathways involving TLS polymerases, whereas further polyubiquitylation of PCNA on the same residue through a lysine-63–linked chain by the Rad5–Mms2–Ubc13 ubiquitin–conjugase–ligase complex activates the Rad5 subpathway [7],[8]. Genetic experiments suggest that the Rad5 branch operates through template switching, where the newly synthesized strand of the undamaged sister duplex serves as a template to bypass the lesion [9]. Rad5, a SWI–SNF family member helicase, most probably directly promotes this process through its fork-reversal activity [10]. The RAD30-encoded DNA polymerase η (Polη) is unique in its ability to efficiently and accurately synthesize through UV-induced cyclobutane pyrimidine dimers [11]. In accordance with its role in the error-free bypass of UV lesions, a defect of Polη in yeast confers an increase in UV-induced mutations, and in humans it causes the cancer-prone syndrome, the variant form of xeroderma pigmentosum [12]–[14]. Besides UV-lesions, Polη can bypass several DNA distorting lesions with varying accuracy [2]. The mutagenic branch involves Rev1 and Rev7, besides Rev3, and the lack of either protein causes immutability [15]. The Rev1 protein is a DNA polymerase with limited ability to insert C residues [16]. Its catalytic activity is dispensable for most induced mutagenesis events, suggesting a mainly structural role for Rev1 [17]. Rev3 together with Rev7 forms DNA polymerase ζ (Polζ) [18]. Rev7 is an accessory protein, whereas Rev3 is the catalytic subunit. Polζ has the ability to efficiently extend from mispaired nucleotides and from nucleotides inserted opposite different DNA lesions [2].
TLS polymerases synthesize DNA with a high error rate and are responsible for introducing mutations into the genome during DNA damage bypass, so their replacement of the replicative polymerase must be tightly regulated. However, our understanding of the polymerase switch at DNA damage sites is elusive. Polη, Rev1, and Polζ were shown to interact with PCNA, and it was suggested that through these interactions TLS polymerases could get access to the replication fork [19]–[21]. Also, the interaction with PCNA was shown to be essential for the in vivo function of all three polymerases. Though PCNA binding can give access to TLS polymerases to the replication fork, the mechanism that allows them to actually take over DNA synthesis from the replicative polymerase during DNA lesion bypass is still unknown.
In this study, we identify Def1 as an indispensable regulator of induced mutagenesis. We show that Def1 promotes the ubiquitylation and subsequent proteasomal degradation of the catalytic subunit of the replicative polymerase after DNA damage treatment. We demonstrate that the noncatalytic subunits of the replicative polymerase are not affected by UV-induced degradation and that they can form a complex with the TLS polymerase Rev1. Based on our results we propose a new model for polymerase exchange at stalled replication forks.
In searching for new factors affecting DNA damage tolerance, we aimed to identify new interacting partners of Rad5. Therefore, we performed tandem affinity purification (TAP) of Rad5 together with its complexes. For that purpose, we introduced a TAP tag, consisting of a calmoduline binding peptide and two IgG binding domains of protein A separated by a TEV protease cleavage site, at the C-terminus of Rad5 at the chromosomal locus. We purified Rad5 and its interacting partners through the two affinity tags under native conditions. To facilitate the formation of damage bypass complexes, we applied 0.02% methyl methanesulfonate (MMS) for 2 h before collecting the cells. Surprisingly, without treatment only one prominent, specific band was visible in the final, highly purified fraction on the Coomassie-stained gel (Figure 1), which was identified by mass spectrometry as the tagged Rad5 itself. However, after MMS treatment two prominent bands appeared on the gel; the lower mobility band was again Rad5, whereas the higher mobility band was identified as Def1. Repetition of the experiment yielded the same result that Def1 copurified with Rad5, but only after treating the cells with MMS. It suggested that either DNA-damage–induced posttranslational modification of Rad5 and/or Def1 or a third, damage-specific factor was necessary to promote the formation of the complex.
Rad5 mediates an error-free DNA damage tolerance pathway under the control of Rad6–Rad18, whereas Def1 has a role in promoting the proteolytic degradation of stalled RNA polymerase II and in telomere maintenance [22],[23]. To further establish a connection between Rad5 and Def1, we analyzed the genetic relations between DEF1 and RAD5, and members of all three branches of the RAD6-governed pathway upon DNA damage. After treating the cells with UV, the sensitivity of the def1 rad6 double deletion strain did not exceed that of the rad6 mutant, pointing to an epistatic relationship between DEF1 and RAD6 (Figure 2A). The higher resistance seen with the double mutant might originate from other functions of these multitask proteins, as the def1 rad18 strain showed the same sensitivity as rad18 (unpublished data), fortifying the involvement of DEF1 in the RAD6-dependent DNA damage tolerance. Surprisingly, the def1 rad5 double deletion strain displayed much higher sensitivity than any of the corresponding single mutants, indicating that DEF1 acted outside of the RAD5-dependent subpathway (Figures 2B). This was verified by the hypersensitivity of the def1 mms2 strain over the single mutants (Figure 2C). Also, the def1 rad30 double mutant was more sensitive to UV than either def1 or rad30, implying that DEF1 functioned independently of RAD30 (Figure 2D). Nevertheless, the def1 rev3 strain exhibited the same sensitivity as the def1 mutant, which indicated an epistatic relationship between DEF1 and REV3 (Figure 2E). We carried out similar experiments using MMS instead of UV as a DNA damage source (Figures 2F–J). Upon MMS treatment DEF1 showed epistasis with RAD6 and REV3, but its deletion further sensitised rad5 and mms2, proving again the involvement of DEF1 in the mutagenic branch of the RAD6-governed DNA damage tolerance. However, DEF1 also showed epistasis with RAD30, as the double mutant was as sensitive as the def1 single mutant. We note that this reflected a real epistatic relationship, as although rad30 itself was not sensitive to MMS, only at very high doses, it was hypersensitive with mms2, but also showed epistasis with REV3 (Figure 2K,L). That means that in the bypass of MMS-induced DNA lesions, RAD30 works together with the members of the REV3 branch. In conclusion, our data strongly suggested that DEF1 participated in the REV3-dependent mutagenic branch of the RAD6–RAD18-regulated DNA damage tolerance.
The TLS polymerases of the REV3 branch are responsible for virtually all damage-induced mutagenesis; consequently, inactivation of either one causes a strong antimutator effect [15]. To prove that DEF1 belonged to the REV3 branch, we measured the rate of UV-induced mutations in def1 strains. In keeping with the results of the epistasis analysis, induced mutagenesis was completely abolished in def1 (Figure 3). In fact, def1 was even more defective than the rev3 strain. Additional deletion of DEF1 in mms2 also eliminated induced mutagenesis, though mms2 by itself causes high mutagenesis, most probably because in the absence of the error-free branch, lesions are channelled to the REV3-dependent mutagenic pathway. Ectopic expression of Def1 in def1 cells restored close to wild-type–level mutagenesis, confirming that the immutability was in fact due to the absence of DEF1. We obtained the same results using MMS instead of UV (unpublished data). From these we concluded that DEF1 played an essential role in induced mutagenesis.
For the REV3 branch to operate, the TLS polymerases of the branch have to take over synthesis from the replicative polymerase stalled at a DNA lesion site, a central but poorly understood step in DNA lesion bypass. Because Def1, unlike other members of the REV3 branch, is not a DNA polymerase, we surmised that it might facilitate the exchange between the TLS and the replicative polymerases. As Def1 played a role in the ubiquitylation of stalled RNA polymerase II [22], we considered the possibility that, similarly, it could mediate ubiquitylation of the stalled replicative DNA polymerase. Ubiquitylation then could lead to polymerase switch by either playing a regulatory role as in the case of DNA-damage–induced ubiquitylation of PCNA [7], or it could result in protein removal through degradation. To test these possibilities, we followed the fate of the replicative polymerase during DNA damage bypass by monitoring Pol3, the catalytic subunit of the replicative DNA polymerase δ (Polδ) during cell cycle in UV-treated, synchronized yeast cultures. In order to facilitate TLS, we first used an mms2 deletion strain. Importantly, we observed a transient decrease in the level of Pol3 upon UV irradiation as opposed to normal growth conditions, and the degree of degradation correlated with the applied UV doses (Figure 4). We could also detect degradation in wild-type cells in the S phase of the cell cycle, as indicated by the expression pattern of the G2/M-specific cyclin Clb2 (Figure 5A). However, in experiments using a def1 deletion strain, we could not detect any decrease in the level of Pol3 (Figure 5B). To investigate whether the observed phenomenon was ultimately under the higher control of RAD6, we performed the same experiment in a rad6 strain and found that Pol3 diminution was also absent (Figure 5C). On the other hand, reduction of Pol3 could be seen in mms2 (Figure 5D) and in rad30 (Figure 5E) backgrounds. These results, in conjunction with the above genetic results, strongly implied that Pol3 diminution was specifically dependent on DEF1, which exerted this function under the control of RAD6, in the mutagenic DNA damage bypass pathway.
The most plausible explanation for the transient decrease of Pol3 would be that Pol3 underwent regulated protein degradation induced by UV. The majority of regulated proteolysis takes place in the proteasome in eukaryotic cells. To resolve whether the decrease in the Pol3 protein level was due to protein degradation mediated by the proteasome, we supplemented the growth media with the proteasome inhibitor MG132. Indeed, in the presence of MG132, the UV-induced degradation of Pol3 could not be observed (Figure 5D). To add further evidence, we applied a temperature-sensitive rpn7 mutant displaying defects in proteasome function at high temperature (37°C) but behaving like wild-type at low temperature (25°C) [24]. Using this mutant we could not detect degradation at the restrictive high temperature contrary to the permissive low temperature (Figure 6A), whereas in the RPN7 strain degradation occurred at both temperatures (unpublished data). These results demonstrated that the proteasome was responsible for the UV-induced degradation of Pol3.
Ubiquitylation is a major signal for proteasomal protein degradation. To show ubiquitylation of Pol3, N-terminally 7 histidine-tagged ubiquitin was expressed in yeast cells and ubiquitylated proteins from cell extracts prepared after irradiating cells with UV were enriched on nickel beads. Indeed, we could detect polyubiquitylated forms of Pol3 upon UV irradiation in wild-type cells, but not in def1 and rad6 cells (Figure 7B and unpublished data).
Polδ is a heterotrimer and consists of two noncatalytic subunits, Pol31 and Pol32, besides Pol3 [25]. Pol31, like Pol3, is essential for cell viability, but Pol32 is a nonessential subunit. Pol3 forms a stable complex with Pol31, and Pol32 is attached to this complex through its interaction with Pol31 [26]. We aimed to examine whether the whole Polδ enzyme was subject to UV-induced proteolysis, or it affected only the catalytic subunit. We found that contrary to Pol3, Pol31 and Pol32 were not affected by UV-induced degradation (Figure 5F and 5G).
Taken together, these results suggested that during DNA damage bypass, Pol31 and Pol32 remained at the stalled fork. We postulated that a TLS polymerase could take the place of Pol3 and carry out lesion bypass in complex with Pol31 and Pol32. To test this idea, we examined whether Pol31 and Pol32 together could form a complex with Rev1 in in vitro assays using purified proteins. We chose Rev1, because it had been suggested to function as a scaffold in TLS, based on its interaction in yeast with Polη and Polζ [27],[28], and in mouse and human cells with Polη, Polι, and Polκ [29],[30]. Also, it has already been shown to interact with Pol32 [31]. In GST pull-down assays we added Pol31 and Rev1 to GST–Pol32 immobilised on glutathione–Sepharose affinity beads, and after incubation bound proteins were released from the beads by glutathione. As shown in Figure 7B lanes 1–4, both Pol31 and Rev1 eluted together with GST–Pol32, indicating that these proteins formed a complex together. In control experiments using GST instead of GST–Pol32, only GST was present in the elution fraction, confirming that the interaction between Pol31, Pol32, and Rev1 was specific (Figure 7B, lanes 5–8). In conclusion, purified Pol31, Pol32, and Rev1 could interact directly and form a stable multisubunit protein complex.
In this study we identified a DNA-damage–induced complex of Rad5 with Def1. Our genetic studies placed DEF1 in the RAD6–RAD18-dependent DNA damage tolerance pathway, where it played an indispensible role during induced mutagenesis. We established that Pol3, the catalytic subunit of the replicative DNA polymerase Polδ, was degraded upon UV irradiation. We presented evidence that degradation of Pol3 was the result of polyubiqitylation-mediated proteasomal degradation, and it was dependent on DEF1 under the higher control of RAD6. Conversely, Pol31 and Pol32, the other two subunits of Polδ, were not degraded. We also demonstrated that Pol31 and Pol32 together could form a stable complex with the TLS polymerase Rev1. Based on these results, we propose a new model for polymerase exchange at stalled replication forks (Figure 8). During replication, when Polδ stalls at a DNA lesion, PCNA gets ubiquitylated by Rad6–Rad18. Monoubiquitylated PCNA activates TLS, for which to occur Pol3 is ubiquitylated by a Def1-dependent manner and removed from the stalled Polδ complex through proteasomal degradation. We hypothesize that a TLS polymerase takes over the place of Pol3 and teams up with the remaining Polδ subunits, Pol31 and Pol32, at the stalled fork to form a new complex capable of executing DNA lesion bypass, suggested by our results showing complex formation of Rev1–Pol31–Pol32, and also by recent studies detecting complex formation of yeast Pol31–Pol32–Rev3–Rev7 and also of their human counterparts [32]–[34]. We surmise that after lesion bypass and deubiquitylation of PCNA, the TLS polymerase is removed from the primer terminus, Pol3 restores Polδ by regaining its place, and replication continues. Importantly, this model gives an explanation for previous genetic results showing that in pol32 cells induced mutagenesis is severely impaired, and that deletion of the N-terminal part of Pol32, responsible for binding Pol31, also abolishes induced mutagenesis [25],[35],[36].
Our data raise an important question: How can the RAD30-encoded TLS polymerase, Polη, operate independently of Def1? Our results imply that Pol3 does not have to be removed from the stalled fork for Polη-dependent UV-lesion bypass to occur. Polη is mainly specialized for the error-free bypass of UV lesions, so it is reasonable to assume that Polη should have preference over the error-prone TLS polymerases in the bypass of UV-induced DNA damages. Polη differs from the other TLS polymerases, Rev1 and Polζ, in its way of binding PCNA. Although Rev1 and Polζ bind the intermolecular interface at the outer face of the PCNA ring [20],[21], Polη, similarly to Polδ, binds the interdomain connector loop of PCNA through its conserved PCNA-interacting peptide motif [19]. Given that PCNA is a homotrimer ring, Polδ and Polη could bind the same PCNA ring simultaneously. We presume that transient conformational changes, probably induced by the stalling of the fork and ubiquitylation of PCNA, could allow Polη to take over synthesis from Polδ, as also suggested by in vitro experiments [37], while both remain attached to PCNA. Because Polη synthesizes opposite pyrimidine dimers with the same kinetics as it does opposite undamaged DNA [38], rapid bypass can occur. Deubiquitylation of PCNA would restore the original conformation and Polδ could continue synthesis. We note that this is in accord with the in vivo finding that Pol32 is not needed for TT dimer bypass carried out by Polη [39]. On the other hand, when the damage poses a kinetic barrier also to the TLS polymerase, for the slower kinetic damage bypass to occur, Pol3 has to be removed so that the TLS polymerases could form a stable complex with Pol31 and Pol32. This would also explain the epistasis of RAD30 with DEF1 in the bypass of MMS-induced DNA lesions, as the efficiency of incorporation by Polη is reduced ∼20-fold opposite 6O-methylguanine and ∼1,000-fold opposite an abasic site [40],[41].
We detected a very stable DNA-damage–induced complex formation of Rad5 with Def1. However, our genetic data placed the two genes into two alternative DNA damage tolerance pathways, both governed by Rad6. We hypothesize that the Def1–Rad5 complex might coordinate the activity of the two subpathways in response to DNA damage. A similar role of Def1 during transcription was suggested, where Def1 assisted in the degradation of the RNA polymerase stalled at DNA damage sites, and probably coordinated the repair mechanisms through its interaction with Rad26 [22].
The high conservation between elements of DNA lesion bypass from yeasts to humans, including the Rad6–Rad18 and Rad5–Mms2–Ubc13 complexes and their enzymatic activities, the TLS polymerases, and PCNA ubiquitylation [42], suggests that DNA-damage–induced selective degradation of the catalytic subunit of the replicative DNA polymerase drives polymerase exchange in higher eukaryotes as well. The role of TLS polymerases in mutagenesis and in cancer makes it highly important to uncover further details of polymerase exchange, to identify and investigate further factors that affect Pol3 degradation, and to check the existence of a similar mechanism in human cells.
The wild-type strain (BY4741) and its single deletion derivatives for the genetic studies were obtained from the Euroscarf collection. Chromosomally C-terminally tagged POL3, POL31, and POL32 with three copies of the hemagglutinin epitope tag (3-HA) were created by a PCR-based strategy [43] in EMY74.7 (MATa, his3-Δ1, leu2–3,–112, trp1Δ, ura3–52) strain, made bar1Δ. Additional deletions were generated by gene replacement. RAD5 was TAP tagged in BJ5464 by the same PCR-based strategy using pBS1539 [44]. BJ5464 was also used for protein overexpression. The rpn7-3 mutant and its corresponding W303 wild-type strain [24] were used in experiments showing the effect of temperature-sensitive inhibition of the proteasome. Polyubiquitylation of Pol3 was shown in MHY500 strain background [45]. For complementation in yeast, Def1 was expressed from the centromeric vector pID394 (p416ADH backbone [46]). For protein purification, Pol31, Pol32, and Rev1 were overexpressed in N-terminal GST fusion from pID370, pID458, and pID460, respectively (pBJ842 backbone [47]). In the plasmid pRS426–pCUP1–His7–Ubiquitin (G76A) [48], the mutation was reversed by site-directed mutagenesis, resulting in plasmid pID198.
For qualitative analysis of sensitivity to MMS, cells were serial diluted and spotted onto YPD plates containing defined amounts of MMS and grown at 30°C for 3–5 d. For quantification, cells were spread onto YPD plates at appropriate dilutions and irradiated with UV light (254 nm) for varying times to apply the specified dosage. Plates were incubated in the dark at 30°C, and colonies were counted after 3–5 d.
UV-induced forward mutation frequencies at the CAN1 locus were measured by comparing the numbers of can1R colonies at given UV doses, selected on synthetic complete medium without arginine and containing canavanine, with the numbers of survivors on complete synthetic medium, exposed to the same UV doses.
Logarithmically growing cells in YPD at 30°C were arrested at A600∶0.8 in G1 by 100 ng/µl α-factor (Sigma) for 2–4 h, washed, resuspended in phosphate buffered saline, and divided into Petri dishes for UV irradiation. Half of the cultures were irradiated with the given UV dose, and the other half served as untreated control. Cells were released back into growth medium containing 50 µg/ml pronase (Calbiochem) to inactivate any residual α-factor. For experiments showing polyubiquitylation of Pol3, the growth media always contained 100 µM CuSO4 to induce 7His–ubiquitin expression. Samples were taken at given time points after UV treatment for whole cell extract preparation. Experiments involving MG132 (Sigma) were done in pdr5 background. MG132 (50 µM) was added to the α-factor synchronized cultures 1 h before UV irradiation. The rpn7-3 mutant and its isogenic wild-type strain were grown at 25°C. To detect the mutant phenotype we followed the protocol described in [24]. Briefly, 50 ml culture of logarithmically growing cells (A600∶0.5) were split. Half of the culture was kept at 25°C and the other half was shifted to 37°C. At A600∶0.8 cultures were synchronised by α-factor for 3 h and processed as detailed above.
Whole cell extracts were prepared according to a trichloroacetic acid (TCA) protein precipitation method [43] except that after TCA precipitation, pellets were washed with ice-cold acetone, air-dried, and resuspended in 1× Laemmli sample buffer before loading to an 8% poly-acrylamide gel. Polyubiquitylated Pol3 was detected using denaturing NiNTA chromatography as described in [49]. Antibodies against HA (Gene Tex), Clb2 (Santa Cruz), PGK (Molecular Probes), and ubiquitin (Santa Cruz) were used. Pol31, Pol32, and Rev1 were overexpressed in N-terminal fusion with GST and purified on glutathione–Sepharose 4B beads following the protocol in [45], with the exception that in the case of Rev1, 0.1% Triton X-100 was added to the lysate after breaking the cells. In the case of Pol31 and Rev1, the GST tag was removed by PreScission protease cleavage in the elution step of purification. For complex formation, GST–Pol32 (3 µg) immobilized on glutathione–Sepharose beads was incubated with purified Pol31 (5 µg) and Rev1 (3 µg), overnight on ice in buffer containing 50 mM Tris/HCl, pH 7.5, 150 mM NaCl, 1 mM EDTA, 1 mM DTT, 10% glycerol, 0.01% Nonidet P-40. Beads were washed five times with the same buffer, and bound proteins were eluted in the same buffer containing 20 mM reduced glutathione. Various fractions were analyzed by SDS/PAGE.
Four liters of yeast culture were grown to logarithmic phase in synthetic complete medium, and at A600∶0.8, half of the culture was treated with 0.02% MMS for 2 h before harvesting. TAP purification was carried out as described [44] with the following modifications: cells were broken in 1× YBB (50 mM Tris/HCl pH:7.5, 50 mM KCl, 100 mM NaCl, 10% sucrose) supplemented with protease inhibitors. After clarifying the lysate with ultracentrifugation for 1 h with 100,000 g, 2-mercaptoethanol was added to 8.5 mM, Nonidet P-40 to 0.01%, and NaCl to 500 mM final concentration, and the lysate was transferred into an IgG Sepharose bead (Amersham)–filled column. In later steps the protocol was followed. Briefly, bound fraction was eluted with TEV protease cleavage. The elution fraction was applied on calmodulin beads, and bound proteins were recovered by eluting in EGTA-containing buffer. Proteins were concentrated and analysed on a 6%–12% gradient sodium dodecyl sulphate polyacrylamide gel stained with Coomassie blue R-250. Excised protein bands were identified by MALDI-TOF mass spectrometry after trypsin digestion. Eleven peptides of the higher mobility band matched yeast Def1 (55% coverage) and they covered 18% of the DEF1 sequence.
|
10.1371/journal.ppat.1003047 | Neutrophil-derived IL-1β Is Sufficient for Abscess Formation in Immunity against Staphylococcus aureus in Mice | Neutrophil abscess formation is critical in innate immunity against many pathogens. Here, the mechanism of neutrophil abscess formation was investigated using a mouse model of Staphylococcus aureus cutaneous infection. Gene expression analysis and in vivo multispectral noninvasive imaging during the S. aureus infection revealed a strong functional and temporal association between neutrophil recruitment and IL-1β/IL-1R activation. Unexpectedly, neutrophils but not monocytes/macrophages or other MHCII-expressing antigen presenting cells were the predominant source of IL-1β at the site of infection. Furthermore, neutrophil-derived IL-1β was essential for host defense since adoptive transfer of IL-1β-expressing neutrophils was sufficient to restore the impaired neutrophil abscess formation in S. aureus-infected IL-1β-deficient mice. S. aureus-induced IL-1β production by neutrophils required TLR2, NOD2, FPR1 and the ASC/NLRP3 inflammasome in an α-toxin-dependent mechanism. Taken together, IL-1β and neutrophil abscess formation during an infection are functionally, temporally and spatially linked as a consequence of direct IL-1β production by neutrophils.
| Invasive infections caused by the human pathogen Staphylococcus aureus result in more deaths annually than infections caused by any other single infectious agent in the United States. Although neutrophil recruitment and abscess formation is crucial for effective host defense against this pathogen, how neutrophils sense and mount an inflammatory response are not completely clear. Using gene expression analysis and in vivo bioluminescence and fluorescence imaging, we found that neutrophil recruitment during a S. aureus cutaneous infection is functionally and temporally linked to IL-1β/IL-1R activation. Surprisingly, neutrophils themselves were determined to be the most abundant cell type that produced IL-1β during infection. Further, neutrophil-derived IL-1β, in the absence of other cellular sources of IL-1β, was sufficient for neutrophil recruitment, abscess formation, and bacterial clearance. Finally, mouse neutrophils produced IL-1β in direct response to live S. aureus in vitro. These findings expand our understanding of the acute neutrophil response to infection in which early recruited neutrophils serve as a source of IL-1β that is essential for amplifying and sustaining the neutrophilic response to promote abscess formation and bacterial clearance. Therapies aimed at promoting IL-1β production by neutrophils may be an effective immunotherapeutic strategy to control S. aureus infections.
| Neutrophil abscess formation represents an important component of the innate immune response, which helps control the spread of an invading pathogen into deeper tissues and systemically [1]. At the site of infection, neutrophils primarily function through the phagocytosis of microorganisms and utilize a variety of antimicrobial mechanisms to mediate pathogen killing [2].
To investigate mechanisms that promote neutrophil recruitment and abscess formation, we chose to use S. aureus cutaneous infection as a model [3]. This gram-positive extracellular bacterium is responsible for the vast majority of skin and soft tissue infections in humans and is a common cause of invasive and often life-threatening infections such as bacteremia, abscesses of various organs, septic arthritis, osteomyelitis, endocarditis, pneumonia and sepsis [4], [5]. S. aureus infection serves as an excellent model system to study neutrophil recruitment since neutrophil abscess formation is required for bacterial clearance in a variety of mouse models of S. aureus infection, including cutaneous infection, bacteremia, septic arthritis and brain abscesses [6]–[8]. The critical role of neutrophils in host defense against S. aureus is also seen in humans, since patients with genetic or acquired conditions with defective neutrophil number or function suffer from recurrent and invasive S. aureus infections in various tissues and organs, including the skin [9].
It is well established that IL-1β plays a central role in initiating the neutrophilic response against S. aureus infections [7], [10], [11]. This is mediated by IL-1β activation of IL-1R/MyD88 signaling, which triggers NF-κB and other signaling molecules that induce proinflammatory mediators and chemokines to promote neutrophil trafficking from the circulation into the infected tissue [9]. Given this essential function of IL-1β, there has been intense interest in understanding how its production is triggered during an infection in vivo. On a cellular level, two different signals are required for IL-1β production in response to S. aureus. The first is production of pro-IL-1β, which is in part mediated by activation of pattern recognition receptors (PRRs) such as TLR2, a cell surface PRR that recognizes S. aureus lipopeptides and lipoteichoic acid [12], [13], and NOD2, a cytoplasmic PRR that recognizes muramyl dipeptide, which is a breakdown product of S. aureus peptidoglycan [14], [15]. The second signal is the triggering of the NLRP3 inflammasome to induce caspase-1 activation and subsequent cleavage of pro-IL-1β into mature IL-1β, the active and secreted cytokine [16]–[18].
The mechanism for inducing IL-1β-dependent neutrophil recruitment in infected tissues in vivo is complex and involves interactions among epithelial cells, stromal cells, resident immune cells, endothelial cells and recruited immune cells. It is known that IL-1β produced at the site of S. aureus skin infection promotes neutrophil recruitment by inducing neutrophil-attracting chemokines and granulopoiesis factors directly via activating IL-1R/MyD88-signaling and indirectly through the production of IL-17 by T cells [3], [11], [19]. A key question is which cell types are responsible for IL-1β production during a S. aureus infection in vivo and how these cells utilize PRRs and the inflammasome to induce its production. The precise mechanism is particularly relevant since many different cells, including keratinocytes, mast cells, Langerhans cells, dendritic cells and monocytes/macrophages, can produce IL-1β in various in vivo models of skin inflammation and infection [20]–[23]. In the current study, we used gene expression analysis and noninvasive in vivo imaging to determine the functional and temporal kinetics, cellular sources and mechanisms by which IL-1β induces neutrophil abscess formation during a S. aureus skin infection.
Mice deficient in IL-1β, IL-1R or MyD88, but not IL-1α, exhibit a severe impairment in neutrophil abscess formation at the site of infection in this mouse model of S. aureus intradermal infection [3], [11], indicating that IL-1β is the major cytokine that initiates the IL-1R/MyD88-dependent pathway for neutrophil recruitment. Given these results and evidence from humans and mice that neutrophils are essential for clearance of S. aureus infections [9], gene expression analysis was performed in an attempt to link IL-1β/IL-1R-dependent gene induction with neutrophil recruitment. To accomplish this, we used a model of S. aureus cutaneous infection in mice, which involves intradermal inoculation of S. aureus (2×106 CFUs of strain SH1000) in the dorsal back skin of mice [3]. Gene expression analysis was first performed on skin biopsy samples from wt and IL-1R-deficient mice at 4 hrs after S. aureus infection and from uninfected skin. This time point was chosen because we previously observed substantially decreased IL-1β protein levels in S. aureus-infected skin of IL-1R-deficient mice compared with wt mice at 6 hrs after infection [3] and the difference in mRNA levels of IL-1β likely preceded the changes in IL-1β protein levels. By using the criteria that upregulated genes were >1.5-fold higher (p-value<0.05) than baseline, there were 1,288 genes upregulated in wt mice and 606 genes upregulated in IL-1R-deficient mice (Fig. S1). Comparing S. aureus-infected skin of wt mice versus uninfected skin, the top 4 induced genes were neutrophil-attracting CXC chemokines, including CXCL1 (KC), CXCL2 (MIP2α), CXCL3 (MIP2β) and CXCL5 (LIX) (Fig. 1A), which bind to CXCR2 on mouse neutrophils to induce chemotaxis [24]. The neutrophil granulopoiesis factors G-CSF and GM-CSF, as well as the proinflammatory cytokines IL-1β, IL-6 and the inflammasome component NLRP3, were also among the top 20 genes. These data indicate that many of the most highly induced genes in S. aureus-infected wt mice were associated with neutrophil chemotaxis, granulopoiesis and IL-1β production.
Pathway analysis (Ingenuity) was then used to categorize the genes upregulated in S. aureus-infected wt versus IL-1R-deficient mice into functional groups. In wt mice, the most significantly upregulated functional pathway was the Cellular Movement group, which included 354 genes (p = 3×10−57) (Fig. 1B). In contrast, in IL-1R-deficient mice, the Cellular Movement functional group included only 256 genes and had lower statistical significance (p = 2×10−25). Categorization of the Cellular Movement functional group by cell type revealed that in wt mice, the Cell Movement of Neutrophils sub-group included 77 genes and was the most statistically significant (p = 7×10−33). Neutrophils were followed in order of number of genes and significance by macrophages (41 genes; p = 2×10−15), T cells (41 genes; p = 9×10−15) and dendritic cells (25 genes; p = 3×10−12). In contrast, in IL-1R-deficient mice, the categorization of the Cell Movement group by cell type was radically different with less numbers of upregulated genes. These differences were most evident for neutrophils, (44 genes; p = 1×10−10), followed by macrophages (28 genes; p = 4×10−9), T cells (29 genes; p = 7×10−9) and dendritic cells (10 genes; p = 1×10−5).
The level of induction of upregulated genes in the Cell Movement of Neutrophils group was then compared between wt and IL-1R-deficient mice using functional network analysis (Fig. 1D). All 77 genes associated with Cell Movement of Neutrophils were significantly upregulated in wt mice. In contrast, over half (45 of the 77 genes), including CXCL2, GM-CSF and IL-1β, were not significantly induced in IL-1R-deficient mice. To confirm these results, the expression of a subset of genes of interest in the Cell Movement of Neutrophils sub-group (Fig. 1D) were evaluated by quantitative real-time PCR (Q-PCR), which included 2 genes that were similarly-induced (IL-6 and G-CSF) and 8 genes that were differentially-induced in wt and IL-1R-deficient mice (MIP-2β, MIP-2α, IL-1β, GM-CSF, CXCR2, G-CSF-R, TLR2 and LTB4R) (Fig. 1E). In agreement with the microarray data (Fig. 1D), we found that the expression of IL-6 and G-CSF between wt and IL-1R deficient mice was similar and the differentially-induced genes had higher expression in wt mice compared with IL-1R-deficient mice. Taken together, these results indicate that genes associated with neutrophil recruitment were upregulated in response to S. aureus skin infection and that the induction of these genes was largely dependent on IL-1β/IL-1R signaling.
To differentiate between the specific anti-S. aureus response versus the response to a live infection, the skin of wt and IL-1R-deficient mice was inoculated with either live or heat-killed S. aureus and Q-PCR was performed on skin samples taken at 4 hrs after inoculation and from uninfected mice (Fig. S2). We found that the top six genes on the microarray were highly expressed in skin samples infected with live S. aureus (ranging from 242- to 5696-fold) or heat-killed S. aureus (ranging from 46- to 1706-fold); however, the level of induction with heat-killed S. aureus was generally a magnitude less (decreased from 4.5-fold to 14.1-fold) than live S. aureus. To determine whether the pattern of gene induction in wt and IL-1R-deficient mice was similar or different in response to live versus heat-killed S. aureus, the same subset of genes of interest in the Cell Movement of Neutrophils sub-group in Fig. 1E was compared (Fig. S2B, C). Live and heat-killed S. aureus had similar expression of IL-6 and G-CSF between wt and IL-1R deficient mice and the differentially-induced genes had higher expression in wt mice compared with IL-1R-deficient mice. However, as with the top 6 induced genes (Fig. S2A), the levels of induction of these genes were lower with heat-killed S. aureus compared with live S. aureus. In summary, the difference in gene expression patterns between wt and IL-1R-deficient mice was consistent between live and heat-killed S. aureus (albeit the live S. aureus resulted in higher gene expression than heat-killed S. aureus), demonstrating that the immune response is more intense in the presence of the live bacterial infection. These results suggest that S. aureus-specific immune responses are significantly impaired in IL-1R-deficient mice. However, we cannot exclude the possibility that the impaired immune responses in the IL-1R-deficient mice were due to a dysregulated immune response in IL-1R-deficient mice in which any stimulus would elicit an aberrant response with a similar pattern of gene expression.
Although IL-1β is the major cytokine that induces IL-1R/MyD88-dependent neutrophil recruitment during a S. aureus skin infection [11], the source and kinetics of IL-1β production during an infection in vivo has remained unknown. Therefore, advanced techniques of in vivo bioluminescence and fluorescence imaging were combined to provide an approximation of the kinetics of IL-1β production and neutrophil recruitment longitudinally over the time course of the S. aureus cutaneous infection. This was accomplished by performing the intradermal inoculation of a bioluminescent S. aureus strain in two fluorescence reporter mouse strains: (1) pIL1-DsRed transgenic mice, which express the red fluorescent protein DsRed under the control of the mouse IL-1β promoter [21], and (2) LysEGFP mice, which possess green fluorescent myeloid cells (mostly neutrophils) due to a knock-in of the EGFP gene into the lysozyme M locus [25]. In vivo noninvasive whole animal imaging was then used to track the S. aureus bacterial burden while simultaneously monitoring IL-1β production or neutrophil recruitment in the same anesthetized mice over the 14 day course of infection. Advantages and limitations of using this strategy of in vivo imaging to quantify these endpoints are described in the Discussion.
Infection with S. aureus resulted in the development of visible skin lesions, which had a maximum size of 0.53±0.11 cm2 by day 3, and healed by day 14 (Fig. 2A). In vivo bioluminescence signals, which closely estimate the bacterial CFUs harvested from the skin lesions during infection [3], [19], peaked on day 1 (up to 1.8±0.7×106 photons/s) and slowly decreased to background levels by day 14 (Fig. 2B). IL-1β-DsRed and EGFP-neutrophil fluorescence signals were significantly higher than uninfected control mice at all time points, peaking on days 3 and 1, respectively (up to 2.7±0.5×1010 and 1.6±0.3×1010 [photons/s]/[µW/cm2], respectively) and decreased to background levels by day 14 (Fig. 2C). In summary, IL-1β-DsRed fluorescence and EGFP-neutrophil fluorescence signals had similar temporal kinetics as they both increased rapidly by day 1 and then decreased along with the in vivo bioluminescent signals over the 14 day course of infection.
Since many different cell types, including keratinocytes, mast cells, Langerhans cells, dendritic cells and macrophages, have the capacity to produce IL-1β in various in vivo models of skin inflammation and infection [20]–[23], it is unclear which of these cell types (or potentially other cell types) contribute to IL-1β production during a S. aureus skin infection. However, in our previous work using the same S. aureus skin infection model in bone marrow chimeric mice, we found that the source of IL-1β was from bone marrow-derived hematopoietic cells because neutrophil recruitment, host-defense and IL-1β production at the site of infection was restored in IL-1β-deficient mice reconstituted with bone marrow from wt mice but not in wt mice reconstituted with bone marrow from IL-1β-deficient mice [11]. Since in vivo fluorescence imaging demonstrated that IL-1β production was detected shortly after infection, histological evaluation of skin lesions at 4 and 24 hrs after S. aureus skin infection in pIL1-DsRed mice was performed. To identify the number of cells that expressed pro-IL-1β, two-color immunofluorescence labeling and confocal laser microscopy was performed using an antibody against DsRed, which is retained within the cytoplasm of IL-1β-expressing cells [21] in combination with mAbs directed against cell-specific markers. At 4 hrs, the earliest IL-1β-expressing cells were found almost exclusively within the dermis at the site of abscess formation (Fig. 3A, C). To identify these early IL-1β-expressing cells, sections were first co-labeled with anti-DsRed and mAbs directed against CD45 (pan-leukocyte marker) or MHCII (antigen presenting cells) (Fig. S3). The vast majority of IL-1β-expressing cells co-localized with CD45 whereas only a few cells co-localized with MHCII, indicating that antigen presenting cells (e.g. dermal dendritic cells and monocytes/macrophages) were not the predominant cell type that expressed IL-1β. This was somewhat surprising since monocytes/macrophages produce large amounts of IL-1β in response to S. aureus or S. aureus components in vitro [16]–[18]. Since neutrophils represent the majority of cells recruited at early time points to the site the S. aureus infection in the skin [3], [11], [19], we next evaluated the expression of IL-1β in monocytes/macrophages and neutrophils using mAbs directed against MOMA2 and 7/4, respectively [26]. We found that only a few IL-1β-expressing cells co-localized with MOMA2 at 4 and 24 hrs after infection (Fig. 3A and S4A). In contrast, the majority of the IL-1β-expressing cells co localized with 7/4 at 4 hrs and especially at 24 hrs after infection (Fig. 3B and S4C).
To quantify the degree of co-localization between IL-1β-expressing cells and the cell-specific markers, image analysis was performed using the Manders' coefficient for a value range of 0 (no pixels co-localize) to 1 (all pixels co-localize) (Fig. 3B,D). The Manders' coefficient between IL-1β-expressing cells and MOMA2+ monocytes/macrophages or MHCII+ antigen presenting cells at 4 hrs was 0.27 or 0.21, respectively (Fig. 3D and Fig. S4), confirming that these cells represented a minority of IL-1β-expressing cells. In contrast, the Manders' coefficient between the IL-1β-expressing cells and 7/4+ neutrophils was 0.56 at 4 hrs and 0.83 at 24 hrs. Although the co-localization of DsRed with the cellular markers does not provide information about how much IL-1β is made per cell, these data suggest that neutrophils represent the most abundant cell type that expresses IL-1β at the site of infection.
Although neutrophils express high levels of 7/4 and monocyte/macrophages express MOMA2 (which includes recently emigrated monocytes and activated macrophages) [26], some subsets of monocytes and macrophages have also been reported to express 7/4 [27]. Therefore, the labeling of 7/4 versus MOMA2 in sections of S. aureus-infected mouse skin was compared (Fig. S5). At both 4 and 24 hrs after infection, there was only a rare occasional cell that expressed both 7/4 and MOMA2. Thus, we conclude that the vast majority of the 7/4+ and IL-1β-expressing cells were neutrophils.
To evaluate the contribution of neutrophil-derived IL-1β in immunity against S. aureus skin infection, bone marrow-derived neutrophils isolated by Percoll density gradient centrifugation from wt or IL-1β-deficient donor mice were adoptively transferred into IL-1β-deficient recipient mice (wt PMN→IL-1β−/− mice or IL-1β−/− PMN→IL-1β−/− mice, respectively) (Fig. 4). Two hrs after adoptive transfer, these mice along with normal wt and IL-1β-deficient mice, were inoculated intradermally with S. aureus. As expected, IL-1β−/− PMN→IL-1β−/− mice and IL-1β-deficient mice developed larger skin lesions and higher in vivo bioluminescence signals compared with wt mice (Fig. 4A and B). The bioluminescent signals could be seen throughout the areas of the infected lesions, indicating that the total lesion size was a reflection of the degree and extent of the bacterial infection. Furthermore, the defects observed in the IL-1β−/− PMN→IL-1β−/− mice or IL-1β-deficient mice were not likely due to impaired neutrophil function in these mice as in vitro assays for phagocytosis, degranulation, oxidative burst and bacterial killing were not significantly different between neutrophils from IL-1β-deficient mice and wt mice (Fig. S6). However, wt PMN→IL-1β−/− mice had lesion sizes and in vivo bioluminescence signals that were similar to normal wt mice, indicating that neutrophil-derived IL-1β is sufficient for host defense against the S. aureus skin infection. To further evaluate if the neutrophils and not other cells (such as the few contaminating monocytes (Fig. S7A) in the adoptively transferred cells played a role in promoting neutrophil recruitment and host defense in the in IL-1β-deficient mice, neutrophils or monocytes were specifically depleted from the adoptively transferred cells by positive selection using anti-Ly6G or anti-CD115 MACS bead separation, respectively (Fig. S7B, C). Depletion of neutrophils reduced the absolute number of transferred Ly6G+ neutrophils from 4.6×106 to 1.8×106 neutrophils/mouse (61% depletion efficiency) and did not decrease the absolute number of transferred CD115+ monocytes (∼1.5×104 before and after depletion). Similarly, depletion of monocytes reduced the absolute number of transferred CD115+ monocytes from 1.3×104 to 3.0×103 (76.9% depletion efficiency) and did not decrease the absolute number of transferred Ly6G+ neutrophils (∼4.6×106 before and after depletion). Although the neutrophil depletion was only 61% complete, the decreased numbers of neutrophils resulted in an inability of adoptively transferred cells from wt mice to rescue the immune impairment in IL-1β-deficient mice. In contrast, monocyte depletion, which decreased the percentage of contaminating monocytes from 0.26% to only 0.06% of the adoptively transferred cells, had no impact on the ability to rescue the immune impairment in IL-1β-deficient mice. These data provide additional evidence that neutrophils and not monocytes in the adoptively transferred cells played a major role in promoting effective neutrophil recruitment and host defense against the cutaneous S. aureus infection.
Histopathological examination of skin biopsies taken one day after infection with S. aureus demonstrated that wt PMN→IL-1β−/− mice and normal wt mice developed large neutrophilic abscesses seen in H&E stained and anti-7/4 labeled sections (Fig. 4C). In contrast, infected skin samples from IL-1β−/− PMN→IL-1β−/− mice and IL-1β-deficient mice had markedly decreased neutrophil recruitment with minimal abscess formation and decreased myeloperoxidase (MPO) activity (which correlates with the degree of neutrophil infiltration) (Fig. 4D). The levels of IL-1β protein expression from infected skin samples at 4 and 24 hrs were evaluated by ELISA and the amount of IL-1β protein at the site of infection in mice adoptively transferred with wt or IL-1β−/− PMN was below the level of detection (data not shown). In contrast, in wt mice at 4 and 24 hrs the levels of IL-1β protein typically exceeded 20 pg/mg tissue weight (Fig. 5C and data not shown). However, immunohistochemistry with an anti-IL-1β mAb identified scattered IL-1β-expressing cells within the abscess at 24 hrs in wt PMN→IL-1β−/− mice but not in IL-1β−/− PMN→IL-1β−/− mice (Fig. S8). These data indicate that adoptively transferred wt neutrophils produced IL-1β at the site of infection and that neutrophil-derived IL-1β is sufficient for promoting effective neutrophil abscess formation and host defense against a cutaneous S. aureus infection.
Certain PRRs have been shown to recognize S. aureus components and initiate innate immune responses, including TLR2, a membrane PRR that recognizes S. aureus lipopeptides and lipoteichoic acid [12], [13], NOD2, a cytosolic PRR that recognizes muramyl-dipeptide (a breakdown product of S. aureus peptidoglycan) [14], [15], and FPRs, which recognize formylated peptides of bacteria [28]. To investigate whether these PRRs contributed to IL-1β production and neutrophil recruitment during a S. aureus skin infection in vivo, we inoculated wt mice and mice deficient in TLR2, NOD2 or FPR1 with S. aureus (Fig. 5). TLR2-, NOD2-, and FPR1-deficient mice all developed larger lesions (up to 4.0-, 2.8- and 2.9-fold, respectively) (Fig. 5A) and higher bioluminescent signals (up to 5.6-, 4.9- and 4.6-fold, respectively) than wt mice (Fig. 5B). Taken together, these results demonstrate that TLR2, NOD2 and FPR1 all significantly contributed to host defense against S. aureus infection in the skin. Furthermore, at 6 hrs after inoculation, S. aureus-infected skin lesions of mice deficient in TLR2, NOD2 or FPR1 had significant reductions in IL-1β protein (80, 62, and 91 percent decrease, respectively) and MPO activity (53, 58, and 47 percent decrease, respectively) compared with wt mice (Fig. 5C, D). Therefore, in addition to having higher in vivo bacterial burden, mice deficient in TLR2, NOD2 and FPR1 also had decreased IL-1β production and neutrophil recruitment during a S. aureus skin infection in vivo. The increased lesion sizes, higher bacterial burden and impaired IL-1β production in TLR2- and NOD2-deficient mice in response to S. aureus skin infection is consistent with previously published studies from our laboratory and others [3], [15].
To determine whether IL-1β production by neutrophils occurred through direct or indirect mechanisms, neutrophils obtained from bone marrow of wt mice and mice deficient in TLR2, NOD2 or FPR1 (purity>99%) were infected with live S. aureus in vitro (Fig. 6). This in vitro infection involved incubating the neutrophils with live S. aureus or a community-acquired MRSA strain (USA300 LAC isolate) (at a multiplicity of infection [MOI] of bacteria to neutrophils of 5∶1) for a total of 6 hrs and gentamicin was added at 60 min from the start of the infection as previously described [16]. The levels of IL-1β protein produced in these cultures were measured using an ELISA that detects both pro-IL-1β and cleaved IL-1β. During this in vitro infection, we observed increased production of IL-1β protein as the MOI increased from 1∶1 to 5∶1 (Fig. S9A, B). The increased production of IL-1β was not due to increased cell death as there was no decrease in the viability of the neutrophils in the presence of S. aureus or MRSA compared with cultures without any bacterial infection (Fig. S9C, D). Furthermore, the lack of any decrease in viability of the neutrophils in the presence of S. aureus or MRSA suggest that the ELISA likely detected mostly cleaved IL-1β rather than pro-IL-1β released into the supernatants from dying cells. Using this in vitro infection, neutrophils from mice deficient in TLR2, NOD2 or FPR1 produced significantly less IL-1β protein (40, 43 and 37 percent decrease, respectively) in response to S. aureus (Fig. 6A), suggesting that activation of TLR2, NOD2 and FPR1 promoted neutrophil production of IL-1β. To provide further evidence that neutrophils and not other contaminating cells such as monocytes produced IL-1β in these cultures, purified neutrophils from pIL1-DsRed reporter mice were evaluated in this in vitro infection and 43% of the Ly6G+ neutrophils expressed IL-1β-DsRed whereas only 0.2% of other cell types (Ly6G− cells) expressed IL-1β-DsRed (Fig. S10). Thus, neutrophils, and not other contaminating cells, were the predominant source of IL-1β in these cultures.
An important step in the production of active IL-1β is the enzymatic processing of pro-IL-1β into its active form. Typically, this cleavage is mediated by caspase-1, which is activated by an intracellular complex of proteins called the inflammasome [29]. However, under certain conditions, cleavage of pro-IL-1β into active IL-1β in neutrophils can be mediated by serine-proteases (such as proteinase 3) or neutrophil elastase rather than inflammasome/caspase-1 activation [30]–[32]. To determine whether IL-1β production by neutrophils in response to S. aureus involved inflammasome activation, we studied neutrophils from mice deficient in ASC, which is required for NLRP3 inflammasome assembly [33]. Neutrophils from wt or ASC-deficient mice were infected with live S. aureus in vitro (Fig. 6B). Neutrophils from ASC-deficient mice had a 58% decrease in IL-1β production in response to S. aureus compared with neutrophils from wt mice, indicating that the majority of IL-1β produced during the S. aureus in vitro infection was dependent on the inflammasome component ASC.
Previous studies in mouse and human monocyte/macrophage cultures demonstrated that processing of pro-IL-1β after exposure to S. aureus was dependent upon ASC/NLRP3 inflammasome and caspase-1 activation, which was induced by S. aureus α-toxin and other pore-forming hemolysins [17], [34]. Therefore, to evaluate whether a similar mechanism of inflammasome activation and IL-1β production was involved in mouse neutrophils, mouse neutrophils were infected with S. aureus or MRSA in vitro in the presence an inhibitor of the NLRP3 inflammasome (glibenclamide) [35], [36], a specific caspase-1 inhibitor (Z-YVAD-FMK) or neutralizing antibodies directed against S. aureus α-toxin [37] (Fig. 6C, D). In mouse neutrophils infected with S. aureus, the amount of IL-1β produced was decreased 62%, 73% and 53% by the NLRP3 inflammasome inhibitor, the caspase-1 inhibitor and the α-toxin neutralizing antibodies, respectively (Fig. 6C). Similarly, in mouse neutrophils infected with MRSA, the amount of IL-1β produced was decreased 44%, 57% and 58% by the NLRP3 inflammasome inhibitor, the caspase-1 inhibitor and the α-toxin neutralizing antibodies, respectively (Fig. 6D). Importantly, addition of inhibitors or neutralizing antibodies did not decrease the viability of neutrophils infected with S. aureus or MRSA (Fig. S11). To confirm that S. aureus or MRSA infection of mouse neutrophils resulted in the generation of cleaved IL-1β, immunoblotting was performed and cleaved IL-1β was only detected in cultures of S. aureus-or MRSA-infected neutrophils and not in uninfected neutrophil cultures (Fig. S12). In these mouse neutrophil cultures, there was no decrease in IL-1β produced in control wells containing DMSO (the vehicle for the NLRP3 inflammasome inhibitor and the caspase-1 inhibitor) or rabbit IgG (the control for the α-toxin neutralizing antibodies) compared with media alone (data not shown). Taken together, these data indicate that the majority of the IL-1β produced was dependent upon activation of caspase-1 via induction of the NLRP3/ASC inflammasome in an α-toxin-dependent mechanism. Furthermore, the dependence of IL-1β production on the NLRP3/ASC inflammasome and caspase-1 provides additional evidence that the IL-1β detected by the ELISA was mostly cleaved IL-1β rather than pro-IL-1β.
It should be noted that it was necessary to use methods that yielded purified cultures of mouse neutrophils for the in vitro infection experiments to minimize monocyte contamination. Mouse neutrophils were positively selected from bone marrow cells using anti-Ly6G magnetic bead separation. This method resulted in 99.1% purity of mouse neutrophils (Fig. S13). Although the positive selection with anti-Ly6G magnetic bead separation may have induced some activation of the mouse neutrophils, this degree of activation was unlikely to play a major role in production of IL-1β because there was minimal IL-1β production observed in cultures of uninfected neutrophils (Fig. 6A–D).
Neutrophil abscess formation is an essential component of innate immunity against many pathogens [1]. In this study, using gene expression analysis and advanced techniques of in vivo fluorescence imaging, we found that neutrophil recruitment during a S. aureus cutaneous infection is functionally and temporally linked to IL-1β/IL-1R activation. Based on our prior work [3], [11], we hypothesized that this association was a result of IL-1β production by hematopoietic cells such as macrophages or dendritic cells and possibly other cells that reside in the skin, such as keratinocytes or mast cells [20]–[23]. Surprisingly, we found that neutrophils were the most abundant source of IL-1β during infection. Neutrophil-derived IL-1β, in the absence of other cellular sources of IL-1β, was critical for host defense since adoptive transfer of IL-1β-expressing neutrophils was sufficient to restore the impaired neutrophil recruitment and abscess formation in S. aureus-infected IL-1β-deficient mice. In addition, mouse neutrophils produced IL-1β in vitro in response to live S. aureus in a mechanism involving the PRRs, TLR2, NOD2 and FPR1, and the ASC/NLRP3 inflammasome. Thus, neutrophil recruitment and IL-1β/IL-1R are functionally, temporally and spatially linked because neutrophils are the predominant source of IL-1β. These findings provide a new paradigm for abscess formation during an infection in which the inflammatory mediators produced by the epithelial, stromal and resident immune cells in the infected tissue may contribute to the recruitment of the very first neutrophils; however, this response is not sufficient for effective abscess formation. Rather, a feed-forward mechanism that involves early recruited neutrophils serving as a source of IL-1β is essential for amplifying and sustaining the neutrophilic response to promote optimal abscess formation and bacterial clearance.
Although a link between neutrophil recruitment and IL-1β during S. aureus infections was previously documented [7], [10], [11], our work defines the primary mechanism by which effective neutrophil abscess formation occurs. These findings provide an explanation for a number of puzzling observations in humans and mice and have important implications for neutrophil-derived IL-1β in potentially contributing to other immune responses during infection and inflammation. First, human pediatric patients with deficiency in TLR/IL-1R signaling molecules, MyD88 or IRAK-4, are predisposed to pyogenic bacterial infections, including S. pneumoniae, S. aureus, and P. aeruginosa, whereas other types of bacterial, fungal and viral infections are exceedingly rare [38], [39]. The reason for this has remained elusive, especially since these patients do not have impaired neutrophil number or function as seen in other conditions predisposed to pyogenic infections such as severe congenital neutropenia or chronic granulomatous disease [9]. Interestingly, during acute or invasive infections, patients with MyD88 or IRAK-4 deficiency develop neutropenia despite having pus in infected tissues [40]. Our findings suggest a pathway beginning with S. aureus-induced inflammation in the skin tissue that results in an initial early recruitment of neutrophils that produce IL-1β. The neutrophil-derived IL-1β is sufficient to amplify and sustain their recruitment that promotes neutrophilia and effective neutrophil abscess formation. Although patients with MyD88 or IRAK-4 deficiency may recruit neutrophils to the site of infection, they cannot respond to neutrophil-derived IL-1β to amplify the neutrophilic response, providing a potential explanation for their selective predisposition to pyogenic infections.
Second, previously published work has demonstrated that neutrophils express pattern recognition receptors, including TLR2 [41], [42], NOD2 [43] and FPRs [28]. Thus, we evaluated TLR2-, NOD2-, or FPR1-deficient mice in response to S. aureus skin infection and found that each of these mice had impaired IL-1β production (Fig. 5C) and neutrophil recruitment after S. aureus skin infection (Fig. 5D). Based on our in vitro infection experiments with neutrophils from TLR2-, NOD2-, or FPR1-deficient mice, there was decreased IL-1β production compared with neutrophils from wt mice (Fig. 6A), suggesting these each of these PRRs on neutrophils directly contribute to the production of IL-1β in response to S. aureus. Since TLR2, NOD2 and FPR1 are activated by different bacterial components, they likely provide overlapping and redundant functions to ensure adequate neutrophil IL-1β production and a deficiency in any one of these PRRs would not have a major impact on host defense. However, TLR2, NOD2 and FPR1 have also been shown to also be involved in other neutrophil functions such as chemotaxis, phagocytosis and oxidative burst [14], [44]–[47]. Thus, the impaired immune response against S. aureus in TLR2-, NOD2-, or FPR1-deficient mice in vivo may not be solely due to decreased IL-1β production but is likely dependent upon the lack of other functional activities of these PRRs on neutrophils as well as on other cell types in the infected skin in these knockout mice. To determine if TLR2 functioned predominantly on neutrophils or other cell types in vivo, we performed an additional experiment in which wt neutrophils were adoptively transferred into TLR2-deficient mice (Fig. S14). This adoptive transfer of wt neutrophils did not rescue the immune impairment in TLR2-deficient mice as observed with the adoptive transfer of wt neutrophils into IL-1β-deficient mice (Fig. 4). Thus, TLR2 activation is needed on other cells to invoke protective mechanisms in vivo.
Third, our previous work found that production of IL-1β was likely from bone marrow-derived hematopoietic cells because neutrophil recruitment, host-defense and IL-1β production at the site of infection was restored in IL-1β-deficient mice reconstituted with bone marrow from wt mice but not IL-1β-deficient mice [11]. Here, we found that adoptively transferred wt neutrophils could rescue the immune impairments in IL-1β-deficient mice, indicating neutrophils are the predominant hematopoietic cellular source of IL-1β that was sufficient for effective neutrophil recruitment and abscess formation. These data further argue against an important role for IL-1β produced by non-hematopoietic cells during the S. aureus skin infection. Although keratinocytes produced IL-1β during the infection as detected by immunohistochemistry (Fig. S15), this amount of IL-1β was not able to promote effective neutrophil recruitment in the absence of IL-1β-expressing hematopoietic cells because wt mice reconstituted with bone marrow from IL-1β-deficient mice show the same impaired neutrophil recruitment response as normal non-irradiated/non-reconstituted IL-1β-deficient mice [11].
Fourth, IL-1 has been shown to play a role in neutrophil recruitment during sterile inflammation. In a mouse model of intraperitoneal injection of necrotic lymphoma cells or acetaminophen-induced liver injury, neutrophil recruitment to the peritoneal cavity or liver was mediated by IL-1α alone or both IL-1α and IL-1β, respectively [48]. Interestingly, in a mouse model of autoimmune inflammatory arthritis, neutrophil recruitment to the inflamed joints was dependent on leukotriene B4 (LTB4) [49]. However, exogenous IL-1β injected into the joints or adoptive transfer of wt neutrophils could restore neutrophil recruitment and arthritis in LTB4-deficient mice, demonstrating that IL-1β-producing neutrophils amplified neutrophil recruitment and arthritis [49], which was similar to what we observed during a S. aureus skin infection. Thus, although IL-1β-producing neutrophils are sufficient for neutrophil recruitment during a S. aureus skin infection, neutrophil recruitment during sterile inflammation is mediated by IL-1α, IL-1β or both IL-1α and IL-1β, depending on the anatomical site and the type of inflammation.
Neutrophil-derived IL-1β may also promote other immune responses at the site of infection. Similar to our findings, a previous report found that NOD2-deficient mice had impaired production of IL-1β during a S. aureus skin infection [15]. This report also found that NOD2-induced IL-1β contributed to production of IL-6, which enhanced neutrophil killing of S. aureus [15]. In our previous work, we found that IL-1R-mediated neutrophil recruitment (through production of the neutrophil-attracting chemokines KC and MIP2) was dependent upon IL-1R-signaling by resident skin cells rather than bone marrow-derived recruited cells [3]. These findings were based upon data using bone marrow chimeric mice in which the impaired host defense and neutrophil recruitment in IL-1R-deficient mice could not be restored in IL-1R-deficient mice reconstituted with bone marrow from wt mice [3]. In contrast, wt mice reconstituted with bone marrow from IL-1R-deficient had no immune impairment. Additionally, as mentioned above, IL-1β production during the S. aureus skin infection was found to be dependent on bone marrow-derived cells rather than resident skin cells [11]. Combining these previous studies with the present findings, a host defense pathway has been discovered whereby neutrophils represent a source of IL-1β, which subsequently activates IL-1R expressed on non-bone marrow-derived resident skin cells to promote effective neutrophil recruitment in host defense during a S. aureus skin infection. Furthermore, we had previously demonstrated that IL-1R activation was required for inducing IL-17A/F production by γδ T cells in infected mouse skin at early time points after S. aureus infection [19]. In this context, IL-17A/F promoted enhanced neutrophil recruitment via induction of neutrophil-attracting chemokines and granulopoiesis factors. Since IL-1β has also been shown to be important in the generation of Th17 cells [50], [51], future studies will be required to determine if neutrophil-derived IL-1β contributes to the IL-6 responses as well as the development of Th17 cells and other IL-17-producing cells following a cutaneous S. aureus infection.
It should also be noted that since IL-1β- and IL-1R-deficient mice ultimately clear these infections, compensatory mechanisms exist that eventually promote bacterial clearance. Similar compensatory mechanisms also may play a role in TLR2-, NOD2- and FPR1-deficient mice as these mice also eventually clear the infection and the cellular composition of 7/4+ neutrophils and MOMA2+ monocytes/macrophages on day 10 after infection in TLR2-, NOD2-, FPR1-deficient mice was similar to the cellular composition in wt mice whereas IL-1β-deficient mice had a paucity of 7/4+ neutrophils at this time point (Fig. S16). These compensatory responses may include activation of other MyD88-dependent receptors such as TLRs, IL-18 or IL-33 because we found that MyD88-deficient mice have a more severe impairment in neutrophil recruitment than IL-1β- or IL-1R-deficient mice [3], [11]. In addition, IL-17 has also been shown to be critical in promoting neutrophil recruitment and antimicrobial responses against S. aureus in various mouse models of infection (cutaneous infection, systemic infections, pneumonia and brain abscesses [52]–[56]) as well as in humans with hyper-IgE syndrome or with a deficiency in IL-17F or IL-17RA [50], [57]–[60]. Although Th17 development is severely impaired in IL-1R-deficient mice in vivo [61], [62] and γδ T cell production of IL-17 is enhanced in the presence of IL-1β [19], [63], the numbers and activity of Th17 and γδ T cells may increase during the course of the S. aureus skin infection and compensate for absence of IL-1β activity. Consistent with this possibility, humans with deficiency in the IL-1R downstream signaling molecules MyD88 or IRAK-4 do not exhibit impaired development of IL-17-producing cells [64].
The use of in vivo imaging in this study to quantify the bacterial burden, neutrophil recruitment and IL-1β production provides an approximation of these endpoints and there are advantages and limitations that should be considered when interpreting this data. First, in vivo bioluminescence imaging provides only a close estimate of in vivo bacterial burden as several factors such as body temperature, metabolic activity of the bacteria in vivo [65], [66] and the presence of reactive oxygen mediators produced by neutrophils at the site of infection that could potentially react with the bacterial luciferase as seen with GFP-labeled bacteria [67], [68]. However, despite these potential confounding factors, we previously demonstrated that in vivo bioluminescent signals directly correlate with ex vivo CFUs harvested at different time points from the S. aureus-infected skin lesions [3], [19], [69]. Thus, the bioluminescent signals and actual bacterial burden is not a perfect correlation; however, it is a noninvasive method that approximates the bacterial burden in vivo that does not require euthanasia of numerous animals at every time point to obtain this information. Second, regarding the use of the LysEGFP mice, lysozyme M is expressed in myeloid cells (including neutrophils and monocytes/macrophages) and the lysozyme M promoter driven EGFP expression is not specific for neutrophils [25]. However, neutrophils from LysEGFP mice have been shown to have much brighter EGFP fluorescence intensity than monocytes or macrophages [70], [71] and we found that F4/80+ macrophages constituted less than 10% of the EGFP-expressing cells during the first 5 days after skin wounding of LysEGFP mice [72], indicating that neutrophils may contribute to more than 90% of the EGFP signals. In addition, the decreasing EGFP signals from days 1 to 10 correlated with the decreasing numbers of 7/4+ neutrophils and not the increasing numbers of MOMA2+ monocytes/macrophages at these time points as detected by immunohistochemistry (Fig. 3B and S16). Furthermore, EGFP was expressed within the cytoplasm and intracellular vesicles of LysEGFP neutrophils and the intensity of EGFP fluorescent signals was not substantially decreased in culture after neutrophil degranulation induced in response to fMLF or PMA (Fig. S17). Thus, the EGFP fluorescent signals more closely approximate of the numbers of neutrophils within the infected skin during the course of infection. With respect to using pIL1-DsRed transgenic mice, there are slightly different kinetics between DsRed fluorescence and IL-1β protein expression measured by ELISA, since DsRed fluorescence in extracts of inflamed skin and in cell culture was induced 6–12 hrs slower and persisted ∼24 hrs longer than IL-1β protein levels [21]. However, the difference in kinetics of DsRed fluorescence signals were less than 24 hrs and thus would be unlikely to impact the approximation of IL-1β production in vivo, since we began our measurements 1 day after infection and the infection takes over 14 days to resolve. Furthermore, there was substantial DsRed fluorescence signal on day 1 during our in vivo S. aureus skin infection (Figs. 2 and 3) and at this time point we previously found that IL-1β was detected from the infected skin by ELISA and that both pro-IL-1β and cleaved IL-1β were detected by immunoblotting [11], [17].
In the present study, we found that processing of pro-IL-1β into active IL-1β by mouse neutrophils was largely dependent upon ASC/NLRP3 inflammasome activation in vitro. We further demonstrated that IL-1β production was dependent upon the activity of α-toxin. These data are consistent with previous work in human or mouse monocyte/macrophage cultures demonstrating that processing of pro-IL-1β during S. aureus cutaneous infections in vivo was dependent upon ASC/NLRP3 inflammasome and caspase-1 activation, which was induced by S. aureus pore-forming toxins (i.e. α-, β- and γ-hemolysins) or digestion of peptidoglycan mediated by lysozyme [17], [18], [34]. Since there was some IL-1β measured in cultures of neutrophils from ASC-deficient mice as well as in cultures of wt neutrophils in the presence of the NLRP3 or caspase-1 inhibitor, the remainder of IL-1β may have been produced through an inflammasome-independent pathway mediated by serine-proteases (such as proteinase 3) or neutrophil elastase [30]–[32]. In addition, the residual IL-1β detected in these cultures may also reflect transcription of pro-IL-1β since the ELISA detects both pro-IL-1β and cleaved IL-1β.
In the adoptive transfer experiments, although the expression of IL-1β protein at the site of infection was extremely low compared with the levels observed in wt mice, scattered IL-1β-producing cells were detected by immunohistochemistry in IL-1β-deficient mice adoptively transferred with wt neutrophils (Fig. S8). While it is tempting to speculate that these adoptively transferred IL-1β-producing neutrophils directly rescued the neutrophil recruitment response at the site of infection, it is also possible the low levels of IL-1β acted indirectly and/or through another anatomical site such as the blood [73]. Future studies will be necessary to further dissect the mechanism by which neutrophil-derived IL-1β contributes to host defense during a S. aureus skin infection. Nevertheless, we show that IL-1β-producing adoptively transferred neutrophils were sufficient to rescue the impaired immunity in IL-1β-deficient mice. These findings provide evidence that neutrophil-derived IL-1β can promote effective neutrophil abscess formation and host defense against a cutaneous S. aureus infection.
IL-1β production in response to live S. aureus cultured with mouse bone marrow-derived macrophages (BMDMs), mouse peritoneal macrophages or human monocytes has previously been described [16]–[18], [74]. These studies used different S. aureus strains, MOIs and culture conditions and the amount of IL-1β produced was generally 15- to 100-fold greater than the levels we observed with our in vitro infection of mouse neutrophils with S. aureus or MRSA. To evaluate whether IL-1β produced by the few contaminating monocytes/macrophages played any role in the adoptive transfer experiments, monocytes were specifically depleted from the adoptively transferred cells and the lack of monocytes had no impact on the ability of the adoptively transferred wt neutrophils to rescue the immune impairment in IL-1β-deficient mice. Furthermore, our in vitro infection experiments used highly purified mouse neutrophils separated with anti-Ly6G MACS magnetic beads (99.1% pure with only 0.1% monocytes) to ensure that we were evaluating neutrophil specific production of IL-1β. To provide additional evidence that neutrophils and not other contaminating monocytes produced IL-1β in these cultures, purified neutrophils from pIL1-DsRed reporter mice were evaluated and 43% of the neutrophils expressed IL-1β-DsRed whereas only 0.2% of other cell types expressed IL-1β-DsRed. Taken together, these data indicate that neutrophils were the predominant source of IL-1β for both the adoptive transfer experiments and the in vitro infection experiments. Finally, a recent study evaluated sorted mouse bone marrow cells to determine which cell type produced the majority of IL-1β in response to LPS in the presence of the known inflammasome activators ATP or nigericin [75]. They found neutrophils were the predominant source of IL-1β as they produced almost 3-fold more IL-1β than F4/80+ macrophages. They further demonstrated that human neutrophils were responsible for half of all IL-1β secreted by human PBMCs. Lastly, they showed that IL-1β production by mouse and human neutrophils involved activation of inflammasome via NLRP3/ASC/caspase-1 axis. These data are consistent with our findings that neutrophils provide a major source of IL-1β during a S. aureus skin infection that is produced in an NLRP3/ASC/caspase-1-dependent manner
In summary, we have identified that neutrophil-derived IL-1β is essential for amplifying the neutrophilic response to promote abscess formation and clearance of a S. aureus skin infection. From a clinical point of view, these findings provide the basis for targeting IL-1β production by neutrophils to improve immunity against pyogenic infections, especially in patients with impaired neutrophilic responses.
All animals were handled in strict accordance with good animal practice as defined in the federal regulations as set forth in the Animal Welfare Act (AWA), the 1996 Guide for the Care and Use of Laboratory Animals, PHS Policy for the Humane Care and Use of Laboratory Animals, as well as UCLA's policies and procedures as set forth in the UCLA Animal Care and Use Training Manual, and all animal work was approved by the UCLA Chancellor's Animal Research Committee (ARC#: 2008-099).
The bioluminescent S. aureus SH1000 strain ALC2906, which possesses the shuttle plasmid pSK236 with the penicillin-binding protein 2 (pbp2) promoter fused to the modified luxABCDE reporter cassette from Photorhabdus luminescens, was used as a representative S. aureus strain [3]. This strain emits bioluminescence signals from live, actively metabolizing bacteria in all stages of the S. aureus life cycle. In some experiments, a community-acquired MRSA strain was used (USA300 LAC isolate [76]), which was kindly provided by Frank DeLeo (National Institute of Allergy and Infectious Diseases, Rocky Mountain Laboratories in Hamilton, MT).
SH1000 cultures were grown in the presence of chloramphenicol (10 µg/ml; Sigma-Aldrich, St. Louis, MO). S. aureus or MRSA was streaked onto tryptic soy agar (tryptic soy broth [TSB] plus 1.5% bacto agar; BD Biosciences, Sparks, MD) and single colonies were placed into TSB and grown overnight at 37°C in a shaking incubator. Mid-logarithmic phase bacteria were obtained after a 2 hr subculture of a 1∶50 dilution of the overnight culture. Bacterial cells were pelleted, resuspended, and washed 3 times in PBS. Bacterial concentrations were estimated by measuring the absorbance at 600 nm (A600) (Biomate 3; Thermo Scientific, Waltham, MA). In some experiments, bacteria was heat-killed (65°C for 30 minutes) prior to infection. CFUs were verified by plating dilutions of the inoculum overnight.
Male mice on a C57BL/6 genetic background were used in all experiments. pIL1-DsRed-reporter mice [21], LysEGFP mice [25], FPR1-deficient mice [77], IL-1β-deficient mice [78] and ASC-deficient mice [33] were generated as previously described. IL-1R1-deficient mice (B6.129S7-Il1r1tm1Imx/J), TLR2-deficient mice (B6.129-TLR2tm1Kir/J) and NOD2-deficient mice (B6.129S1-Nod2tm1Flv/J) and wt C57BL/6 mice were obtained from Jackson Laboratories (Bar Harbor, ME). All mouse colonies were maintained in autoclaved cages under specific-pathogen free conditions.
The mice were shaved on the back and inoculated intradermally with mid-logarithmic growth phase S. aureus (2×106 CFUs) in 100 µl of sterile saline using a 27-gauge insulin syringe as previously described [3]. Measurements of total lesion size (cm2) were made by analyzing digital photographs of mice using the software program Image J (http://rsbweb.nih.gov/ij/).
Skin punch biopsy (8-mm) specimens from uninfected or lesional skin were taken at 4 hrs after S. aureus intradermal inoculation from wt and IL-1R-deficient mice and homogenized (Bio-Gen Pro200; Pro Scientific, Oxford, CT). RNA was isolated using TRIzol reagent (Invitrogen, Grand Island, NY) and purified using the RNeasy Mini kit (Qiagen, Valencia, CA). The UCLA Microarray Core performed probe synthesis and hybridization to the GeneChip Mouse Genome 430 2.0 Array (Affymetrix, Maumee, OH) according to the manufacturer's protocol. Image files were processed using the invariant set method for probe selection during normalization and the model-based expression method of pooling information across arrays using dCHIP (DNA-Chip Analyzer) gene expression software (www.dchip.org) [79]. Genes were considered upregulated in S. aureus-infected skin at 4 hrs compared with uninfected skin according to the criteria: fold-change >1.5, p-value<0.05. Functional group and network analysis was performed using Ingenuity Pathway Analysis software (version 6.0; Ingenuity Systems, Redwood City, CA) as previously described [80]. The raw gene expression data for this study are available through the Gene Expression Omnibus database (http://www.ncbi.nlm.nih.gov/geo/) under accession number GSE36826.
Total RNA from homogenized (Pro200 Series homogenizer [Pro Scientific]) 8-mm skin biopsy specimens taken at 4 hrs from skin inoculated with live or heat-killed S. aureus and uninfected skin was extracted by the use of TRIzol reagent (Invitrogen), followed by DNase treatment (Invitrogen) according to the manufacturer's recommendations. Real-time quantitative real-time PCR (Q-PCR) reactions were performed as previously described [19]. TaqMan Gene Expression Assays primers and probes sets for a subset of genes of interest in the Cell Movement of Neutrophils sub-group, including IL-6, G-CSF, MIP-2β, MIP-2α, IL-1β, GM-CSF, CXCR2, G-CSF-R, TLR2 and LTB4R and the normalizer GAPDH were purchased from Applied Biosystems (Foster City, CA). The relative quantities of mRNA per sample were determined using the ΔΔC(T) formula as previously described [3]
Mice were anesthetized via inhalation of isoflurane and in vivo bioluminescence imaging was performed using the IVIS Lumina II imaging system (Caliper Life Sciences, a PerkinElmer Company, Alameda, CA) as previously described [3]. Data are presented on color scale overlaid on a grayscale photograph of mice and quantified as total flux (photons/s) within a circular region of interest using Living Image software (Caliper).
pIL1-DsRed mice and LysEGFP mice were anesthetized with inhalation isoflurane and in vivo fluorescence imaging was performed (sequentially after in vivo bioluminescence imaging) using the IVIS Lumina II imaging system (Caliper). DsRed fluorescence was measured using: excitation (535 nm), emission (575–650 nm) and exposure time (0.5 s). EGFP fluorescence was measured using: excitation (465 nm), emission (515–575 nm) and exposure time (0.5 s). Data are presented on color scale overlaid on a grayscale photograph of mice and quantified as total radiant efficiency ([photons/s]/[µW/cm2]) within a circular region of interest using Living Image software (Caliper).
For histological analysis, lesional 8-mm punch biopsy skin specimens were embedded in Tissue-Tek OTC compound (Sakura Finetek) and cut into 4 µm sections by the UCLA Tissue Procurement and Histology Core Laboratory, according to guidelines for clinical samples.
Frozen sections were fixed in acetone, air-dried, and rehydrated in PBS. Sections were permeabilized with 0.1% saponin in PBS and non-specific binding was blocked with 2% goat serum (Invitrogen) and mouse IgG2a (10 µg/ml; clone UPC 10, Sigma-Aldrich) in PBS. Sections were subsequently labeled with primary antibodies specific for DsRed (rabbit anti-DsRed antibody; Clontech, Mountain View, CA) in combination with mAbs specific for neutrophils (anti-7/4 [Ly-6B.2]; 5 µg/ml; AbD Serotec, Raleigh, NC), monocytes/macrophages (anti-MOMA2; 5 µg/ml; AbD Serotec), antigen presenting cells (anti-MHC II; 5 µg/ml; clone 2G9; BD Biosciences) or total leukocytes (anti-CD45; 5 µg/ml; clone 30-F11; BD Biosciences) or appropriate isotype controls. Secondary antibodies included goat anti-rabbit IgG-Alexa 568 and goat anti-rat IgG-Alexa 488 (Invitrogen, Carlsbad, CA). All specimens were imaged on a Leica SP2-1P FCS Confocal Microscope (Leica Microsystems, Heidelberg, Germany) as previously described [19]. Representative images of isotype controls are shown (Fig. S4B, D). Quantification of co-localization was performed using the Manders' coefficient for a value range of 0 to 1 in which 0 = no pixels co-localize and 1 = all pixels co-localize using Definiens Tissue Studio software (Definiens, Parsippany, NJ). The Manders' coefficient was determined from 4 different mice per group after averaging 2–3 fields of view per specimen.
Detection of 7/4+ or MOMA2+ cells on frozen sections of lesional skin was performed with the anti-7/4 mAb or the anti-MOMA2 mAb as described above, followed by the biotinylated goat anti-rat IgG polyclonal antibody (5 µg/ml; Vector Labs, Burlingame, CA) or corresponding isotype control antibodies. To detect IL-1β protein expression, a biotinylated anti-mouse IL-1β (20 µg/ml: clone 1400.24.17; Thermo Scientific) or corresponding isotype control mAb was employed. All procedures were performed using the immunoperoxidase method as previously described [11].
MPO activity in lesional skin specimens was determined using an established MPO activity assay. Briefly, 8-mm punch biopsies were weighed and homogenized (Bio-Gen Pro200; Pro Scientific) in a buffer containing potassium phosphate (50 mM, pH 6.0) and hexadecyltrimethylammonium bromide (0.5%; Sigma-Aldrich). To measure MPO levels, 140 µl assay buffer, containing o-Dianisidine dihydrochloride (0.168 mg/ml; Sigma-Aldrich) and hydrogen peroxide (0.05%; Sigma-Aldrich), was added to 10 µl of homogenized supernatant and the change in absorbance (A490) was determined at 40 s intervals for 2 min using the Synergy 2 microplate reader (BioTek, Winooski, VT). Purified MPO (Sigma-Aldrich) was used to generate a standard curve and data are presented as MPO activity (U/mg tissue weight).
For mouse neutrophils, the expression of the neutrophil-specific marker Ly6G (FITC-conjugated rat anti-Ly6G mAb; 5 µg/ml; clone 1A8, IgG1; BD Pharmingen) and the monocyte-specific marker CD115 (the M-CSF receptor) (PE-conjugated rat anti-CD115 mAb, 2 µg/ml; clone AFS98; eBioscience), CD11b (APC-conjugated rat anti-CD11b mAb; 2 µg/ml; clone M1/70; BD Pharmingen) and corresponding fluorescently-conjugated isotype control mAbs were used. Using these antibodies, Ly6G+ CD115− or Ly6G+ CD11bhigh represented mouse neutrophils and Ly6G− CD115+ or Ly6G− CD11blow represented mouse monocytes as previously described [81]–[83]. In some experiments, purified neutrophils from pIL1-DsRed reporter mice were used. pIL1-DsRed neutrophils were co-labeled with the anti-Ly6G mAb and prepared for flow cytometry as described above.
For adoptive transfer experiments, neutrophils were obtained from the bone marrow of IL-1β-deficient or wt mice using Percoll density gradient centrifugation. Briefly, marrow cavities of the tibias and femurs of 8-week old mice were flushed with complete RPMI 1640 containing 10% FBS. After hypotonic lysis of red blood cells, mature neutrophils were isolated by centrifugation for 30 min at 10°C and 1600 g over a discontinuous Percoll gradient consisting of 50% (vol/vol), 55% (vol/vol), 62% (vol/vol) and 81% (vol/vol) Percoll (Sigma-Aldrich, St. Louis, MO) in PBS.
The purity of the adoptively transferred cells was determined by flow cytometry using the neutrophil specific marker Ly6G (anti-Ly6G mAb, clone 1A8) and the monocyte-specific marker (anti-CD115 mAb, the M-CSF receptor). These markers have been shown to distinguish between mouse neutrophils and monocytes by flow cytometry [81], [82]. The adoptively transferred cells were 90% neutrophils (Ly6G+ CD115− cells), 0.26% monocytes (Ly6G− CD115+ cells) and 9.1% Ly6G− CD115− cells, which likely represented other granulocytes (eosinophils or basophils) or residual red blood cells not lysed with the lysis buffer (Fig. S7A). After washing extensively in saline, 5×106 adoptively transferred neutrophils in 100 µl of sterile saline were injected intravenously into IL-1β-deficient mice two hrs prior to intradermal inoculation with S. aureus. In some experiments, neutrophils or monocytes were specifically depleted prior to adoptive transfer using either an anti-Ly6G or an anti-CD115 MicroBead Kit and MACS magnetic bead separation (61.4% and 76.9 percent depletion efficiency, respectively) according to the manufacturer's protocols (Miltenyi Biotec, Inc., Auburn, CA) (Fig. S7B, C). In another set of adoptive transfer experiments, neutrophils were obtained from the bone marrow of TLR2-deficient or wt mice and adoptively transferred into TLR2-deficient recipient mice and the mice were infected with S. aureus according to the same procedures as described above (Fig. S14).
For all in vitro cultures with mouse neutrophils, neutrophils were obtained from the bone marrow of TLR2-, NOD2, FPR1- and ASC-deficient mice, pIL1-DsRed reporter mice or wt mice by anti-Ly6G MACs magnetic bead separation according to the manufacturer's protocols (Miltenyi Biotec, Inc.). Purity of the mouse neutrophils was determined by flow cytometry (see above) and these cultures contained 99.1% Ly6G+ CD11bhigh neutrophils. There were very few Ly6G− CD11b+ monocytes (0.1%) and the remaining cells (0.6%) were Ly6G− CD11b− cells (Fig. S13).
Murine neutrophils (from TLR2-, NOD2-, FPR1-, ASC-deficient or wt mice) were cultured in RPMI 1640 complete media supplemented with 10% heat-inactivated FBS at a density of 1×105 cells per 200 µl/well in a 96-well plate. These neutrophil cultures were infected with live S. aureus (SH1000 strain) or MRSA (USA300 LAC isolate) at a multiplicity of infection (MOI) of bacteria to neutrophils of 5∶1, 2∶1 or 1∶1 at 37°C and 5% CO2 in a humidified incubator for 6 hrs. Gentamicin (20 µg/ml) was added to the cultures at 60 minutes after infection according to previous methods to study inflammasome activation in response to live S. aureus in vitro [16]. Using these culture conditions, the MOI of 5∶1 for S. aureus or MRSA resulted in the highest production of IL-1β compared with MOI of 2∶1 or 1∶1 (Fig. S9A, B). The levels of IL-1β in wt mouse neutrophils did not differ more than 15% between experiments. There was also no decrease in neutrophil viability in any of the cultures with the different MOI of S. aureus or MRSA compared with neutrophils cultured in the absence of any bacteria (Fig. S9C, D). Therefore, the MOI of 5∶1 was used in all in vitro culture experiments. In some experiments, specific inhibitors were also added to the culture at the same time as S. aureus or MRSA. These include, the NLRP3-inhibitor, glibenclamide (100 µM; Imgenex) [35], [36], the caspase-1 inhibitor Z-YVAD-FMK (20 µM; Millipore), anti-staphylococcal α-toxin antiserum (1% vol/vol; Sigma-Aldrich) [37], or respective vehicle controls (DMSO or normal rabbit IgG).
After in vitro infection of neutrophils with S. aureus or MRSA, cell viability was determined using the CellTiter 96 AQueous One Solution Cell Viability Assay (Promega Corporation, Madison, WI) according to the manufacturer's instructions.
Protein levels of IL-1β from lesional mouse skin were obtained from tissue homogenates (Pro200 Series homogenizer [Pro Scientific]) of 8-mm skin punch biopsy specimens performed at 4 and 24 hrs after S. aureus skin inoculation using a commercially-available ELISA kit (R&D Systems). Levels of mouse IL-1β protein in culture supernatants were determined by using a commercially available ELISA kit (R&D Systems, Minneapolis, MN) according to the manufacturer's instructions.
For detection of pro-IL-1β (35 kDa) and cleaved IL-1β (17 kDa) by immunoblotting, purified mouse neutrophils from wt mice were cultured in RPMI 1640 media supplemented with 10% heat-inactivated FBS at a density of 1×106 cells per 500 µl/well in a 24-well plate. These neutrophil cultures were infected with live S. aureus (SH1000 strain) or MRSA (USA300 LAC isolate) at an MOI of bacteria to neutrophils of 5∶1 at 37°C and 5% CO2 in a humidified incubator for 6 hrs and gentamicin (20 µg/ml) was added to the cultures at 60 minutes after infection. Following incubation, cells were lysed using the M-PER Mammalian Protein Extraction Reagent (Thermo-Fisher) supplemented with a Protease Inhibitor Cocktail (Sigma-Aldrich). Cell lysates were diluted in SDS-PAGE sample buffer, boiled for 5 min, and proteins were separated by SDS-PAGE. Pro-IL-1β and cleaved IL-1β protein were detected by immunoblotting using a goat anti-mouse IL-1β polyclonal antibody (1∶2000 dilution; catalog #: AF-401-NA; R&D Systems, Minneapolis, MN) followed by HRP-conjugated chicken anti-goat-HRP (1∶1000 dilution; catalog #: HAF019; R&D Systems).
All assays were performed using anti-Ly6G magnetic bead enriched neutrophils obtained from bone marrow cells of wt or IL-1β-deficient mice as described above. Phagocytosis was measured using pHrodo S. aureus BioParticles (Invitrogen), according to the manufacturer's instructions. Briefly, 1×105 neutrophils were incubated with fluorophore-conjugated S. aureus bioparticles at 37°C for 1 hr. Cells were then stained with FITC-conjugated anti-Ly6G and pHrodo-positive neutrophils were quantified by flow cytometry. For neutrophil degranulation, 1×105 neutrophils were stimulated for 30 min. at 37°C with 1 µM fMLF. Lactoferrin release was quantified from the supernatant using a mouse Lactoferrin ELISA kit (Biotang, Inc., Waltham, MA). Release of neutrophil reactive oxygen species was measured using the Phagoburst kit (Orpegen Pharma, Heidelberg, Germany) according to the manufacturer's instructions. Briefly, 5×105 neutrophils were treated with 1 µM fMLF for 10 min at 37°C. The generation of reactive oxygen species was measured by flow cytometry by gating on 10,000 neutrophil events and determining the proportion of these cells positive for the conversion of the substrate dihydrorhodamine-123 to fluorescent rhodamine-123. Finally, neutrophil killing assays were performed by opsonizing S. aureus with 10% serum from C57BL/6 wt mice and adding the opsonized bacteria to purified neutrophils at a 1∶1 ratio (2×105 neutrophils∶2×105 CFU bacteria) for 45 min at 37°C. After incubation, neutrophils were diluted in H2O (pH 11) to lyse the neutrophils and serial dilutions were plated on TSB agar plates to enumerate viable bacterial CFU. As negative a control, bacteria were also incubated in media without neutrophils.
Neutrophils were obtained from the bone marrow of LysEGFP mice using Percoll density gradient centrifugation. Neutrophils were washed once and resuspended in 1 ml of RPMI 1640 (Gibco) supplemented with 5 µg bisBenzinamide Hoescht 33342 trihydrochloride, a nuclear counterstain, (Sigma-Aldrich) for 30 minutes at room temperature. Neutrophils were subsequently attached onto glass slides using a Shandon Cytospin IV (Thermo Scientific) and imaged using an Olympus ×61 fluorescence microscope. To determine whether EGFP fluorescent signals were altered after neutrophil degranulation, neutrophils from LysEGFP mice were left unstimulated or stimulated with 1 µM fMLF or 100 ng/ml PMA (both from Sigma-Aldrich) for 15 minutes at 37°C. Cells were labeled with a biotinylated anti-mouse Ly6G (University of California San Francisco Monoclonal Antibody Core) with streptavidin-PE (Caltag) and anti-mouse PE-Cy7 CD11b (Biolegend) and analyzed on a Beckman Coulter FC500 flow cytometer.
Data were compared using Student's t test (2-tailed). All data are expressed as mean ± SEM (standard error of the mean) where indicated. Values of *p<0.05, †p<0.01, and ‡p<0.001 were considered statistically significant.
|
10.1371/journal.pcbi.1003741 | Structure-Based Druggability Assessment of the Mammalian Structural Proteome with Inclusion of Light Protein Flexibility | Advances reported over the last few years and the increasing availability of protein crystal structure data have greatly improved structure-based druggability approaches. However, in practice, nearly all druggability estimation methods are applied to protein crystal structures as rigid proteins, with protein flexibility often not directly addressed. The inclusion of protein flexibility is important in correctly identifying the druggability of pockets that would be missed by methods based solely on the rigid crystal structure. These include cryptic pockets and flexible pockets often found at protein-protein interaction interfaces. Here, we apply an approach that uses protein modeling in concert with druggability estimation to account for light protein backbone movement and protein side-chain flexibility in protein binding sites. We assess the advantages and limitations of this approach on widely-used protein druggability sets. Applying the approach to all mammalian protein crystal structures in the PDB results in identification of 69 proteins with potential druggable cryptic pockets.
| Advances reported over the last few years and the increasing availability of protein crystal structure data have greatly improved structure-based druggability approaches. These algorithms predict our ability to discover small molecule drugs for protein targets and can help in identifying promising new biological targets for small molecule drug discovery. However, in practice, nearly all druggability estimation methods are applied to protein crystal structures as rigid proteins, with protein flexibility often not directly addressed. The increasing interest in finding small molecule drugs to protein-protein interfaces makes this issue particularly acute since these interfaces tend to have substantial flexibility compared to traditional enzyme targets. Here, we apply an approach that accounts for light protein backbone movement and protein side-chain flexibility in protein binding sites. We present the results of applying this method to all publicly available mammalian protein crystal structures.
| The majority of small molecule drug discovery efforts towards new, unprecedented biological targets do not progress past high-throughput screening or hit-to-lead optimization due to lack of pursuable chemical matter [1], [2]. To counter this, drug discovery groups increasingly use druggability analysis methods to estimate the amenability of new targets to small molecule drug discovery efforts. In prioritizing new targets, druggability analysis results are then considered along with the strength of evidence that affecting the target will lead to human therapeutic benefit [3]. The results also inform the use of structure-based drug design resources and alternative approaches, such as those involving pro-drugs and covalent interactions, for targets that are expected to be very difficult.
In a drug discovery setting, small molecule druggability is commonly defined as whether a small molecule can bind a desired biological site with good, nanomolar range potency, and, at the same time, also have good, drug-like properties conducive to oral bioavailability and clinical progression [3]–[6]. Thus, the concept refers to chemical tractability of the target. The term, bindability, is also used [7], although the term may not capture the desire for the optimized compound to have drug-like properties. We emphasize that a binding site is not necessarily druggable simply because a ligand binds; the ligand additionally needs to have reasonable drug-like properties and potency. The concepts of ‘druglikeness’ and ‘druggability’ as we use it here cover the most common strategies for small molecule drug discovery, and alternative strategies (e.g., involving covalent adducts, metal chelation, prodrugs, and non-oral delivery) can also be useful in prosecuting targets that are found to be likely ‘undruggable’ when only weak, non-covalent interactions are considered [4].
Druggability estimation has historically been based on precedence, that is, whether there are known drugs targeting the protein or one of its homologs [2], [3]. However, this type of data is scarce or non-existent for many newer protein targets. Advances reported over the last few years allow us to leverage the increasing availability of protein crystal structure data using structure-based druggability approaches. There are at least ten published methods for estimating druggability this way [3], [6], and the body of work is extremely consistent in finding that druggable sites are those that have particular ranges of size, curvature, and hydrophobic character [3]–[6]. These descriptors largely characterize aspects of receptor desolvation [4], and atomistic simulations using molecular dynamics have now shown desolvation to be relevant and sufficient for predicting druggability [8], [9].
Many current structure-based methods for druggability estimation are remarkably accurate if the potential small molecule binding site is largely rigid [3], [6]. Binding sites are not always rigid though, and druggability methods are less accurate if the protein readily changes conformation upon small molecule binding. This is particularly true of protein-protein interface and allosteric sites, where druggable pockets often become exposed only with protein movement [10]–[12]. These ‘cryptic pockets’ are large and shallow when bound to their biological peptide or protein partners, but tend to have high hydrophobic character, and, crucially, have flexibility such that larger, deeper pockets more typical of druggable binding sites are energetically accessible [10], [11].
To begin to address these sites, we apply an approach to modeling conservative movements in pockets using comparative protein modeling approaches coupled closely with structure-based druggability analysis. The approach models relatively light protein motions, involving side-chain flexibility and local protein backbone movements, and maintains reasonable prediction accuracy in retrospective validation studies. It allows us to take pockets that a rigid-protein druggability analysis would deem to have some drug-like properties, but not have sufficient drug-like size, and assess whether local protein motion can result in the pocket having all the drug-like properties, including drug-like size. The approach is computationally efficient enough to enable mining of the structural proteome while taking into account light protein flexibility. Applying the method to roughly 18,000 mammalian protein crystal structures in the PDB results in prediction of one percent of proteins as containing likely druggable cryptic pockets.
We combine a druggability scoring model with protein modeling and docking methods to first identify candidate pockets that have drug-like physicochemical properties but may lack sufficient drug-like size, and then seek out energetically accessible side-chain and backbone motions near these potential pockets using protein modeling approaches.
For determining whether a target pocket has drug-like physiochemical properties, we use an adaptation of a validated druggability score [13]-[15], which we call Dscore+. Dscore+ is a modification of Dscore [13], and we've found that this modification results in good correlation with 19F NMR hit-rates for five newer targets prosecuted at Amgen [14]. Dscore+ is computed from physiochemical descriptors generated from a program, SiteMap, and is a weighted sum with contributions from the degree of pocket enclosure (a surrogate for pocket curvature), pocket size, and the balance between hydrophobic and hydrophilic character in the binding site [13]. In this work, we further validate Dscore+ as a druggability score, and tune the pocket identification parameters in the program, SiteMap, to better identify pockets that may become more druggable with protein motion. In addition to assessing the physiochemical properties of the pocket, we assess reasonable drug-like size by considering the volume of the pocket.
The method consists of the three steps depicted in Figure 1. If a protein site is found to meet a minimal Dscore+ threshold, then residues surrounding the site are put through an iterative flexible protein docking and protein modeling workflow known as induced-fit docking [16] in order to model protein flexibility, including light, local backbone movement and side-chain rotamer conformations. Induced-fit docking is done twice, first using a small naphthalene molecule to reorder side-chains and expose small hydrophobic clefts, and then again using a larger, tetra-substituted naphthalene molecule to further open the cleft if the protein structure allows. We note that others have used fragment docking to assess druggability of pockets in static structures [17], but we are using docking for the different purpose of inducing flexibility in pockets.
We chose naphthalene and a larger tetra-substituted naphthalene because they are hydrophobic and aromatic—known features of drug-like molecules. Additionally, naphthalene is rigid so docking is fast. The tetra-substituted naphthalene molecule we use is a natural progression from naphthalene, and includes four substituents (ethyl, propyl, and cyclohexyl at two positions) that we thought could help in opening pockets. These simple-minded choices performed reasonably well in validation studies, and limited experimentation with a few other molecules gave similar or worse results. In particular, use of benzene in place of naphthalene resulted in a large false positive rate because benzene is small and much more promiscuous, fitting into many small sites. For the larger molecule, we tried five molecules similar to tetra-substituted naphthalene and the results were not substantially different. It is certainly plausible that more systematic experimentation with a larger number of ligands could result in improved performance.
Other approaches that address the issue of protein flexibility for druggability assessment use computational solvent mapping or molecular dynamics simulations, or both. Kozakov et al. used computational solvent mapping with 16 small organic fragment probes to identify small pocket ‘hot spots’ where multiple probes bind in the simulations [18]–[19]. Alternate side-chain conformers are then modeled for selected residues adjacent to the hot spots, and the resulting modeled sites are subjected to a round of computational solvent mapping. ‘Hot spots’ having all 16 probes bound are identified as druggable [19]. Bakan et al. used molecular dynamics simulations for solvent mapping with a different set of small fragment probes and found that the probes bind known allosteric sites during the simulation [20]. Brown and Hajduk earlier showed that molecular dynamics simulations can capture pocket dynamics that result in a more druggable binding pocket in Bcl-xl, while preserving much of the known binding site rigidity of Akt-PH and FKBP [21]. Tools have very recently become available to facilitate druggability assessment in molecular dynamics trajectories [22]–[24]. However, the typically-used short molecular dynamics simulations on the order of 10–30 ns likely do not fully capture protein flexibility relevant to drug binding [25], which typically occurs on much longer timescales [26].
Our approach to flexibly treating potentially druggable binding sites is substantially less compute-intensive, which is important for our goal of analyzing the structural proteome. For a single binding site where flexibility is modeled, our approach requires between one and two hours for most individual protein structures on a current scientific workstation with a four core CPU. In contrast, a 30 ns molecular dynamics simulation on a single protein would require about a week, and the computational solvent mapping approach using FTMAP requires about half a day for each protein binding site since FTMap must be run for each discrete side-chain configuration and each configuration requires about two hours on a single CPU core [27]. Our approach runs relatively quickly due to the use of comparative protein modeling techniques instead of more resource-intensive methods that attempt to directly simulate biophysical phenomena. Our approach finds very few sites that open up significantly—less than two percent of protein structures across the mammalian structural proteome. The false positive rate is reasonably low; we see a 12% false positive rate in our protein-protein interaction validation set and a 0% false negative rate. In contrast, it is likely possible to find all known druggable sites using molecular dynamics, including some that are undetectable by our method. However, separating the signal from the noise is challenging [20], [21] since there is a tendency for many pocket openings to be observed. To be clear, our approach does not reveal any new pockets that do not already exist within the rigid structure. It does, however, locate small sites that do not meet the druggability criteria initially but can meet the druggability criteria when conservative protein flexibility is modeled.
While accurate modeling of protein motion continues to be difficult and the subject of substantial research, we found the approach we present here to be sufficiently accurate and efficient for the purposes of mining the structural proteome. We note that previous efforts we are aware of to identify the “druggable genome” rely on sequence-similarity to known druggable proteins [2]. Structure-based druggability analyses are based on “first principles” and are thus complementary to precedence-based sequence-similarity approaches.
We applied the method to two widely-used druggability validation sets to check its performance and measure any increase in false positive rate due to allowance of protein flexibility.
The first validation set is a published set that covers a variety of targets, and consists of 27 targets: 17 druggable targets and 10 difficult targets [4]. A histogram of the druggability scores, Dscore+, based on the original crystal structures, with no flexibility modeling, is shown in Figure 2a. The plot supports a Dscore+ >1.3 threshold for druggable versus difficult targets, with higher scores roughly indicating more druggable sites. Modeling protein flexibility for target sites that meet a threshold of Dscore+>1.3 results in an increase in Dscore+, but the increase is relatively systematic (mean = 0.4, σ = 0.3) and appears to be consistent enough that a useful differentiation between difficult and druggable targets is retained.
We also investigated the effect of flexibility modeling on targets with scores of Dscore+≤1.3. For these additional targets, we again find an increase in scores by an average of 0.4 (σ = 0.3). Thus, difficult and druggable targets in the validation set can still be distinguished after flexibility modeling, although the distinction is less crisp than it was when scoring rigid structures. Comparing the score distributions for druggable and difficult targets using the two-sample Kolmogorov-Smirnov (K-S) statistic finds that the scores are significantly different from each other, both with and without flexibility modeling (p-values of 8×10−4 and 1×10−6, respectively). When we subtract 0.4 from each score that includes protein flexibility, the score distributions remain similar within each set. For druggable targets, the means of the with- and without-flexibility scores are 1.6 and 1.6 respectively, with variances of 0.05 and 0.07. For difficult targets, the means of the with- and without-flexibility scores are 1.2 and 1.1 respectively, with variances of 0.04 and 0.03.
Taken together, the results suggest that a Dscore+ threshold of 1.7 (i.e., 1.3+0.4) should be applied to sites resulting from flexibility modeling, and this threshold is depicted in Figure 2b by a red line. We will show that this threshold value is strongly supported by analysis of a larger number of proteins (109 proteins) from the mammalian proteome druggability results. While the threshold is determined empirically and the increase is not ideal, we can rationalize the increase as due to modeling of protein flexibility with an impetus towards making the pocket more hydrophobic.
Turning to pocket volumes, the method should not lead to all pockets increasing significantly in volume, consistent with the belief that some sites are inherently flexible while others are less so. In this first validation set, which is composed largely of enzyme active sites, the average volume before and after flexibility modeling are both about the same (420 Å3 and 360 Å3, respectively, with standard deviations of 190 Å3 and 130 Å3), and are both within the drug-like range, as discussed later. In the second, validation set of protein-protein interfaces, we will see that the binding site volumes change more significantly. Resulting volumes tend towards a volume of around 300–400 Å3, if the flexibility of the protein allows, and this appears to be related to the size of the second ligand (tetra-substituted naphthalene) used in the induced-fit docking step of the flexibility modeling. In the mammalian proteome analysis, we find that only one percent of proteins analyzed have cryptic pockets that change substantially from a volume substantially below the drug-like range (≤100 Å3) to a drug-like volume (160–800 Å3). In developing our approach, the drug-like volume range was initially set roughly to 150–600 Å3 based on our judgment, and later refined to 160–800 Å3 based on quantitative analysis of the mammalian proteome results.
The second validation set addresses protein-protein interaction (PPI) targets, and includes six targets from the 2P2I database and Wells et al. (2007): Bcl-xL, HDM2, IL-2R, HPV E2, ZipA, TNFα [10], [12]. Bcl-xL and HDM2 are classified as druggable since oral small molecule inhibitors have advanced to clinical trials. We argue that the remaining potentially high-value targets are difficult. While pioneering small-molecule inhibitors have been reported, we note that substantial efforts made over the last 15 years to identify inhibitors of TNFα, IL-2R, HPV E2, and ZipA have not resulted in reported small molecule clinical compounds [10]. Given just the protein crystal structures, with no information on location of binding sites, the method successfully opens the relevant binding pockets for Bcl-xL, HDM2, and TNFα and scores them as druggable based on Dscore+ and volume considerations, as shown in Table 1. Calculated values of Dscore+ and volume that fall within the defined drug-like range are highlighted in bold. In the cases of Bcl-xL and TNFα, light flexibility results in small molecule binding pockets with roughly 50% and 100% larger volumes, respectively. Bcl-xL would have been classified as difficult based on the 2bzw PDB structure without additional flexibility modeling because the pocket size (112 Å3) is too small by any reasonable criteria for drug-like volume size. Flexibility modeling results in small changes in the binding site that, together, increases the volume of the pocket to a reasonable volume (172 Å3). Bcl-xL is perhaps the one clear example where a protein pocket opens substantially and the druggability is known (i.e., orally administered small molecule inhibitors have progressed to clinic).
We also analyzed all targets listed in 2P2I where crystal structures are provided, but some targets have either unclear experimental druggability because efforts on the targets are more recent, or known inhibitors involve metal chelation. The results for these additional targets are included in Table S1.
Comparing Dscore+ and pocket volume calculation results with and without protein flexibility modeling finds that Bcl-xL (PDB IDs: 2bzw, 2yxj, 3qkd, 4ehr) and a minority of HDM2 structures (PDB IDs: 1rv1, 3lbk) would have been missed without the additional flexibility modeling to open up pockets to a drug-like volume. Interestingly, one IL-2Rα structure (PDB ID: 1pw6) has a Dscore+ that places the target in the low end of the druggable score range, but the pocket volume does not satisfy the drug-like criteria, and this remains the case after protein flexibility modeling. Protein flexibility modeling does not always open pockets significantly.
With TNFα, the known pocket at the trimer interface was identified as the top pocket in the apo-structure, and flexibility modeling resulted in a binding site with good druggability score and good drug-like volume. This result is consistent with the scores obtained using the co-complex structure with SPD-304 [28]. However, the best reported inhibitor has only single digit micromolar range potency against TNFα [28], and although there is not really sufficient data currently to determine this, it is possible that the calculations overestimate the druggability of the pocket.
With Bcl-xL, comparison of a BAD peptide-bound structure (PDB ID: 2bzw) with a small molecule-bound structure (PDB ID: 2yxj) shows that two residues, Phe105 and Leu130, adopt alternate conformations, and the helix around Leu108 becomes disordered to create the ligand binding pocket [29], as shown in Figure 3. The target serves as a good illustrative example of our complete approach. First, a potentially druggable site is identified regardless of whether it is too small to hold a drug-like molecule. This is followed by induced-fit docking of naphthalene to the identified site, which moves residues, including Phe105, as shown in Figure 3b. A second induced-fit docking of the larger (molecular weight of 363 Da) tetra-substituted naphthalene (TSN) results in a total of four models, where we see additional movements in addition to Phe105. A representative model is shown in green in Figure 3c, and shows a Leu130 rotamer change and backbone movements around Leu108 resulting in loss of the alpha-helical secondary structure. However, the modeled structures still differ from the ligand-bound crystal structure, as shown in Figure 3d, and the model, in this case, is effectively a hybrid of the peptide-bound structure and the known ligand-bound structure. Thus, the modeling approach, in the case of Bcl-xL, successfully allows backbone motion and reproduces some of the known side-chain and backbone movements in the resultant models. We note that the TSN molecule makes similar interactions compared to the ligand, ABT-737, in the ligand-bound crystal structure. Despite not reproducing all of the atomic details of the ABT-737 crystal structure, the flexibility modeling captures many key features and the inherent flexibility of the pocket that results in an increased binding site volume and increased druggability score.
To investigate the behavior of the flexibility modeling approach on targets where protein flexibility is known based on crystal structures, we applied the method to a set of protein crystal structures with binding site flexibility from Huang and Jacobson [17], where we've selected targets that show RMSDave >1.5 Å for binding site residues in different crystal structures of the same protein. We compare these results with their published druggability results, which do not account for flexibility, in Table 2. While the two methods perform similarly on non-PPI targets (first six targets in Table 2), the flexible druggability method performs better on PPI targets (last five targets in Table 2). In particular, their docking-based druggability model predicts IL-2 and HPV E2 to be druggable (docking hit rate>0.36) based on some structures, but predicts the same targets to be very difficult based on other structures. The flexibility modeling approach results in classification of both targets as very difficult, consistent with what is known, as previously discussed.
Examining the variation in scores between different crystal structures of the same target finds that while both the static protein and flexible protein methods yield similar score variation for non-PPI targets, they have substantially different variation with PPI targets. In particular, the docking hit-rate method shows large variation in score among structures of IL-2 (107%), MDM2 (69%), and HPV E2 (323%) compared with a median variation of 21% in all 11 targets. The flexibility modeling method, on the other hand, results in score variation on PPI targets that is consistent with that found with non-PPI targets.
Overall, the docking method has a median score variation of 21% with a standard deviation of 94% in the dataset, while the flexibility modeling approach has a median score variation of 13% with a standard deviation of 10%. Yet, when the PPI targets are removed, the two methods have comparable score variation. Taken together, the flexibility modeling method appear to provide more reliable, consistent predictions at PPI interfaces, and this makes sense because PPI interfaces are much more likely to involve substantial protein flexibility [11]. Accounting for protein flexibility in a conservative manner, as we have done, leads to more consistent druggability predictions.
We next applied the flexibility modeling approach to all publicly-available crystal structures containing mammalian proteins to estimate the number of druggable targets and identify potential druggable cryptic pockets. Analyzing the over 18,000 structures in the Protein Data Bank (PDB) [30] as of June 30, 2012, required approximately 35,000 CPU-days of calculation (in aggregate) on our internal Linux clusters. The analysis covered not only the crystal structures as deposited in the PDB, but also all individual monomers in the case of multimer assemblies. Five percent of PDB files generated an error, due mostly to structures containing only Cα atoms (i.e., no protein side-chains) or involving large biomolecular assemblies, since we stopped calculations on a particular PDB entry if it ran for more than fourteen days.
The results are summarized in Table 3, where we also mapped PDB chains to SWISS-PROT ID's to determine the number of proteins represented. Of the 17,834 PDB entries analyzed, 42% had at least one site that met the Dscore+ >1.3 threshold for further protein flexibility modeling. Twenty percent of mammalian proteins in the PDB have a potentially druggable pocket. Of the 5,739 PDB entries (1,134 proteins) that have a predicted druggable pocket, about two–thirds would be predicted druggable based on the original, static crystal structure.
To identify druggable pockets with the greatest likelihood of biological relevance, we winnowed the list to protein sites in 2095 PDB structures where a small molecule could potentially disrupt a known intermolecular interaction. The interacting partner should be transiently-bound (as opposed to obligately-bound) and can be a protein, natural co-factor, natural ligand, or synthetic ligand. These sites are either at protein-protein interfaces or contain a small molecule of molecular weight less than 1000 Da. Including these criteria gives us higher-confidence druggability predictions and may remove many false-positives, but could result in removing sites that are functionally relevant but perhaps not well-characterized. For sites at protein-protein binding interfaces, we assessed whether the relevant protein-protein interaction is an obligate or transient interaction based on a published database, Interevol [31]. Sites involving obligate dimer interfaces were removed, but sites without any prediction or assessment were retained; there was no annotation for over half of the protein interfaces we considered.
Overall, we identified predicted druggable pockets in 2,095 PDB structures representing 730 unique proteins. In Figure 4, we depict the breakdown of predicted druggable pockets at these intermolecular interfaces with pockets where a bound ligand would disrupt a protein-protein interaction (including protein-peptide interactions) shown in blue, and pockets where a bound ligand would disrupt a protein-ligand interaction shown in red. The purple overlap region indicates protein pockets that are at both a protein-ligand and protein-protein interface. Some of these pockets are adjacent to small peptides, which can be classified as both a ligand and protein by our definition; ligands are defined as any molecule with molecular weight of 1000 kDa or less, while proteins are defined as non-HETATM molecules. The number of unique proteins in the purple region is much higher because a given protein may not only have a co-crystal structure solved bound to a protein partner, but also bound to a small molecule ligand (or vice-versa).
To identify cryptic pockets, we looked at potential druggable pockets that were small (volume ≤100 Å3) in the static structure, as long as the initial cavity was not fully buried (enclosure ≤96%). Less than 20% of these, representing 105 structures, met the flexible druggability criteria, opening up to at least 160 Å3 with flexibility modeling. These targets representing 69 unique proteins are provided in Table S2.
To compare the mammalian PDB results to a positive control set, we mapped known oral drugs from MDDR (2008 release) to ligands in known crystal structures. Of the 421 oral drugs administered in tablet form, 109 could be mapped to PDB co-crystal structures that had crystallographic resolution ≤2.5A. The 102 pockets with ligand overlap to known co-crystalized ligands (ligand overlap >0) are plotted by volume in Figure 5A, where volume is computed using SiteMap. Targets at the low end of the volume range, with volumes of 160 Å3 or less, include four complexes with large FK-506 natural product analogs that are not captured by the drug-like binding site definition in use. Targets with volumes of 800 Å3 or greater are largely natural product complexes as well (macrocyclic antibiotics, reservatrol, and others). The results therefore suggest a drug-like volume range of between 160 and 800 Å3 is appropriate for the approach we use. We note that volume calculations are highly sensitive to the algorithm used, and so these volume ranges should be established independently for different implementations of our approach.
The range of druggability scores for the known oral tablet drug set versus all pockets is shown in Figure 5B, where the top histogram represents the oral tablet drug set. The distributions are overlapping, and while the means of 1.7 and 1.4 are significantly different (p = 6×10−15 based on the two-sample K-S test), the 95% confidence intervals overlap. While the large-scale data shows there is room for improvement in the separation of druggable and difficult targets, the 1.3 Dscore+ cut-off we use is nevertheless useful for identifying druggable pockets in rigid proteins, and removing 60% of pockets from further analysis with more resource-intensive flexibility modeling. Applying flexibility modeling to the MDDR targets also results in a shift in Dscore+ range, shown in Figure 5C, similar to what is seen with the smaller general validation set. The shift seen here further supports use of a Dscore+>1.7 cut-off in conjunction with protein flexibility modeling.
To assess the effect of our flexibility modeling approach on pocket volumes, we looked at all pockets at intermolecular interfaces before and after flexibility modeling and show the results in Figure 6. A diagonal white line indicates no change in volume. While the modeling method likely overpredicts volume increases in pockets, the majority of pockets that increase in volume increase by less than 50 Å3. The vertical and horizontal white lines in Figure 6 indicate the 160 Å3 volume cut-off, and it is clear that most pockets under the cut-off remain under the cut-off. As expected, pockets with volumes closer to the cut-off, with volumes of 100–160 Å3, are the most likely to increase to over 160 Å3 with protein flexibility modeling.
In Figure 6, pockets with original volumes less than about 200 Å3 tend to get larger, while pockets with original volumes greater than about 400 Å3 tend to get smaller. The likely rationale is that the tetra-substituted naphthalene ligand used in the flexibility modeling approach induces smaller pockets to grow to accommodate the ligand, while it induces larger pockets to shrink to better enshroud the ligand. These tendencies are, however, dependent on the inherent flexibility of the protein structure.
While the analysis provides a good set of putative druggable proteins in the mammalian structural proteome, we are not blind to deficiencies in this analysis. The prediction error rate in the large mammalian structural proteome analysis is hard to know, and we discuss the limitations in the next section.
The automated approach to protein flexibility we report here is useful for identifying druggable targets in the structural proteome. We are aware of three areas for further improvement.
The first is related to pocket selection. Pocket selection is based on geometric considerations, and the pockets are subsequently scored for druggability using Dscore+ as well as, potentially, protein flexibility modeling. Ideally, the pocket selection and scoring would be done simultaneously to yield pockets that maximized the druggability score [32]. This issue, for instance, has an effect on scoring of phosphodiesterase active sites such as those in PDE-4D and PDE-5, where protein residues at one end of the catalytic site are very polar, and known oral inhibitors do not interact in this region [32]. Figure 7a shows the binding site including the polar region that results in a low druggability score, Dscore+ = 1.4, which is not representative of the druggability of the binding site. Removing the polar region shown in Figure 7b results in a more representative druggability score, Dscore+ = 1.6. We were not successful in adjusting our protocol to account for this, and thus we may miss druggable sites that are similarly amphipathic in nature. In addition, despite our efforts at tuning the pocket identification algorithm, SiteMap does, in about 2% of cases, return candidate pockets with volumes over 800 Å3, the drug-like size limit that we use. Currently we simply remove these pockets, which may cause us to miss druggable sites. Future work to resolve these issues include modifying SiteMap to identify only pockets of the desired volume (160–800 Å3), allow for pockets that overlap with each other, and account for properties such as hydrophobicity in the pocket definition process.
A second area relates to the false positive rate, that is, the fraction of pockets identified as druggable that are not truly druggable. Even though we restrict protein flexibility to side-chain motion and localized backbone movement, the protein flexibility modeled and our selection of proble molecules are biased towards increasing the hydrophobicity of the pocket under analysis, and relaxation of the resultant structures may improve results. In addition, the degree of protein flexibility modeled is probably more than that present in reality. In this work, we empirically compensated for these issues by measuring the impact of flexibility modeling in Figures 2 and 4, which led to the use of a Dscore+≥1.7 criteria. While our flexibility modeling approach demonstrates statistically significant discrimination of difficult and druggable targets, we also plan to explore approaches such as protein relaxation [33], to remove the need for an empirical correction. The flexibility modeling approach is more likely to exaggerate the flexibility of smaller proteins due to fewer stabilizing interactions within the protein. For the cryptic pocket analysis, we restricted our results to those proteins that are greater than 100 amino acids in length (which translates to about 10 kD).
Lastly, we need to consider that the protein structures observed in crystal structures, in a minority of cases, may not be the biologically relevant constructs or complexes. Crystal structures may be synthetic constructs or portions of proteins, which, in the context of the full-length protein, have predicted binding sites occluded. Similarly, biological obligate dimers not seen in the crystal structures can occlude the binding site. Co-factors can also affect the druggability of binding sites; here, we only account for selected, particularly tight-binding co-factors such as metals and hemes. We analyze both biological assemblies defined in the PDB as well as the individual monomer components to account for binding to intact complexes as well as unbound partners. Other partially dissociated complexes may exist however. In addition, we are looking at binding, and not functional effects of binding; weak binding at an allosteric site is sometimes sufficient to generate the desired inhibition or activation of biological activity [32].
We leverage advances in druggability assessment and modeling of protein flexibility to create an approach that allows light flexibility in the protein backbone and side-chains. The method improves the accuracy of druggability assessments when tested on two validation sets representing general pharmaceutical targets and protein-protein interactions of pharmaceutical interest. Combining this with the wealth of crystal structures available in the PDB allows us to find new protein binding sites that are potentially druggable by small molecules. Searching for such sites is thought to be like finding needles in a large haystack, and a systematic, automated approach is thus useful. Accurate modeling of protein flexibility continues to be difficult and the subject of substantial research. Even so, our approach is useful in exploring induced druggable pockets and provides a substantial number of hypotheses. For applications focused on analysis of protein pockets, the approach we take is computationally efficient and may be complementary to comprehensive analyses of static crystal structures [34]. Finally, we have long been intrigued by the possibility of combining the druggability data with biological target disease-relevance data. This has recently been done on a small-scale with cancer targets [35], although protein flexibility was not accounted for in the druggability assessment. Inclusion of protein flexibility in such assessments can help to provide more accurate target assessment.
Protein structures were downloaded from the biounits repository at the RCSB based on criteria that the structure (1) contains protein, (2) is categorized as deriving from the class Mammalia, and (3) has an x-ray crystal structure resolution ≤2.5A.
Ligands, defined as having molecular weight ≤1000 Da, are removed with the exception of heme groups, zinc, and magnesium (PDB het groups HEM, MHM, HEV, VER, SRM, HEO, HEB, HEC, HDM, HDD, DDH, ZN, MG). Protein structures were prepared using Schrödinger Protein Preparation Wizard (version 2012, Schrödinger LLC, New York, NY), on the command line with the following options: –watdist 0, –fillsidechains, –rehtreat, –mse, –noepik, –noimpref. These options assign bond orders, add hydrogens, remove all waters, create zero-order bonds to metals, create disulfide bonds for close cysteines, mutate selenomethionines to methionine, fill in any missing side-chains with Prime (v3.1, Schrödinger LLC, New York, NY), and optimize hydrogen placement and polar residue flips using PropKa. Validation test runs using restrained minimization to a heavy atom RMSD of 0.3 Å, a procedure known as “Impref”, did not change which sites were found and did not significantly change druggability scores on the validation dataset proteins, so we chose to increase workflow speed by avoiding this step in the protein preparation.
Next, initial potential druggable surface patches were identified using Schrödinger SiteMap (v2.6, Schrödinger LLC, New York, NY), the results of which are used to compute Dscore+. We run Sitemap with a fine grid (0.35 Å spacing) and “loose” definition of hydrophobicity. In this study, all calculations were performed from the command line with options that return the 5 largest SiteMap sites, in order of the number of site points they contain. Our modified settings allow more shallow binding sites to be found and include binding site regions with slightly weaker vdW interaction energy. We used the following non-default Sitemap parameters: maxdist = 10, enclosure = 0.4, maxvdw = 1.0, dthresh = 5.0, mingroup = 7, nthresh = 7, grid = 0.35, modphobic = 0. The smaller value of maxvdw (default is 1.1 kcal/mol) and the less restrictive definition for modphobic of zero together allow gridpoints with slightly weaker vdW interaction energy to be included as sitepoints. The smaller enclosure score (default is 0.5) and larger maxdist value (default 8.0 Å) allow more shallow binding sites to be found. The enclosure score is computed by drawing radial rays from each sitepoint, and the score is the fraction of rays that strike the receptor surface within a distance of 10 Å (maxdist), averaged over the sitepoints. Decreasing dthresh from the default (6.5 Å) and increasing nthresh from the default (3) causes SiteMap to return smaller, more compact sites than it otherwise would when using a fine grid. When considering a gridpoint for inclusion in a site, there must be at least nthresh other points within 1.76 Å (square root of d2thresh) for it to be considered. When considering whether two sites should be joined, the closest points in the two sites must be closer than dthresh. The parameter, “mingroup”, is the only parameter here that limits the number of sites found; this is the minimum number of points in a site-point group required to constitute a site (default = 7). We found that including sites with less than seven points in combination with a fine grid of 0.35A resulted in merging of many very small pockets to form long, stringy sites that were not realistic as small molecule binding sites. Overall, these modified SiteMap settings allow us to find shallow pockets with less hydrophobic character than is possible to find with default settings.
From the SiteMap results, sites identified with a druggability score, Dscore+, of greater than 1.3 are taken as candidate sites regardless of volume, where Dscore+ is defined as Dscore + 0.3*hydrophobic, as previously described [14], [15], and druggability scores are rounded to the first decimal place. The choice for the 1.3 value is discussed in the Results and Discussion section. Dscore is computed from physiochemical descriptors generated by SiteMap, and is a weighted sum with contributions from three components, (1) degree of site enclosure, (2) pocket size defined by the number of site points included in the site; site points are x, y, z coordinates that are outside the protein, are reasonably enclosed, and have a vdW interaction potential over a defined threshold are clustered into sites, and (3) a negative contribution from the hydrophilic score, which limits the impact of hydrophilicity in charged and highly polar sites [13].
To identify binding sites with potential flexibility, we used an iterative protein-modeling and docking approach [16], available in the Schrödinger Suite (2012 release, Schrödinger LLC, New York, NY) as the induced fit docking (IFD) workflow, and applied this using the ligands in Figure 8 to each candidate site. In this study, all calculations were performed from the command line with the default IFD parameters, except the variable, OUTER_BOX, which is always set to 25 Å, since we were docking the same small ligands. We defined the variable, BINDING_SITE, by a single sitepoint which is placed at the centroid of all SiteMap sitepoints from the candidate site. First, we used IFD to dock a naphthalene molecule, 1, to the top 5 sites found by SiteMap and kept the best two naphthalene poses for each site. If poses were returned for naphthalene, we then used IFD again to dock a tetra-substituted naphthalene molecule, 2, to the same pocket, now opened up by naphthalene. SiteMap was then applied to score the sites in the four top-scoring structural model results (typically, at least ten models were generated per site). Increasing the number of models can result in better predictions of binding site conformations, but we chose to produce a smaller but reasonable set of four models to reduce the compute time required to process all mammalian crystal structures.
To analyze the druggability and protein-protein interaction validation data, we automatically compared each SiteMap site to the corresponding ligand-bound structure using the Phase (version 3.4, Schrödinger LLC, New York, NY) command-line utility phase_volcalc to compute the overlap (measured in Å3) between the SiteMap sitepoints and the bound crystal ligand. After the IFD steps, we used the same utility to compute the overlap between the tetra-substituted naphthalene and the bound crystal ligand. This value is positive when there is direct overlap between the two sets of atoms. For the validation studies only, we identified the relevant protein biological assembly based on the known literature, and only retained those assemblies or protein monomers that are biologically meaningful. The calculations were otherwise performed automatically.
For calculations run on all mammalian PDB structures, we used a purely automated procedure applying the method to the first “biological unit” as defined in the PDB. Calculations were performed on commodity cluster hardware running RedHat Enterprise Linux. Failed calculations were re-run up to five times, including at both Amgen and Schrödinger facilities, to ensure that failures were not the result of compute infrastructure issues. To identify protein-protein interaction interfaces, we checked whether any of the TSN molecules modeled into a predicted druggable site also overlapped with another protein chain in the crystal structure. Overlap was defined as at least one atom of the TSN molecule being within 2 Å of the additional protein chain, where hydrogens were included. To identify protein-ligand interfaces, we used the previously-described volume overlap calculation. Finally, to analyze the results of the mammalian proteins in the PDB for obligate dimers, we used the Interevol database, publicly available at http://biodev.cea.fr/interevol/interevol.aspx [31]. We downloaded the database (July 2012 release) and joined the data with our PDB results by matching both the four-letter PDB code and any chain identifier. PDB IDs were translated to gene ID's using the SWPROT database [36], and all gene annotations were performed using bioDBnet [37] Structure figures were produced using PyMOL version 1.4.1 [38].
To map MDDR drugs to PDB co-crystal structures, we first identified all oral drugs in MDDR that were annotated as ‘marketed’ and delivered orally as tablets or pills. PipelinePilot (ver. 8.5., Accelrys Software, San Diego, CA) was used to identify identical compounds based on structural identity when compared with the SMILES strings included in the HET code file downloaded from RCSB LigandDepot [30]. PDB codes were then identified that corresponded to matched HET codes.
Matlab version 7.9 (R2009b, The Mathworks Inc., Natick, MA) was used to generate Figures 2, 5, and 6, and also to calculate statistical means, variances, and two-sample Kolmogorov-Smirnov test results for the general validation set.
Calculations were performed on Intel Xeon CPU (2.7GHz) multi-core processors running RedHat Enterprise version 6. CPU timings quoted in the paper are per single core.
We thank Nigel Walker, Philip Tagari, Yax Sun, Paul Kassner, Mike Ollman, and Astrid Ruefli-Brasse at Amgen, and Woody Sherman and Alessandro Monge at Schrödinger for their support and helpful discussions.
|
10.1371/journal.pntd.0004261 | Time-Dependent Transcriptional Changes in Axenic Giardia duodenalis Trophozoites | Giardia duodenalis is the most common gastrointestinal protozoan parasite of humans and a significant contributor to the global burden of both diarrheal disease and post-infectious chronic disorders. Although G. duodenalis can be cultured axenically, significant gaps exist in our understanding of the molecular biology and metabolism of this pathogen. The present study employed RNA sequencing to characterize the mRNA transcriptome of G. duodenalis trophozoites in axenic culture, at log (48 h of growth), stationary (60 h), and declining (96 h) growth phases. Using ~400-times coverage of the transcriptome, we identified 754 differentially transcribed genes (DTGs), mainly representing two large DTG groups: 438 that were down-regulated in the declining phase relative to log and stationary phases, and 281 that were up-regulated. Differential transcription of prominent antioxidant and glycolytic enzymes implicated oxygen tension as a key factor influencing the transcriptional program of axenic trophozoites. Systematic bioinformatic characterization of numerous DTGs encoding hypothetical proteins of unknown function was achieved using structural homology searching. This powerful approach greatly informed the differential transcription analysis and revealed putative novel antioxidant-coding genes, and the presence of a near-complete two-component-like signaling system that may link cytosolic redox or metabolite sensing to the observed transcriptional changes. Motif searching applied to promoter regions of the two large DTG groups identified different putative transcription factor-binding motifs that may underpin global transcriptional regulation. This study provides new insights into the drivers and potential mediators of transcriptional variation in axenic G. duodenalis and provides context for static transcriptional studies.
| Giardia is the most common gastrointestinal protozoan parasite of humans. This parasite causes diarrheal disease and is correlated with post-infectious conditions such as irritable bowel syndrome. In the absence of a vaccine, treatment is limited to drugs such as metronidazole, against which clinical resistance is reported. Effective control of Giardia requires a detailed understanding of its biology, and in turn, complete characterization of the standard in vitro culture system. Using RNA sequencing assisted by informatics to functionally annotate hypothetical proteins, we investigated transcriptional changes in axenic Giardia trophozoites at three growth phases over 96 hours. We found two large groups of differentially transcribed genes that indicate changes in the antioxidant system and central carbon metabolism over time. A putative novel signaling pathway may act together with putative transcription factor-binding motifs to regulate these transcriptional changes. Our results suggest that dissolved oxygen in Giardia culture medium may cause oxidative stress early during in vitro growth and that oxygen depletion may limit the efficiency of glycolysis in the declining phase. This work enhances our understanding of the transcriptional flexibility and metabolism of Giardia in vitro.
| Giardia duodenalis (syn. G. lamblia or G. intestinalis) is a gastrointestinal protozoan parasite, and a major cause of chronic infectious diarrhoea in the developed and developing world. G. duodenalis infects approximately one billion people world-wide, causing 200–300 million reported clinical cases each year [1]. G. duodenalis is proposed to account for ~15% of cases of childhood diarrhoea in developing countries [2]. High rates of chronic diarrhoea in the first two years of life is significantly associated with physical and cognitive 'stunting,' and predisposes sufferers to a variety of adult-onset metabolic disorders [3]. In particular, infection with G. duodenalis is associated with post-infectious gastrointestinal disorders such as irritable bowel syndrome, chronic fatigue, and obesity [4,5]. Control of giardiasis depends primarily on chemotherapeutic treatment with one of two major drug classes: nitroheterocyclics (e.g., metronidazole) and benzamidazoles (e.g., mebendazole) [6,7]. Treatment failure rates as high as 30% are reported for these compounds [8,9], and in vitro resistance to widely used chemotypes is documented in isolates from treatment-refractory patients (reviewed in [8,10]). The recently reported increasing incidence of metronidazole treatment-failure in travellers returning to the United Kingdom [11], and toxicity associated with most nitroheterocyclics [6], highlight the need for continued development of anti-giardial drugs. This in turn requires a thorough understanding of the molecular biology of the parasite. Aside from its medical importance, G. duodenalis is thought to belong to one of the earliest eukaryotic lineages, and therefore serves as a useful model for studies of eukaryotic features such as secretory and organellar protein trafficking [12], cellular differentiation [13,14] and RNA interference [15–17].
G. duodenalis can be cultured in complex, host cell-free media, which is a rarity among parasitic protists and of great advantage for conducting molecular research. Axenic culture provides an excellent system in which to explore the biology of G. duodenalis over time and in response to external stimuli, drug perturbation [18–21] and other stressors [22,23]. System-level transcriptomic investigations based on microarray or serial analysis of gene expression, have established that G. duodenalis trophozoites exhibit clear transcriptional responses to encystation medium [13], protein folding stress [23] and the presence of intestinal epithelial cells [24,25]. More recently, RNA sequencing (RNA-seq) has been used to identify transcriptional start-sites [26], 3’ un-translated regions and polyadenylation variants in mRNA [27], and to compare transcription between different G. duodenalis assemblages [27]. RNA-seq has also recently been applied to investigate the transcriptional response to oxidative stress in trophozoites [28], and to ultraviolet irradiation in trophozoites and cysts [29]. However, considering the complex nature of the standard culture medium (TYI-S33) for G. duodenalis trophozoites, and variation between laboratories in how this medium is prepared, comparing studies is challenging, particularly given that each study represents a static time-point observation. Understanding how transcription varies over time in TYI-S33 medium is important to provide context to single time-point studies. In terms of the axenic growth of the parasite, such research can also provide insight into the changes in the metabolic behaviour and demands of G. duodenalis during different growth phases.
A major challenge for genomic investigations of divergent organisms such as G. duodenalis relates to the vast numbers of functionally un-annotated gene products. Indeed, around 50% of the proteins predicted in this protist lack functional information. In lieu of in vitro characterization, computational protein structure-prediction approaches can provide substantial insight into the putative function of hypothetical proteins [30]. Here, we used RNA-seq coupled with structural homology-based protein annotation, to investigate the longitudinal transcriptional behavior of G. duodenalis assemblage A (WB isolate) trophozoites under standard laboratory conditions in TYI-S33 medium. This represents the first high-resolution, longitudinal transcriptional data set for this protist. We hypothesize that differential transcription will be evident between log, stationary, and declining growth phases, and that these changes will reflect the metabolic preferences of G. duodenalis and the pressures of resource exhaustion.
Giardia duodenalis trophozoites (assemblage A, strain WB-1B; [31]) were generously provided by Drs Jaqueline Upcroft and Peter Upcroft and maintained in axenic culture in filtered, complete modified TYI-S33 medium in close-capped t25 flasks (Falcon) according to standard protocols [32]. The growth kinetics of attached trophozoites was charted over 96 hours (h; S1 Fig), from which the following growth phases were estimated: lag (0–24 h), log (24–60 h), stationary (60–72 h) and declining (72–96 h). As attached and suspended populations exhibited generally similar growth dynamics, we focused on the attached population in order to enrich samples for viable trophozoites and minimize the risk of contamination with degraded mRNA from dead cells. In our hands the generation time of WB1B was 5.4 ±1.2 h during log phase. Samples for sequencing were generated on four different weeks as follows. Nine t25 flasks (64 mL total capacity) were filled with 56 mL of medium, inoculated with 105 trophozoites from confluent t25 flasks, and incubated at 37°C. At 48, 60 and 96 h after inoculation, three flasks were selected at random, and supernatant and suspended cells were discarded and replaced with ice-cold phosphate-buffered saline (PBS). Flasks were incubated on ice for no more than 5 minutes to ensure detachment of trophozoites, and the suspensions were transferred to 50 mL falcon tubes and pelleted at 680 x g for 5 min at 4°C. Supernatants were discarded, and pellets were combined through re-suspension in 1 mL of PBS before transfer to a 1.5 mL Eppendorf tube. The suspension was pelleted (770 x g, 2 min, 22–24°C), re-suspended in 1 mL of TriPure reagent (Roche), and stored at -80°C.
RNA was extracted from TriPure reagent according to the manufacturer’s instructions within four weeks of sample preparation. The dried RNA pellet was re-suspended in reverse-osmosis deionized water (H2O) and treated with Turbo DNAse (Ambion) according to the manufacturer’s instructions. The DNAse-treated RNA was electrophoresed, and large and small subunits of nuclear rRNA bands were examined as a proxy for RNA integrity. RNA concentration was estimated by fluorometry (Qubit) and further quality control was performed using a BioAnalyzer (Agilent). Polyadenylated RNA was purified from 10 μg of total RNA using Sera-mag oligo(dT) beads, fragmented to a length of 100–500 bases, reverse transcribed using random hexamers, end-repaired, and adaptor-ligated, according to the manufacturer's instructions (Illumina). Ligated products (~300 bp) were excised from agarose and PCR-amplified. Products were purified over a MinElute column (Qiagen) and paired-end sequenced (100 bp; non-normalised cDNA) using the Ilumina HiSeq 2000 (Yourgene Biosciences, Taiwan).
Adapters were trimmed from raw reads using Trimmomatic [33] (sliding window: 4 bp, minimum average PHRED quality: 20; leading and trailing: 3 bp; minimum read length: 40 bp), and overlapping read pairs were merged using SeqPrep (downloaded 2 June 2014 https://github.com/jstjohn/SeqPrep) with default parameters. The merged reads were combined with unpaired and non-overlapping paired reads from Trimmomatic output, and all were mapped as single-ended reads to the accepted G. duodenalis gene models (assemblage A genome, WB strain, release 3.1; GiardiaDB.org; [27][34]), using RSEM [35]. Transcripts-per-million (TPM) for each gene were averaged across replicates from each growth phase, and used to rank genes according to relative transcriptional abundance. Expected counts for each gene were submitted to EBSeq [36], incorporating median normalization, and DTGs were identified using a false discovery rate (FDR) of <0.05. Fold-change in transcription between growth phases was calculated using the normalized mean expected counts output from EBSeq. As a further filter, only those genes with at least 10 mean expected counts in at least one growth phase were included in further analysis.
Feature detection was calculated as a function of mapped read depth, using the counts module in QualiMap (v1.0) with the–k 10 flag (denoting a minimum mapped read threshold of 10) [37]. Saturation plots were displayed in Excel (Microsoft). Heat maps were generated in R (v3.0.2) using the heatmap module. KEGG BRITE terms associated with peptides in the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (release 69.0), were transferred to the closest homolog in G. duodenalis using BLASTp (lower expect threshold of 10−5; [38]). Gene ontology (GO) terms for the predicted G. duodenalis proteome were retrieved from GiardiaDB.org; and sensitive structure-based homology searches were performed for DTGs annotated as ‘hypothetical’ or ‘deprecated,’ and for peptides encoded by highly transcribed (top 100) genes, using I-TASSER software (v3.0; [30,39]). Briefly, I-TASSER generates putative three-dimensional structural models from amino acid sequences, incorporating predicted secondary structure and the consensus model from multiple threading programs, followed by iterative molecular dynamics simulation to minimize free energy [30]. The closest structural homolog available in the Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB; rcsb.org), and consensus GO terms associated with the ten best structural matches, are then inferred for the putative model. Bar charts and box plots describing transcriptional abundance were generated using Excel (Microsoft) and Prism software (GraphPad). For bar charts, mean normalized expected counts are plotted with standard error derived from pre-normalised expected counts for biological quadruplicates. Putative transcription factor (TF)-binding motifs in the promoter regions of DTGs were identified within 400 bp upstream of the start codon in non-overlapping (i.e., non-coding) regions using DREME [40]; cf [41]. Promoters from a DTG group of interest were interrogated using promoters from another DTG group as the background. Interacting TFs for homologous motifs in yeast (all available databases) were predicted using TOMTOM [42]. For each motif, the density of the 5’ nucleotide position (both forward, and reverse complement) was calculated as a function of promoter length, and displayed together with a histogram of the corresponding promoter lengths using R.
The coefficient of variation (SD ÷ average; denoting variation in transcript abundance between biological replicates) was calculated for all transcribed genes encoding variant-specific surface proteins (VSPs). Pearson correlation was used to compare transcriptional abundance (FPKM and RPKM) values from independently generated transcriptomic data sets. Fisher exact tests were used to determine significantly over-represented (i.e., enriched) KEGG BRITE terms within DTG groups. GO enrichment analysis was performed using the BinGO module in Cytoscape [43,44] by firstly providing a background of GO terms specific to G. duodenalis, incorporating all GO terms from GiardiaDB.org and those terms from I-TASSER with confidence scores ≥0.3. The enriched Biological Process GO terms within DTG groups were then identified using Fisher exact testing (FDR < 0.05). To further minimize false positives in this analysis, resultant GO terms with fewer than ten associated genes in the background, were discarded. Fisher permutation tests were also used to compare the average transcription level of genes of interest between growth phases.
Open reading frames for genes in this article, quoted according to GiardiaDB.org: GL50803_87577; GL50803_7195; GL50803_10403; GL50803_33769; GL50803_27266; GL50803_23756; GL50803_16568; GL50803_27266.
We compared the transcriptomes of the attached population of G. duodenalis trophozoites harvested at 48, 60 and 96 h during axenic culture. These time points corresponded to log phase (~ 28 x 104 attached cells/cm2), stationary (confluent monolayer on the flask wall; 72 x 104 cells/cm2), and declining phases (detaching; 42 x 104 cells/cm2) respectively (Figs 1A and S1). In total, 473.76 million RNA-seq reads were produced. For each of 12 samples, an average of 25.25 million high-quality reads (i.e. 92% of all filtered reads) were mapped to the G. duodenalis gene models (WB strain [27,34]). The average coverage depth was 405-times (S1 Table), and novel transcript detection as a function of the number of mapped reads was saturated for all samples (S2 Fig). After filtering, transcripts were detected from 7,637 ORFs (78.9% of all predicted ORFs), of which 2,347 had functional annotations (98.8% of all annotated), 3,208 were hypothetical (90.5% of all hypothetical), and 2,082 were deprecated (55.3% of all deprecated). When the relative transcriptional abundance (TPM) for genes at each growth phase was compared, values ranged across four orders of magnitude (S3 Fig). The five most highly transcribed genes across the three phases were a translation elongation factor (EF1-beta), ornithine carbamoyltransferase, glyceraldehyde-3-phosphate dehydrogenase, and two ribosomal proteins (S15A and P2C). Among the top 100 most highly transcribed genes in each phase were arginine deiminase, carbamate kinase, enolase, variant-specific surface proteins (VSPs), peroxiredoxin-1ai, two putative thioredoxins and two protein disulfide isomerases (S2 Table). The normalized mean expected counts for genes at each growth phase were displayed in a heat map, and a progressive trend of up- or down-regulation was identified for the majority of the detected genes (Fig 1B). During the axenic growth and decline of G. duodenalis trophozoites over 96 hours, 754 genes were differentially transcribed at statistical significance (S3 Table); including 438 genes (47% hypothetical; 6% deprecated) down-regulated at the declining phase relative to the log and stationary phases; and 281 genes (49% hypothetical; 19% deprecated) up-regulated at the declining phase relative to both of the earlier phases. On average, fewer than ten genes were differentially transcribed in other comparisons (Fig 1C), and thus statistical analyses were restricted to the two large DTG groups, hereafter termed ‘down-regulated’ and ‘up-regulated’ in the declining phase.
The median fold change for differentially transcribed down- and up-regulated gene groups was 1.85 (IQR = 0.77) and 2.12 (IQR = 0.8) respectively. Using I-TASSER [30], putative structures could be generated for 224 of 234 (95.2%) hypothetical and deprecated proteins within the down-regulated group, and 141 of 190 (74.2%) such proteins in the up-regulated group. The lower availability of structural models for the up-regulated group is due to the presence of 31 genes (16.3%) encoding peptides of >1500 amino acids, which are too large for analysis in the I-TASSER software [30].
GO enrichment analysis revealed significant over-representation of 113 ‘Biological Process’ terms in the down-regulated gene group, of which 49 were unique to this group, including ‘energy derivation by oxidation of organic compounds,’ ‘signaling,’ and ‘locomotion’ (S3 Table). Enriched KEGG BRITE annotations included the NEK kinase family, threonine peptidases, ubiquitin conjugating enzymes (E2) and the T1 proteasome family (S4 Table). The down-regulated genes annotated with oxidoreductase activity (GO:0055114 and/or GO:0016491) included glutamate synthase, 6-phosphogluconate dehydrogenase, alcohol dehydrogenase and nitroreductase-1 as well as a number of hypothetical proteins with predicted structural similarity to glutamate synthase, hydroxylamine oxidoreductase, thiol-cycling enzymes (protein disulfide isomerase and thiol:disulfide protein dsbA), a ferretin-like Dps-like peroxide resistance protein and a nickel-binding superoxide dismutase (Table 1). Further investigation of down-regulated genes associated with ubiquitin-conjugating and protease activity, revealed four ubiquitinylating enzymes (two paralogs of a 28.4 kDa E2, a 17 kDa E2 and the ubiquitin ligase UBC3), and four beta-subunits of the 20S proteasome (S4 Fig).
We identified an annotated and a hypothetical glutamate synthase in the down-regulated gene group. Interrogation of putative structures for each protein revealed that the hypothetical glutamate synthase (GL50803_87577) was most structurally similar to a bacterial glutamate synthase beta sub-unit (PDB code: 2VDC), whereas surprisingly, the annotated glutamate synthase (GL50803_7195) was most similar in structure to trimethylamine (TMA) dehydrogenase from Methylophilus methylotrophus (PDB code: 1DJQ). Putative homologs of a TMA sensor protein, a Rap modulator protein (C-terminal domain only), and a phosphotransfer protein were also present in the down-regulated gene group. The former two are present as single-copy orthologs in other G. duodenalis assemblages (B and E; GiardiaDB.org). These findings suggest the presence of a two-component-like signaling system in G. duodenalis. Histidine kinases are integral to two-component systems, and further interrogation of putative structures for hypothetical proteins revealed putative histidine kinase activity for GL50803_10403 (Fig 2).
The up-regulated gene group was enriched for 107 ‘Biological Process’ GO terms (43 unique), including ‘gene silencing,’ ‘mitosis,’ and ‘microtubule-based process’ (S3 Table). Enriched KEGG BRITE annotations in this group related to carbon fixation, oxidoreductases and the cytoskeleton (S4 Table). Further investigation of oxidoreductase-related genes in the up-regulated group revealed two pyruvate:ferredoxin oxidoreductase (PFOR) paralogs, and a hypothetical protein with structural homology to a GntR transcription factor (S2 Table).
To contextualize the oxidoreductase-related genes that were differentially transcribed between growth phases, we investigated the transcription of other genes involved in the antioxidant system and glycolysis. Progressive but non-significant decreases in transcriptional abundance were identified for thioredoxin reductase, three putative thioredoxins, three peroxidredoxins and five thioredoxin domain-containing protein disulfide isomerases. Similar patterns were observed for transcripts encoding the oxygen-consuming enzymes NADH oxidase (GL50803_33769; [45]) and flavodiiron protein [46,47] (Fig 3A). The collective transcription of annotated antioxidant enzymes was significantly lower in the declining phase compared to earlier phases (Fisher permutation test, 1000 iterations, p = 0.037; Fig 3A).
The presence of genes encoding glycolytic oxidoreductases in both down- and up-regulated gene groups suggested changes in central carbon metabolism during in vitro growth. A hexose transporter, and the non-oxidative glycolytic enzyme glucose-6-phosphate isomerase, were also down-regulated over time. Indeed, transcription of the majority of genes representing glycolytic enzymes in both the pentose phosphate and Embden-Meyerhoff pathways, was lowest in the declining phase (Fig 4). Conversely, transcription of a number of glycolytic enzymes downstream of pyruvate increased over time. In addition to significant up-regulation of PFOR-coding genes in the declining phase, genes encoding one of three ferredoxins (GL50803_27266) and a glutamate dehydrogenase also showed greatest transcriptional abundance in this phase (Fig 4).
Given the robust transcriptional changes in the antioxidant system, we compared the transcriptome for each growth phase with independently generated transcriptomes for WB strain trophozoites cultured under aerobic and anaerobic conditions, as reported by Ma’ayeh et al [28]. When all transcribed genes were considered, our data correlated most closely with the anaerobic transcriptional profile. The transcriptional abundance of annotated antioxidant genes at log and stationary phase, however, correlated best with the aerobic transcriptome, and in the declining phase we observed a shift to stronger correlation with the anaerobic transcriptome (S5 Fig, panels A & B). When glycolytic genes were investigated, there was a trend of increasing correlation with the anaerobic profile over time, and the correlation with the aerobic profile declined after stationary phase. Further dissection of this result revealed separate underlying trends, wherein the genes upstream of pyruvate diverged markedly from the aerobic profile after stationary phase, but little change was seen in genes downstream of pyruvate (S5 Fig, panels C-E).
Eleven genes encoding VSPs were identified in the up-regulated group as opposed to only two such genes in the down-regulated group. Further investigation of the transcription levels of 193 transcribed VSPs in our data revealed seven prominently transcribed genes, whereas the rest of the population was relatively lowly transcribed (S6 Fig). Aggregate VSP transcription increased over time, driven by increases in the seven most highly transcribed genes (Fig 3B and 3C), which were consistently highly transcribed in all replicates (S7 Fig). Interestingly, a marked decrease in inter-experimental variation was evident for highly transcribed VSPs in the declining phase relative to earlier phases (S8 Fig). These results prompted investigation of the other major class of membrane proteins in G. duodenalis, the 60 high-cysteine membrane proteins (HCMPs), whose aggregate transcription in contrast to the VSPs, decreased progressively over time, driven by declining abundance of the most highly transcribed gene quartile (Fig 3B).
We hypothesize that DNA-binding transcription factors (TFs) may mediate the dynamic variation in gene transcription evident in G. duodenalis during axenic growth. The down-regulated gene group did not contain annotated TFs, but did include a putative homolog of the Neisseria gonorrhoeae MtrR repressor among three hypothetical protein-coding genes with GO annotations relating to transcriptional regulation (GO:0001071; Fig 2 and S2 Table). The up-regulated group contained an E2F-like TF (GL50803_23756; [48]), a putative TF (GL50803_16568) and four hypothetical proteins annotated with GO:0001071 including the aforementioned GntR homolog (S2 Table). In order to identify putative TF-binding motifs within the promoter regions of genes in the two large DTG groups, we used a similar method to Xu et al. [41] for analysis of the related diplomonad Spironucleus salmonicida. Totals of 283 and 178 non-overlapping promoter regions of 8 bp to 400 bp were available for the down- and up-regulated gene groups, respectively. The motif AWTTW was significantly over-represented in the promoters of down-regulated genes relative to promoters of up-regulated genes, and the motif GRGGTM was over-represented in the up-regulated gene promoters in the same way (Table 2). After correction for multiple comparisons, no significant matches to known TF-binding motifs in yeast were found for either motif using TOMTOM. An analysis of the position of each motif within promoter regions revealed robust positional conservation for the AWTTW motif within 100 bp of the start codon, but not for GRGGTM. There was no evidence to suggest that these results are biased by the proportion of non-overlapping genes available for analysis in each group (76% and 72%, respectively); moreover, the motif positional densities remain when only the (artificially truncated) 400 bp promoter regions are considered, indicating little effect of promoter length on motif location (Fig 5).
We studied the transcriptional dynamics of axenic G. duodenalis trophozoites over the log, stationary, and declining phases (48, 60 and 96 h) of in vitro growth. We also generated putative structures for the majority of hypothetical and deprecated proteins encoded by DTGs, gaining insights into the function of divergent gene products that lie beyond reach of conventional sequence-based annotation tools. The majority of genes were transcribed in at least one of the three growth phases investigated, and clear transcriptional variation was evident, implicating effective and directed mechanisms of transcriptional regulation. Strikingly, and contrary to our original hypothesis, trophozoites in non-sparged axenic culture exhibited little difference in transcription between log and stationary phases. Indeed, at both phases, we observed high transcription of a number of genes involved in managing oxidative stress. It should be noted that in order to limit the influence of dead or dying cells on our transcriptomic data, we confined our analyses to adherent cells, and discarded suspended cells prior to sequencing. The suspended population would likely be enriched for both dividing cells and non-viable/dying cells, and it would be of interest to analyze the transcriptomes of these different populations if they could be fractionated. As trophozoites consume oxygen [49,50] and are cultured in a closed system, the pronounced shift in oxidoreductase transcription between log and stationary phases, and the declining phase of axenic growth, is consistent with declining oxygen tension in TYI-S33 medium, and might be associated with a concomitant decline in glucose catabolism and/or entry into a metabolically quiescent state. The present work also found informatic evidence for a two-component–like signaling system that might mediate cytosolic redox or metabolite sensing, and identified different promoter motifs within DTG groups which, together, could contribute to transcriptional regulation.
G. duodenalis features an elaborate antioxidant system that utilizes NAD(P)H to reduce intracellular oxygen and associated reactive oxygen species (ROS). This system both protects iron-containing enzymes, such as PFORs, from oxidative inactivation [51,52] and allows maximal ATP production from glycolysis [53]. As a ‘microaerophile’, G. duodenalis thrives under dissolved oxygen (dO2) concentrations between 5–25 μM, above which dO2 becomes cytotoxic [54]. In accordance with the standard trophozoite maintenance protocol [32], we did not sparge dO2 from the TYI-S33 medium for this experiment, and medium is expected to contain dO2, particularly during the log phase of trophozoite growth. The abundant transcription of genes encoding peroxiredoxin, oxygen-consuming NADH oxidases, thioredoxins and other protein disulfide isomerases is consistent with previous reports [27–29,55]. However, the dynamics of antioxidant transcription in G. duodenalis under standard culture conditions has not been studied previously. It is likely that the antioxidant system is constitutively highly transcribed to manage transient increases in gut dO2 [28,56]. Nevertheless, against this background, we observed down-regulation of the vast majority of antioxidant-coding genes over time, suggesting global regulation of antioxidant transcription.
A number of our results indicate that G. duodenalis trophozoites may be under oxidative stress at early stages of in vitro culture. Firstly, although it was not possible to directly measure dO2 without perturbing the standard culture system, correlations with independently generated transcriptional profiles for trophozoites under aerobic and anaerobic conditions [28] can be considered as a proxy measure of dO2 tension. At successive growth phases, antioxidant and glycolytic gene transcription shifted from resembling the aerobic transcriptional profile, to the anaerobic profile. The correlation was influenced more strongly by antioxidant transcription than glycolytic enzyme transcription, however glycolytic genes upstream of pyruvate appeared most sensitive to changes in oxygen tension. This strongly suggests that oxygen tension decreases progressively in axenic culture. Furthermore, genes that were down-regulated in our data at the declining phase, encode putative antioxidant proteins, such as an iron-independent superoxide dismutase, and a ferritin-like (iron-sequestering) protein. TYI-S33 medium is supplemented with ammonium ferric citrate, as iron is essential for the activity of enzymes such as PFOR. However iron can also react with dO2 to generate ROS. Thus the greater transcriptional abundance of putative iron-sequestering and iron-independent antioxidant proteins at early growth phases, may represent a response to iron- and oxygen-induced oxidative stress. In addition, PFOR enzymes are sensitive to oxidative inactivation [51,52,57], and both PFOR paralogs showed a significant increase in transcription in the declining phase, which is consistent with lower oxygen tension. Lastly, the greater abundance of transcripts encoding elements of the ubiquitin system and proteasomal components at earlier phases, may reflect heightened turn-over of oxidized proteins [58]. However as the proteasome is also important for the demands of protein folding in actively dividing cells, this result requires further investigation.
In the context of global changes in antioxidant transcription, and a suggested decline in dO2, we observed inverse transcriptional patterns for the high-cysteine membrane proteins (HCMPs), and variant-specific surface proteins (VSPs), which may indicate differential sensitivity to dO2. HCMPs contain disulfide motifs that are common in antioxidant proteins and can both oxidize and reduce substrates such as misfolded proteins, and reduce ROS [59]. HCMPs localize to the nuclear envelope, endoplasmic reticulum, and possibly to the trophozoite plasma membrane [24,59]. The greater aggregate transcription of HCMPs early during in vitro growth might relate to protecting membranes from peroxidation [59]. Conversely, aggregate VSP transcription increased over time, which was largely due to progressively greater transcription of seven highly abundant VSP-coding genes, rather than induction of new genes. G. duodenalis trophozoites transcribe multiple VSP-coding genes, all but one of which are degraded by RNA-interference [17], and thus only a single VSP gene product is displayed on the trophozoite membrane at any time. Our results support previous reports that WB trophozoite populations express relatively few VSPs [60], which may be due to slower VSP turnover in this strain [61]. Specific VSPs are proposed to modulate trophozoite sensitivity to host defenses such as intestinal proteases [62], and VSP suppression is associated with nitroimidazole resistance [19]. Given that the variation in VSP transcription between replicates is particularly low for highly transcribed VSPs in the declining phase, it would be interesting to test whether VSPs are under selection in TYI-S33 medium, or whether the transcriptional increase merely reflects, for example, lower competition from HCMPs for membrane occupancy.
The functionally reduced mitochondria, or mitosomes, of G. duodenalis do not contain enzymes for the tricarboxylic acid cycle or oxidative phosphorylation [34], and this protist is dependent on glycolysis and fermentative metabolism to generate energy from glucose. ATP is generated by direct phosphorylation of AMP and ADP [63]. Electrons liberated during glycolysis are accepted by NAD and NADP, forming NAD(P)H, which must be re-oxidized. Under anaerobic conditions, pyruvate is diverted to ethanol to regenerate NAD, whereas NADP is regenerated through a ‘shunt’ incorporating glutamate dehydrogenase and alanine aminotransferase. Conversely, in the presence of dO2, oxidoreductases in the antioxidant system consume NAD(P)H to neutralize dO2, ROS, and oxidized biomolecules. Under these conditions, pyruvate is not required for NAD(P) regeneration, and can be further oxidized to acetate. This ‘micro-aerobic’ metabolism maximizes the ATP that is generated from glycolysis.
Here, we observed a trend of decreasing transcriptional abundance for the majority of genes encoding enzymes in the Embden-Meyerhoff and pentose phosphate glycolytic pathways, that convert glucose to pyruvate (Fig 4). TYI-S33 medium contains very high glucose concentrations (55 mM) [32], and thus it is highly unlikely that glucose is exhausted at the declining phase. Instead, the apparent down-regulation of glycolytic pathways may be due to declining dO2 availability, which could limit the efficiency of glycolysis. If glycolysis is progressively down-regulated as dO2 declines, G. duodenalis may rely on alternative energy sources. This protist is capable of converting arginine, aspartate and alanine to pyruvate [64,65]. The arginine dihydrolase pathway provides a ready source of ATP in G. duodenalis and is highly transcribed and stable across all time points studied here. In contrast to glycolysis, which must be coupled to NAD(P)+ regeneration mechanisms, the catabolism of aspartate to pyruvate is effectively redox-neutral in that it does not generate excess NAD(P)H [65]. We observed progressive, but non-significant increases in transcription of two genes involved in aspartate catabolism (Fig 4), which may reflect greater reliance on aspartate for energy as dO2 declines. Significant up-regulation of PFORs at the declining phase, and the concomitant down-regulation of alcohol dehydrogenase-coding genes, suggests that a substantial amount of pyruvate is converted to acetate for ATP production even under limited dO2 availability (Fig 4). This is supported by reports of acetate production under complete anaerobiasis in G. duodenalis [64], indicating that pyruvate flux through acetyl-CoA synthase is a resilient mechanism of ATP generation.
The identification of a putative trimethylamine (TMA) dehydrogenase and elements of a putative TMA-NO sensing system (discussed below) in the down-regulated genes, raise the possibility that G. duodenalis can utilize TMA-NO as a terminal electron acceptor. TMA-NO metabolism has been demonstrated for a variety of gut commensals [66], and the acquisition of key metabolic genes in G. duodenalis via horizontal gene transfer is well documented [34, 67, 68]. Lastly, the significant up-regulation of glycerol kinase at the stationary and declining phases relative to log phase, could indicate a glycerol-dependent ATP generation pathway in G. duodenalis. The metabolically similar protist Entamoeba histolytica has been shown to divert glycolytic intermediates to glycerol when central glycolytic enzymes are experimentally inactivated, and glycerol kinase is suggested to function in ATP generation [69]. Thus, it would be interesting to investigate the potential of this ‘glycerol shunt’ as an alternative source of ATP in G. duodenalis.
The major shift in transcription between the log and stationary, and the declining growth phases, is likely to involve signaling between cytosolic proteins and nuclear transcription factors. Intriguingly, the observed transcriptional down-regulation of genes related to glycolysis might be mediated by redox-dependent TFs. A putative GntR homolog is up-regulated at the declining phase. In Corynebacterium glutamicum, GntR is reported to repress transcription of the gene encoding 6-phosphogluconate dehydrogenase, which is associated with the production of NADPH in the pentose-phosphate pathway. Notably, the gene encoding 6-phosphogluconate dehydrogenase is transcriptionally down-regulated at the declining phase in our data-set, and as mentioned above, transcription of glycolytic genes upstream of pyruvate, such as 6-PGDH, seem to vary more greatly in response to changes in dO2 tension (S5 Fig, panel D). Taken together, these findings suggest that declining dO2 may inhibit glycolysis and lead to the accumulation of intracellular NAD(P)H. The apparent metabolic shift away from glycolysis may be mediated, at least in part, by redox-sensitive TFs such as the GntR homolog.
Conserved kinases have been classified in G. duodenalis, many of which localize to the cytoskeleton and might participate in regulation of cell structure or motility. Other kinases, particularly within the massively expanded NEK kinase family, are predicted to lack catalytic activity [70]. Although complete signaling pathways have not been resolved in G. duodenalis, here we have identified several differentially transcribed structural homologs of proteins that participate in two-component signal transduction. Two-component systems feature histidine kinases and associated sensor domains that detect changes in cytoplasmic or extracellular environmental conditions. Conformational changes in the sensor domain induce histidine autophosphorylation proximal to the kinase domain, and the charged phosphate is subsequently transferred to an effector protein either directly, or via phosphotransferase intermediates. In many cases, the effector translocates to the nucleus to elicit a transcriptional response [71]. Two-component systems are prevalent in bacteria, fungi, plants and free-living protozoa [72], but are little documented in parasitic protozoa and reportedly absent from Plasmodium falciparum [73–75]. The putative sensor protein in our data is structurally similar to a transmembrane protein that detects trimethylamine (TMA) in the bacterial periplasm. The predicted G. duodenalis structure is truncated however, which may indicate a role in intra-membrane or cytosolic chemotaxis. We also identified a histidine kinase homolog, which could represent a kinase class that was previously thought to be absent in G. duodenalis [70]. Furthermore, the down-regulated gene group included a gene encoding a hypothetical protein with putative structural homology to the MtrR repressor, a TF that is regulated by two-component signaling in Neisseria gonorrheae [76] (Fig 2).
Although we did not identify a dimerization domain in the putative histidine kinase or a candidate for the response regulator, the sensor domain, Rap modulator and MtrR homologs are conserved between assemblages B and E of G. duodenalis, emphasizing the functional importance of these gene products among members of the species complex (Fig 4). Given the absence of two-component systems from metazoans [74], this putative pathway warrants detailed investigation as a possible target for chemotherapeutic intervention.
In further support of coordinated transcriptional regulation, we performed the most comprehensive and sensitive promoter motif search on this species to date, and identified motifs that are enriched in the promoter regions of down- and up-regulated DTG groups. The motif enriched in the down-regulated group (AWTTTW), occurs close to the translation start codon and is likely to be related to the AT-rich transcription initiator motifs reported in previous studies [26,77,78]. At present, little is known about TFs in G. duodenalis other than those involved in inducing encystation [48,79–82]. It is conceivable that, in the presence of abundant nutrients, transcription is relatively tightly regulated via TF binding to initiator regions. Subsequently, in the declining phase when trophozoites appear to be metabolically stressed, and display physiological changes such as a loss of cytoadherence, less energy might be expended on transcriptional regulation. The GRGGTM motif, which is enriched in the promoters of up-regulated genes, exhibits little positional conservation, perhaps due to the relatively low number of positive promoter sequences (64 vs 209 for the AWTTTW motif). Although neither the of the motifs identified in down- and up-regulated genes matched to known TF-binding motifs in yeast, the MtrR and GntR putative TFs identified in these groups are worthy candidates for further investigation, which could link redox/metabolite sensing and transcriptional regulation.
By characterizing the dynamics of the mRNA transcriptome of axenically cultured G. duodenalis trophozoites across growth phases, we identified major changes in gene transcription that relate to central carbon metabolism and the antioxidant system. We also identified a putative signaling pathway and promoter motifs upstream of DTGs that might contribute to transcriptional regulation. We show that transcriptional behaviour of G. duodenalis trophozoites differs markedly over time in axenic culture, likely reflecting the exploitation and depletion of essential nutrients in a closed culture system. Importantly, the present data indicate that culturing G. duodenalis under micro-aerobic or entirely anaerobic conditions (through sparging medium) is likely to have a significant impact on transcriptional behaviour, particularly in relation to the oxidative stress response. Therefore, we suggest that the modularity of antioxidant transcription in this protist be further tested under precisely defined atmospheres, together with metabolomic profiling. The consistently high transcription of these pathways during log and stationary phases—when trophozoites are typically harvested for experimentation—is also relevant for contextualizing single time-point studies in these parasites.
Consistent with previous findings that G. duodenalis exhibits specific stress responses and considerable transcriptional flexibility, the present study indicates that major shifts in transcription in G. duodenalis might be regulated at least in part through key transcription factors (i.e., MtrR, GntR), and a number of distinct motifs consistent with TF-binding sites, that are enriched in the promoter regions of down- and up-regulated DTGs respectively. The possible role of a two-component-like signal transduction system is particularly interesting and could link cytosolic redox or metabolite sensing and transcriptional changes. This work has been greatly enhanced by the large-scale prediction of putative structures and structural homologs for hypothetical and deprecated proteins, among which were novel antioxidant and signaling proteins among many others. This approach has great potential for illuminating the functions of vast numbers of under-annotated and un-annotated gene products in other important pathogens. The present study also provides a starting point for re-examination of the constituents of the standard trophozoite culture medium, and a reference against which future studies of targeted alterations could be compared. For example, the reaction of dissolved oxygen (dO2) with iron may be a source of oxidative stress at early phases during in vitro culture of G. duodenalis trophozoites, and thus concentrations of ammonium ferric citrate and ascorbic acid might need to be reconsidered, particularly as work in other systems has linked ascorbic acid with intracellular iron concentrations [83]. A major finding is that glucose is likely to be in excess in the standard culture media, but its utilization as a carbon source might be limited by the availability of dO2, inducing trophozoites to rely on less efficient energy generation pathways in the declining phase. In the future these findings and the detailed longitudinal transcriptomic information presented here, should be used in conjunction with targeted metabolomic investigations in G. duodenalis, with the aim of creating a completely defined medium. Whereas the present study has used the G. duodenalis assemblage A genome strain (WB), single time-point profiling of assemblages B [84] and E [85] indicate different transcriptional patterns, which could result from different genomic organization [27,60]. Therefore, similar longitudinal transcriptional investigations of other assemblages are required to better characterize the degree of transcriptional flexibility and the drivers and mediators of transcriptional responses across the species. This work should facilitate a better understanding of the transcriptional flexibility and metabolic preferences of G. duodenalis under standard culture conditions, and contribute to the development of a completely defined medium for more refined investigations into the biology of this important model eukaryote and pathogen.
|
10.1371/journal.pgen.1005617 | DCA1 Acts as a Transcriptional Co-activator of DST and Contributes to Drought and Salt Tolerance in Rice | Natural disasters, including drought and salt stress, seriously threaten food security. In previous work we cloned a key zinc finger transcription factor gene, Drought and Salt Tolerance (DST), a negative regulator of drought and salt tolerance that controls stomatal aperture in rice. However, the exact mechanism by which DST regulates the expression of target genes remains unknown. In the present study, we demonstrated that DST Co-activator 1 (DCA1), a previously unknown CHY zinc finger protein, acts as an interacting co-activator of DST. DST was found to physically interact with itself and to form a heterologous tetramer with DCA1. This transcriptional complex appears to regulate the expression of peroxidase 24 precursor (Prx 24), a gene encoding an H2O2 scavenger that is more highly expressed in guard cells. Downregulation of DCA1 significantly enhanced drought and salt tolerance in rice, and overexpression of DCA1 increased sensitivity to stress treatment. These phenotypes were mainly influenced by DCA1 and negatively regulated stomatal closure through the direct modulation of genes associated with H2O2 homeostasis. Our findings establish a framework for plant drought and salt stress tolerance through the DCA1-DST-Prx24 pathway. Moreover, due to the evolutionary and functional conservation of DCA1 and DST in plants, engineering of this pathway has the potential to improve tolerance to abiotic stress in other important crop species.
| Drought and salt are two of the most serious threats to food production worldwide, and research on stress tolerance in crops is important for future food security. In this study we identified DCA1, a transcriptional co-activator of DST that is conserved in the world’s three major crops. DCA1 participates in stress tolerance by controlling stomatal aperture through modulation of H2O2 homeostasis in guard cells. This finding not only increases our understanding of the molecular mechanism by which plants withstand harsh environmental conditions, but it may also facilitate future molecular breeding and genetic engineering of drought- and salt-tolerant crops.
| How to feed a growing population that is expected to reach roughly 9 billion by the middle of this century is among the major challenges of our time [1]. Modern agriculture has greatly improved food production [2], but progress towards avoiding the negative effects of climate change and diminishing soil conditions has been insufficient. Most worryingly, many of the plants upon which we depend for food production are particularly sensitive to environmental stress [3]. Droughts are likely to be more frequent as global warming accelerates, and rising sea levels will result in the loss of productive agricultural land to water infiltration and increased soil salinity. Together these unfavorable factors pose a huge threat to food security, and studying drought and salt tolerance in crops is becoming increasingly urgent.
Many previous studies on drought and salt tolerance in plants have mainly focused on the model species Arabidopsis thaliana. Years of effort have revealed that the responses of plants to water stress are controlled by complex regulatory signalling events mediated by Ca2+ [4, 5], abscisic acid (ABA) [6–8], reactive oxygen species (ROS) [9, 10], ion transport [4, 11–13], and the activities of transcription factors (TFs) [14]. Stomata are often described as the guardians of leaves which are of course the photosynthetic organs in which food manufacturing occurs. The main function of stomata, which are formed from pairs of guard cells, is to allow gases to move rapidly into and out of the leaf. However, this evolved trait poses a problem for plants since they face the predicament of taking up CO2 through stomata while attempting to minimize water loss through these pores. The ability to effectively control the balance between photosynthesis and transpiration in accordance with the external environment is an impressive evolutionary achievement.
ROS such as H2O2 are a product of the incomplete reduction of molecular oxygen, and ROS production is widely considered a symptom of cellular dysfunction. ROS participate in cell death and may also function as signalling molecules. Increasing evidence suggests that ROS function as second messengers in stomatal aperture control [9, 15]. In 1996, McAinsh et al. first reported that ROS induced stomatal closure and inhibited stomatal opening by increasing the concentration of cytosolic free calcium ([Ca2+]cyt) in Commelina communis [16]. A detailed follow-up study reported that ABA-stimulated ROS accumulation induced stomatal closure via the activation of plasma membrane calcium channels in Arabidopsis [9].
Since plants are sessile organisms, they must have highly evolved sophisticated mechanisms to detect and respond to environmental perturbations. Changes in the expression of stress-related genes are an important part of the plant response to environmental stress. Numerous transcription factors including APETALA 2/ethylene-responsive element binding factor (AP2/ERF), dehydration responsive element binding protein (DREB)/C-repeat-binding factor (CBF), ABA-responsive element binding protein (AREB)/ABA-responsive element-binding factor (ABF), No apical meristem, Arabidopsis transcription activation factor and Cup-shaped cotyledon (NAC) are associated with plant abiotic stress responses [14]. Regulation of target genes by TFs is a highly complex and delicately balanced process. In most cases TFs do not function alone but recruit partner proteins (cofactors) to form transcription initiation complexes [17]. Cofactors are transcription factor interacting proteins that either activate or repress the transcription of target genes, and numerous examples have been reported in animals including humans, but few have been identified in plants. Arabidopsis HAIRY MERISTEM (HAM) family proteins were recently found to act as conserved interacting cofactors with the transcription factor WUSCHEL (WUS) to drive downstream transcriptional programs that help promote shoot stem cell proliferation [18]. Another study demonstrated that HYPOXIA RESPONSE ATTENUATOR1 (HRA1) interacts with ethylene-responsive factor group VII transcription factor (ERF-VII TF) RAP2.12 to negatively modulate its activity under hypoxia [19].
Efforts have also been made to determine the underlying physiological, genetic and molecular mechanisms mediating drought and salt tolerance in crops such as rice. Stress-responsive NAC 1 (SNAC1) is a member of the rice NAC TF family that is specifically induced in guard cells in drought conditions, and overexpression of SNAC1 in rice significantly enhanced drought tolerance [20]. We previously isolated the C2H2 zinc finger transcription factor DST that negatively regulates stress tolerance in rice [21]. DST regulates stomatal aperture by modulating the expression of genes related to ROS homeostasis. However, these studies are fragmentary. The exact mechanisms by which these TFs regulate the expression of target genes remain unknown. In the present study, we identified the CHY zinc finger protein DCA1, an interacting partner of DST. Homologs of DCA1 in rice and Arabidopsis were recently shown to increase stomatal opening and were upregulated by heat stress [22]. However the exact molecular function of this protein, the pathways involved and the phenotypes of plants in which DCA1 is modified remain unknown. In this research, we revealed that DCA1 forms a heterologous tetramer with DST and positively regulates DST activity. This transcriptional complex regulates the expression of Prx24, an H2O2 scavenger preferentially expressed in guard cells. Together, the DCA1-DST-Prx24 pathway contributes to stomatal movement via regulating ROS homeostasis under stress conditions. These findings may facilitate the engineering of crops with improved drought and salt tolerance.
Most TFs do not function alone. Rather, they recruit intermediary proteins (cofactors) to initiate transcription effectively. In humans, several hundred putative co-regulators have been identified [17]. To identify cofactors of DST, we performed yeast two-hybrid screening with a ZH11 (Oryza sativa L. japonica. cv. ZhongHua 11) leaf cDNA library. A version of DST with the self-activation domain removed (amino acids 1–72 removed from the N-terminus, including the zinc finger domain, ΔN) was used as bait. Dozens of potential targeting proteins were found including a U-box domain containing protein, Elongation Factor 1-alpha, a protein similar to Calcium-dependent protein kinase-related kinase, and the CHY zinc finger protein DCA1. To further verify the interaction between these proteins, we used DCA1 and DST as bait and prey, respectively (Fig 1A), and their interaction was confirmed through additional yeast two-hybrid assays. DCA1 was able to interact with DST independently of its C2H2 zinc finger domain which likely functions in DNA binding (Fig 1B). The DST-DCA1 interaction was also confirmed via bimolecular fluorescence complementation (BiFC) assays in Arabidopsis protoplasts in vivo, in which DST was fused to the amino-termini half of yellow fluorescent protein (YFP) (DST-nYFP) and DCA1 was fused to the carboxy-termini half of YFP (DCA1-cYFP). The YFP signal was mainly restricted to nuclei when DST-nYFP was cotransformed with DCA1-cYFP (Fig 1C). Consistent with these results, an in vitro pull-down assay using recombinant DST fused to maltose binding protein (DST-MBP) and His2-tagged DCA1 (DCA1-His2) that were expressed in E. coli showed that DST-MBP but not MBP was able to pull down DCA1-His2 (Fig 1D). These results confirm the direct interaction between DCA1 and DST both in vitro and in vivo.
Careful analysis of the structure of DCA1 identified a CHY zinc finger domain and a C3H2C3 or RING-H2 domain (S1A Fig). A database search for homologous proteins identified conserved CHY zinc finger domains in various different species, but the C3H2C3 domain appeared not to be conserved (S1B Fig). To characterize DCA1, we investigated the subcellular localization of DST-YFP and DCA1-YFP fusion proteins in Arabidopsis protoplasts. YFP fluorescence was apparent in both the cytoplasm and nuclei of control protoplasts transformed with the YFP-vector. In contrast YFP fluorescence was mainly restricted to nuclei in DST-YFP and DCA1-YFP transformants (Fig 2A–2C). Investigation of the tissue-specific expression of DCA1 by qRT-PCR revealed a similar expression pattern to that of DST, with both genes expressed at relatively high levels in leaves and culms (Fig 2D and S2A Fig). We also tested whether the expression of DCA1 was regulated by drought and salt stress using qRT-PCR and found that DCA1 is upregulated rapidly after 1–3 h of drought or salt treatment, with expression peaking at 12 h and normal expression resumed 24 h after treatment (Fig 2E and 2F).
To dissect the physiological functions of DCA1 in plants, we generated transgenic rice (ZH11) overexpressing DCA1 under the control of the enhanced cauliflower mosaic virus 35S promoter. qRT-PCR assays confirmed that DCA1 expression was enhanced in all eight overexpression lines, compared with vector-only control ZH11 plants (CK). Two transgenic lines, 35S::DCA1-7 and 35S::DCA1-8, were selected for further analysis (Fig 3A), and the tolerance to drought and salt stress was investigated. Twelve-day-old CK, 35S::DCA1-7 and 35S::DCA1-8 seedlings grown in normal conditions were treated with 100 mM NaCl for 10 days or with 18% PEG for 13 days. As shown in Fig 3B–3D, almost all of the 35S::DCA1 seedlings died, while approximately 70% of the vector-only control plants survived, following a subsequent 7 day recovery in normal conditions. The relative chlorophyll content decreased to ~13% in 35S::DCA1-8 plants vs. ~27% in vector-only control plants after drought treatment (S3A Fig). Fresh weight similarly decreased to ~52% in 35S::DCA1-8 plants vs. ~80% in vector-only control plants after drought treatment (S3C Fig). These results indicated that 35S::DCA1 plants were extremely sensitive to drought and salt stress compared with CK plants. Since there are many differences between laboratory and field stress conditions, we examined the responses of the 35S::DCA1 overexpression plants to drought stress in soil and obtained similar results. Specifically, 35S::DCA1 plants were much more sensitive to drought stress than CK plants (Fig 3E). In a previous study, we failed to obtain plants overexpressing DST, and their phenotypes under stress conditions remain unknown [21]. However after many years of effort, various 35S::DST lines have been prepared and qRT-PCR analysis revealed that expression of DST was only increased 1–2-fold in these plants (S5A Fig). This may be because strong overexpression of this gene causes death. 12-day-old CK, 35S::DST-1 and 35S::DST-2 seedlings grown under normal conditions were stressed with either 100 mM NaCl or 18% PEG and allowed to recover for 7 days. Almost all of the 35S::DST seedlings died, while approximately 70% (NaCl) and 80% (PEG) of the vector-only control plants survived (Fig 4). The relative chlorophyll content decreased to ~7% in 35S::DST-1 plants vs. ~27% in vector-only control plants (S3A Fig). Similarly, fresh weight decreased to ~53% in 35S::DST-1 plants vs. ~80% in vector control plants (S3C Fig). These results showed that 35S::DST plants were much more sensitive to drought and salt stress than controls, consistent with the results obtained for 35S::DCA1 (Fig 3). We also checked whether the reduced stress tolerance of 35S::DCA1 plants was caused by changes in DST expression, but the results suggested this was not the case (S5B Fig). We concluded that DCA1 may therefore function in cooperation with DST in the stress response.
A Tos17 mutant (NF7038) of DCA1 was obtained from the Rice Genome Resource Center. The Tos17 fragment was inserted into the second intron of DCA1, and this reduced DCA1 expression levels by ~80% compared with wild-type Oryza sativa L. japonica. cv. Nipponbare. These results indicate that this mutant (designated dca1) was a genuine DCA1 knockdown mutant (Fig 5A–5C).
We next evaluated whether dca1 plants possessed enhanced tolerance to drought or salt stress. 12-day-old seedlings of Ni (Nipponbare) and dca1 grown under normal conditions (Fig 5D and 5G) were treated with 100 mM NaCl for 9 days (Fig 5E) or 18% PEG for 10 days (Fig 5H) and recovered for 7 days. Approximately 10% of Ni plants survived following a subsequent 7 day recovery period after salt stress, compared with 62% of dca1 plants (Fig 5F). Similarly, the survival rate following recovery from drought stress was significantly increased in the dca1 mutant compared with Ni plants (Fig 5I). The relative chlorophyll content decreased to ~5% in Ni vs. ~16% in dca1 plants following drought treatment (S3B Fig), and the fresh weight decreased to ~43% and ~89% in Ni and dca1 plants, respectively (S3D Fig). These results indicated that dca1 plants were more tolerant to abiotic stress than Ni. Since dca1 is a Tos17 insertion mutant, we knew there may be several copies of the insertion, and Southern blotting revealed more than 10 copies (S4A Fig). We therefore designed an artificial mircoRNA to elicit knockdown of DCA1, and expression of DCA1 in the silenced variant was reduced by 50–80% compared with CK plants. DCA1 microRNA transgenic plants were also more tolerant of drought and salt stress than CK plants (S4B, S4C, S4E and S4F Fig). Dehydration treatment of 66-day-old Ni and dca1 plants grown in soil under normal conditions for 14 days and rewatered for 16 days showed that dca1 was more tolerant to soil dehydration than Ni (Fig 5J–5L). Moreover, dca1 seeds could be harvested from the surviving tillers after re-irrigation (S5C and S5D Fig). Together these results indicated that DCA1 knockdown plants exhibited improved tolerance to drought and salt stress, consistent with the phenotype of the DST mutant dst [21].
We previously demonstrated that DST is a transcription factor with transactivational activity [21]. However, we found that DCA1 does not exhibit this activity directly, since yeast harboring the DCA1-BD construct did not grow on SD/-Ade-His-Leu-Trp medium (Fig 1B). We therefore conducted dual luciferase assays in Arabidopsis protoplasts to confirm the direct effects of the DCA1-DST interaction on gene expression. Seven copies of the GAL4 binding sequence are located before the translational start site of the REN reporter and act as putative cis-acting elements. Compared with empty-vector controls, the REN/LUC ratio was moderately upregulated by DST alone but markedly upregulated when DCA1 and DST were both present, indicating a positive role for this interaction in regulating DST transcriptional activity (i.e., DCA1 enhances the transcriptional activity of DST; Fig 6A and 6B). Peroxidase 24 precursor (Prx 24) is an H2O2 scavenger and target gene of DST, as verified by ChIP and electrophoretic mobility shift assay (EMSA) [21]. Our qRT-PCR results showed that Prx24 was downregulated in dst, dca1 and DmiR (DCA1 artificial microRNA) leaves (Fig 6C and 6D and S4D Fig) and upregulated in 35S::DST and 35S::DCA1 leaves (S6A and S6B Fig).
Although we know that Prx24 is a downstream gene of this transcriptional regulation module, phenotypes of transgenic plants under stress conditions remain unknown. So we generated overexpression plants of this gene in ZH11. Prx24-overexpressing plants became sensitive to drought and salt stress to an extent comparable to DCA1 and DST-overexpressing plants in a ZH11 background (S6C Fig). We identified the DST binding sequence (DBS), 5’-TGCTANNATTG-3’, using a bacterial one-hybrid system [21], and EMSA results showed that His1-DST but not the His1 tag alone could bind to B3, a DNA fragment containing this conserved site (DBS) (Fig 6E). We next wanted to determine if DCA1 could also bind to the DBS, and further EMSA results revealed that His2-DCA1 and His1-DST could bind to the DBS, while His2-DCA1 alone could not (Fig 6E). It is possible that the B3 sequence that we used did not include a DCA1 binding site, or perhaps DCA1 was not associated with DNA at all. To verify our hypothesis, we used the longer promoter sequence of Prx24 described in our previous research [21], and EMSA results were similar and confirmed that DST but not DCA1 could bind to the fragment (S6D Fig). Together, our combined genetic, physiological and molecular evidence indicated that DCA1 functions as an interaction co-activator in DST-mediated transcriptional programs.
Since DST negatively regulates stomatal closure to improve drought and salt tolerance [21], we investigated whether stomatal aperture was also altered in 35S::DCA1 and dca1 plants using Cryo scanning electron microscopy (Cryo SEM). Three levels of stomatal opening in different plants were observed (Fig 7A and S7A Fig) and statistical analysis revealed that 6% of stomata were completely closed in 35S::DCA1 plants vs. 21% in CK plants, while 67% were completely open in 35S::DCA1 plants vs. only 48% in CK plants. The percentage of partially open stomata was 27% and 31% in 35S::DCA1 plants and CK plants, respectively (Fig 7B). In contrast, there was no significant difference in stomatal density and guard cell size between 35S::DCA1 and CK plants (Fig 7C and S7B and S7C Fig). Following dehydration stress, 35S::DCA1 plants lost more water than CK plants (Fig 7D). We next investigated stomatal aperture following drought stress, and statistical analysis revealed that 14% of stomata were completely closed in 35S::DCA1 plants vs. 28% in CK plants, while 37% of stomata were completely open in 35S::DCA1 plants vs. only 27% in CK plants (S7F Fig). The percentage of partially open stomata was 49% and 45% in 35S::DCA1 plants and CK plants, respectively (S7F Fig). The ratio of completely closed stomata was 11% vs. 6% in dca1 mutant and Ni (wild-type) plants, and 60% of stomata were completely open in the dca1 mutant compared with 67% in wild-type plants (Fig 7E). The percentage of partially open stomata was similar in dca1 mutant and wild-type plants (Fig 7E). At the same time, low magnification Cryo SEM images showed that no significant changes occurred in stomatal density in dca1 plants (Fig 7F), and there were also no significant changes in guard cell size (S7D and S7E Fig). Meanwhile, the rate of water loss of detached leaves was lower in the dca1 mutant plants than in wild-type plants (Fig 7G). When under drought stress, statistical analysis revealed that 17% of stomata were completely closed in Ni plants vs. 25% in dca1 plants, while 38% of stomata were completely open in Ni plants vs. only 27% in dca1 plants (S7G Fig). The percentage of partially open stomata was 44% and 48% in Ni plants and dca1 plants, respectively (S7G Fig). Together these results suggest that the enhanced drought tolerance observed in the dca1 mutant is mainly due to increased stomatal closure, which minimizes water loss.
The phytohormone ABA is known to induce stomatal closure [23], and measurement of the endogenous ABA content revealed no significant difference between dst mutant and wild-type plants [21]. H2O2 can also induce leaf stomatal closure [16, 24, 25], and Huang et al. observed increased stomatal closure in the dst mutant as a result of H2O2 accumulation [21]. We therefore examined the H2O2 content in the stomata of DCA1-overexpressing and mutant plant leaves using the fluorescent dye 2’, 7’-dichlorodihydrofluorescein diacetate (H2DCFDA). We found that overexpression plants contained ~20% less H2O2 in their guard cells than did CK plants (Fig 8A and 8B). However, mutant plants accumulated more H2O2 in their guard cells than did Ni plants (Fig 8C and 8D). To explore the molecular mechanism behind this phenomenon, we isolated guard cell protoplasts (GCP) and mesophyll cell protoplasts (MCP) from ZH11, dst, Ni and dca1 plants to examine the expression levels of DCA1, DST and Prx24. The results revealed higher levels of Prx24 and DST expression in GCPs than MCPs of ZH11, but DCA1 expression was not elevated (Fig 8E). Prx24 was expressed at relatively lower levels in dca1 mutant and dst mutant GCPs than in CK plants (Fig 8F), and since Prx24 functions as a ROS scavenger, these results indicated that the H2O2 content in stomata may be affected by DCA1-DST.
While carefully observing the phenotypes of DST-complemented (DSTpro::DST-GFP in dst) plants under normal and stress conditions, we found that the dst mutant was not fully complemented. Specifically, the phenotype of DSTpro::DST-GFP under stress conditions was intermediate between those of wild-type and the dst mutant plants (Fig 9A–9D). We hypothesized that perhaps the mutated form of DST exhibiting reduced transcriptional activity could still function in the transcriptional complex. A BiFC assay in Arabidopsis protoplasts revealed that DST interacts with itself in vivo (Fig 9E), and an in vitro pull-down assay showed that MBP-DST was able to pull down the His2-DST protein (Fig 9F and S8A Fig). Meanwhile, the mutant site of DST did not influence its dimerization (Fig 9G). These results confirm that our hypothesis was correct, and indicate that DST functions as a dimer. In addition, the complex containing the mutant isoform could not fully execute its transactivation function, which may explain why the dst mutant was not fully complemented. We also examined whether DCA1 functions in the same manner with DST, and pull-down assays revealed that it does not (S8B Fig). Further analysis revealed that the DST-DST interaction was negatively influenced by stress in Arabidopsis protoplast cells (S9A, S9B and S9E Fig). This may be due to decreased levels of DST monomer or decreased DST dimerization. To investigate further, DST self-promoter promoted DST-nYFP and DST-cYFP were cotransformed into Arabidopsis protoplasts and cultured for 9 h under normal conditions. Half of the cells were maintained in normal conditions, while the other half were treated with NaCl and cycloheximide (CHX), an inhibitor of protein biosynthesis. After 3 h, the relative fluorescence ratios of the cells were similar (S9C, S9D and S9F Fig). These results demonstrate that DST was inhibited by stress, but not degraded or dissociated into monomeric form, and this was verified by qRT-PCR experiments (S2B and S2C Fig). These results, when combined with those from the EMSA experiments (Fig 6E), indicate that DST and DCA1 may form a putative heterotetrameric complex that participates in gene regulation.
In this study, we identified the DST interacting protein DCA1. Overexpression of DCA1 increased the plant’s sensitivity to abiotic stress, and 35S::DST plants were also highly sensitive to salt and drought stress. Knockdown of DCA1 improved the tolerance to stress, as also observed in the dst mutant. These results indicate that DCA1 has a positive effect on the transcriptional activity of DST, which was confirmed by dual luciferase assays. The qRT-PCR results showed that DST was suppressed under stress conditions while DCA1 was induced (Fig 2E and 2F and S2B and S2C Fig). Hsu, et al. (2014) identified ZFP34, a homolog of DCA1, and discovered that ZFP34 participates in the heat stress response. Overexpression of this gene in rice and its homolog in Arabidopsis increased stomatal opening, and ZFP34 mutants exhibited decreased stomatal opening in both rice and Arabidopsis [22]. Heat stress is usually accompanied by drought, since transpiration is induced when plants experience high temperatures. DCA1 expression is therefore likely to be upregulated by drought, salt and heat stress. Water loss caused by increased transpiration may aggravate the impact of drought stress, and plants have evolved very sophisticated molecular mechanisms to overcome this paradoxical phenomenon. One possible molecular mechanism may be the DCA1-DST pathway, since DCA1 may induce the opening of stomata to reduce the internal temperature during heat stress. At the same time downregulation of DST may cause closure of stomata to prevent excessive moisture loss. This seemingly contradictory phenomenon ensures that the plant can adapt to various environmental changes in different ways, which builds robustness into the plant responses to abiotic stress.
The DCA1 protein contains two zinc ion binding sites, a CHY zinc finger domain and a RING-H2 domain. The CHY domain is highly conserved among homologous proteins from different species, but the RING-H2 domain is not highly conserved. Therefore, the CHY domain may be the major functional part of this protein family, but the exact function was unknown until now. In the present study, we found that DCA1 functions as a transcriptional co-activator. However, the means by which DCA1 promotes the activity of DST is not yet elucidated. DCA1 may alter the conformation of DST through the protein-protein interaction, or may help DST to recruit other transcription initiation factors such as RNA polymerase, or may even function as a regulator of chromatin structure. Many RING finger domain proteins function as ubiquitin ligases, although many are also involved in other roles such as Breast Cancer 1 (BRCA1) which participates in transcriptional regulation, DNA damage repair and chromatin remodeling [26, 27]. Mex-3 homolog D (MEX3D) is an RNA binding protein in C. elegans that has a negative regulatory action on Bcl-2 expression at the posttranscriptional level [28], and peroxisome biogenesis factor 10 (PEX10) is involved in import of peroxisomal matrix proteins [29]. Thus, DCA1-like proteins may also perform a range of different functions.
Previous studies focusing on H2O2 signalling in guard cells during stress conditions have indicated that NADPH oxidases such as Atrboh D and Atrboh H may play major roles in this process [10]. Indeed, the atrbohD/F double mutant exhibited reduced stomatal closure and ROS production compared with wild-type plants. We previously identified the transcription factor DST that binds to the DBS element to regulate the expression of H2O2 metabolism-associated genes in stomata, and extended this work by identifying the DST interacting CHY zinc finger protein DCA1 in the present study. This transcriptional complex appears to regulate the expression of Prx 24, an H2O2 scavenger that we found to be more highly expressed in guard cells than in mesophyll cells. Our results reveal how plants prolong H2O2 signalling by reducing the catabolism of H2O2 through the DCA1-DST-Prx24 pathway. Based on these results, we propose a simple model to explain the role of the DCA1-DST-Prx24 pathway in regulating the status of guard cells and stress tolerance. In this model, DCA1 functions as an interacting co-activator of DST to form a heterotetrameric complex that regulates H2O2 homeostasis and stomatal aperture, ultimately affecting stress tolerance (Fig 10).
Drought and salinity have had devastating effects on food production throughout human history, and the problems are increasing with the growing population. During 2010–2012, drought occurred on most continents, and millions of hectares of crops were destroyed, from wheat in Europe to corn in the US to rice in China [30, 31]. In 2014, the state of California in the US suffered its worst drought on record [32]. Engineering the stomatal response to improve water use efficiency represents a powerful tool for enhancing drought tolerance in plants. Indeed, the constitutive reduction in stomatal opening in the rice dca1 and dst mutant enhances drought tolerance. Given the high conservation of DCA1 among plants, our research suggests that inhibiting the functional association of DCA1-DST may provide one route to unlocking improved drought and salt tolerance in a variety of crops.
All rice genetic stocks used in this study were in the Zhonghua 11 (ZH11) japonica variety background except for dca1, which is in the Nipponbare (japonica variety) background. 35S::DCA1 plants were generated via transformation with a construct produced by inserting an amplified ORF fragment containing DCA1 from ZH11 cDNA between the CaMV35S promoter and the 3’OCS terminator. 35S::DST plants were generated using a similar method. The dca1 mutant (NF7038) was obtained from the Rice Genome Resource Center.
We generated knockdown transgenic plants using an artificial microRNA (designed by http://wmd3.weigelworld.org/cgi-bin/webapp.cgi). Target site on DCA1 was 5’- acgtgatgagagcgcatcatc-3’. The construction method was modified from Norman Warthmann et al. [33]. For the miRNA construct, three modification PCRs were performed with primers G-11491 + DmiR II, DmiR I + DmiR IV and DmiR III + G-11494 on ZH11 cDNA as template, yielding three fragments, respectively (S1 Table). The three resulting fragments were gel purified (Qiagen) and then fused by one PCR with the two flanking primers G-11491 + G-11494 on a mixture of 1 ul from each of the previous three PCRs as template. The fusion product was again gel purified (Qiagen), cloned into the overexpression vector to generate DCA1 knockdown transgenic plants.
Seeds were submerged in water at room temperature for 48 h, followed by germination for 18 h at 37°C. Seeds were then sown in a bottomless 96-well plate that was placed in a container of Yoshida’s culture solution and incubated in a growth chamber under a 13 h light (25°C)/11 h dark (23°C) photoperiod. For salt treatment, 20-day-old seedlings were transferred to a culture solution containing 100 mM NaCl. For PEG treatment, 20-day-old seedlings were transferred to a culture solution containing 18% (w/v) PEG4000. For soil drought experiments, 2-week-old seedlings that were cultured in the growth chamber were transplanted to containers with soil and grown in a greenhouse at 24–30°C and 50–60% relative humidity. About 50 days later, water was removed from the containers for the dehydration treatment.
Screening for DST interacting proteins was performed according to the manufacturer’s instructions (Clontech). Briefly, libraries were prepared using total RNA extracted from ZH11 leaves that were approximately 20 days old. The DST fragment lacking the self-activation domain (amino acids 1–72) was cloned into the pGBKT7 BD vector to generate pGBKT7/△72 (S1 Table). Screening was performed on SD/–Leu/–Trp medium containing 2.5 mM 3-aminotriazole (3-AT). To confirm positive interactions, full-length ORFs of each target gene were cloned into the pGADT7 AD vector and cotransformed with pGBKT7/△72 into Y2HGold Competent Cells. To verify positive interactions, the full-length DST ORF was cloned into the pGADT7 AD vector, and target genes were cloned into the pGBKT7 BD vector, and Y2HGold Competent Cells were cotransformed with 100 ng of each vector pair. Cells were grown at 30°C for 3 days on SD/–Ade/–His/–Leu/–Trp medium containing 1–10 mM 3-AT (depending on the bait—prey combination).
DST (GQ178286) and DCA1 (Os10g0456800) coding regions were cloned into YFP expression vectors, BiFC vectors or dual luciferase system vectors. For subcellular localization experiments, 10 μg of YFP-DST, YFP-DCA1 or YFP expression plasmid was transformed into Arabidopsis protoplasts and observed under a confocal microscope (Carl Zeiss). For BiFC experiments, 5 μg each of YFP-N- and YFP-C-tagged protein expression vector was cotransformed. For the dual luciferase transient transcription activity assay, a 7× GAL4 binding sequence was inserted into the pGreenII 0800-LUC vector to generate the LUC reporter construct. The coding sequence of DST was inserted into pGreen GAL4BD to produce the effector construct, and DCA1 was inserted into pGreenII 62-SK (S1 Table). Firefly LUC and Renilla luciferase (REN) activities were measured using the dual luciferase reporter assay system (Promega). The LUC:REN ratio represents transcriptional activity.
The DST ORF was inserted into pET32a, pCold TF and pMAL-c5x vectors. The DCA1 coding sequence was inserted into pCold TF and pMAL-c5x (S1 Table). The tag in pET32a was designated His1 and that in pCold TF was designated His2. The pMAL-c5x constructs were transformed into TB1 competent cells, and pET32a and pCold TF vectors were transformed into DE3 competent cell.
The pull-down assay was performed according to the instructions for the MagneGST Pull-Down System (Promega) with some modifications. Briefly, E. coli cells expressing DST-MBP and MBP were lysed with BugBuster Protein Extraction Reagents (Novagen) and centrifuged. The supernatant was incubated with Anti-MBP Magnetic Beads (NEB) for 30 min at 4°C and washed five times with MBP column buffer. Then, supernatants containing DCA1-His2 and His2 proteins were incubated with approximately the same amount of MBP and DST-MBP binding beads for 2 h at 4°C. The magnetic beads were then washed five times with MBP column buffer. The mixture was resuspended in SDS loading buffer, boiled for 3 min, separated by 10% SDS-PAGE and immunoblotted with anti-MBP antibody for target proteins and anti-His antibody for pull-down proteins (CWbiotech).
Leaves from 2-week-old seedlings were harvested and total RNA was extracted using TRIzol reagent (Invitrogen) according to the manufacturer’s instructions. 1–2 μg of total RNA was used for cDNA synthesis with oligo (dT) primer and Superscript Reverse Transcriptase (Invitrogen). Quantitative RT-PCR was performed using FastStart Universal SYBR Green Master (Roche) and an ABI 7300 Real-time PCR System (Applied Biosystems). The expression of DST, DCA1 and Prx24 was normalized against actin. Oligonucleotide primers are listed in S1 Table.
Leaves from 20-day-old seedlings grown under normal conditions or treated with 12% PEG for 6 hours were detached and frozen in liquid nitrogen. Images of stomata were obtained using a Cryo (QUORUM) scanning electron microscopy (JEOL) system (Cryo SEM).
H2O2 production in guard cells was detected using H2DCFDA (Molecular Probes) as described by Huang [21]. Epidermal strips were peeled from leaves of 20-day-old rice seedlings using tissue forceps, washed with loading buffer (10 mM Tris-HCl, 50 mM KCl, pH 7.2) and incubated in staining buffer (loading buffer containing 50 mM H2DCFDA) for 10 min at 25°C in the dark. Strips were washed twice with distilled water to remove excess staining buffer, and fluorescence was measured using a confocal laser-scanning microscope (Olympus) with excitation at 488 nm and emission at 525 nm. All images were taken under identical conditions. To quantify fluorescence, the guard cell region was selected and the mean fluorescence value was calculated using the microscope’s software. Each sample contained at least eight independent leaves, and approximately 10 randomly selected stomata were analyzed per leaf.
30 fully expanded one-month-old rice leaves were cut into ~0.3 cm sections and blended with 200 mL Lysis buffer (5 mM CaCl2, 0.5 mM ascorbic acid, 0.1% PVP-40, 10 mM MES, pH 6.0) for 90 s using a blender (Waring blender model LB20ES). The blended material was filtered through a 200 μm nylon mesh, washed once with 200 ml ddH2O, suspended in 100 ml suspension buffer (0.25 M mannitol, 1mM CaCl2, 0.5mM ascorbic acid), filtered through a 30 μm nylon mesh and transferred into a flask with 25 mL of an enzyme mixture containing 0.7% Cellulysin cellulase, Trichoderma viride (Sigma), 0.1% (w/v) PVP-40, 0.25% BSA in 55% (v/v) basic medium (5 mM MES, 0.5 mM CaCl2, 0.5 mM MgCl2, 10 μM KH2PO4, 0.5 mM ascorbic acid (Sigma), 0.55 M sorbitol, pH 5.5), and 45% (v/v) deionized water. The flask was placed in a shaking water bath at 24°C with the shaking speed set to 120 rpm. After 3 h of digestion, 75 mL basic medium was added, and the mixture was shaken for an additional 5 min. The mixture was filtered through a 70 μm nylon mesh and washed twice with 50 mL basic medium. MCPs were collected by centrifuging the filtrate at 200 g for 5 min. Epidermal fragments were transferred to another flask with 50 mL of basic medium (1.5% Cellulase RS (Yakult, Japan), 0.02% Pectolyase Y23 (Yakult, Japan), 0.25% BSA, pH 5.5) and incubated at 24°C in a shaking water bath (100 rpm). After incubation for 3 h, the mixture was filtered through a 30 μm nylon mesh, washed twice with 50 mL basic medium, and GCPs were collected at 200 g for 5 min. Two different transcription inhibitors, actinomycin D and cordycepin, were used to prevent changes in gene expression during protoplast isolation.
EMSA was performed as described by Huang [21] using His1 tag, DST-His1, His2 tag and DCA1-His2 recombinant fusion proteins. Biotin-labeled oligonucleotides were obtained by PCR amplification using biotin-labeled primers (Invitrogen). The binding reactions (containing biotin-labeled DNA and different combinations of proteins) were performed with a Light Shift Chemiluminescent EMSA kit (Pierce) according to the manufacturer’s instructions. Gel electrophoresis was performed on a 10% native polyacrylamide gel (79:1 acryl/bis). After blotting onto a nylon membrane, DNA was attached to the membrane using a UV light cross-linker instrument. Blots were imaged using a Tanon 5200 imaging system (bioTanon).
Peptide sequences were obtained from Phytozome (Phytozome v9.1, http://www.phytozome.net/) and sequence alignment was performed using the Clustal Omega program (http://www.ebi.ac.uk/Tools/msa/clustalo/).
|
10.1371/journal.pgen.1000902 | GTPase Activity Plays a Key Role in the Pathobiology of LRRK2 | Mutations in the leucine-rich repeat kinase 2 (LRRK2) gene are associated with late-onset, autosomal-dominant, familial Parkinson's disease (PD) and also contribute to sporadic disease. The LRRK2 gene encodes a large protein with multiple domains, including functional Roc GTPase and protein kinase domains. Mutations in LRRK2 most likely cause disease through a toxic gain-of-function mechanism. The expression of human LRRK2 variants in cultured primary neurons induces toxicity that is dependent on intact GTP binding or kinase activities. However, the mechanism(s) underlying LRRK2-induced neuronal toxicity is poorly understood, and the contribution of GTPase and/or kinase activity to LRRK2 pathobiology is not well defined. To explore the pathobiology of LRRK2, we have developed a model of LRRK2 cytotoxicity in the baker's yeast Saccharomyces cerevisiae. Protein domain analysis in this model reveals that expression of GTPase domain-containing fragments of human LRRK2 are toxic. LRRK2 toxicity in yeast can be modulated by altering GTPase activity and is closely associated with defects in endocytic vesicular trafficking and autophagy. These truncated LRRK2 variants induce similar toxicity in both yeast and primary neuronal models and cause similar vesicular defects in yeast as full-length LRRK2 causes in primary neurons. The toxicity induced by truncated LRRK2 variants in yeast acts through a mechanism distinct from toxicity induced by human α-synuclein. A genome-wide genetic screen identified modifiers of LRRK2-induced toxicity in yeast including components of vesicular trafficking pathways, which can also modulate the trafficking defects caused by expression of truncated LRRK2 variants. Our results provide insight into the basic pathobiology of LRRK2 and suggest that the GTPase domain may contribute to the toxicity of LRRK2. These findings may guide future therapeutic strategies aimed at attenuating LRRK2-mediated neurodegeneration.
| Parkinson's disease (PD) is the second most common neurodegenerative disorder. PD is considered to be caused by a combination of risk factors including environmental exposure, age, and a positive family history for disease. Several genes have been unambiguously implicated in PD. However, our knowledge is still limited about these genes and the disease mechanisms involved. Mutations in the LRRK2 gene account for up to 40% of PD in certain populations. Since a single-cell model, baker's yeast, has been employed successfully to study the function of genes related to PD and other neurodegenerative disorders, we developed a yeast model of LRRK2 cytotoxicity in this study to investigate the function of LRRK2. We dissected the LRRK2 protein into different fragments including the various functional domains and found that fragments including the GTPase domain of LRRK2 are toxic. This toxicity can be modulated by alterations in GTPase activity and correlates with defects in cellular trafficking. These truncated LRRK2 variants induce similar toxicity and trafficking defects in both yeast and primary neuronal models. This yeast model reveals an important role of GTPase activity in the basic pathobiology of LRRK2 and may guide future therapeutic strategies for PD.
| Parkinson's disease (PD (OMIM #168600)) is a common neurodegenerative movement disorder that is characterized by muscular rigidity, bradykinesia, resting tremor and postural instability [1],[2]. Although typically a sporadic disease, mutations in the leucine-rich repeat kinase 2 (LRRK2, PARK8, OMIM #607060, GenBank #AY792511) gene have been identified as a cause of late-onset, autosomal dominant familial PD that is clinically and neurochemically indistinguishable from sporadic PD [3]–[7]. Importantly, LRRK2 pathogenic mutations also contribute to sporadic PD [4],[8]. Mutations in LRRK2 are the most common cause of familial and sporadic PD identified to date [9]. The LRRK2 gene encodes a large protein of 2527 amino acids that contains multiple domains. These include a LRRK2-specific repeat region, multiple leucine-rich repeats, a Ras of Complex (Roc) GTPase domain, a C-terminal of Roc (COR) domain, and a protein kinase domain belonging to the tyrosine kinase-like protein kinase family [10],[11]. LRRK2 exhibits kinase activity whereby it can undergo autophosphorylation and can phosphorylate generic substrates [12]–[18]. However, physiological substrates for the kinase activity of LRRK2 have not yet been identified. The GTPase domain of LRRK2 can mediate GDP (guanosine-5′-diphosphate)/GTP (guanosine-5′-triphosphate) binding as well as GTP hydrolysis albeit at a relatively slow rate compared to other small GTPases such as Ras [14], [15], [19]–[22]. Intriguingly, GTP binding markedly enhances the kinase activity of LRRK2 and is an essential requirement for kinase activity [14],[15],[21],[22]. It is unclear at present how the GTP binding and GTP hydrolysis activities of LRRK2 are regulated. Disease-associated mutations located throughout the LRRK2 protein have been shown to variably alter GTP binding, GTP hydrolysis or kinase activity [14]–[24]. Thus, alterations in both GTPase and protein kinase activity are clearly important for the development of PD due to LRRK2 mutations.
A number of useful models have been developed to investigate the pathobiology of LRRK2 disease-associated variants, including Drosophila, transgenic mice and primary neuronal models. Studies in cultured primary cortical neurons reveal that the exogenous expression of pathogenic mutant forms of full-length human LRRK2 (i.e. G2019S, R1441C and Y1699C) induces marked neuronal toxicity relative to the wild-type protein [21]–[23]. Wild-type LRRK2 can also induce neuronal toxicity but to a lesser degree. LRRK2-induced toxicity in this neuronal model is dependent on intact GTP binding and kinase activity [21]–[23]. In Drosophila models, expression of human LRRK2 variants induces selective dopaminergic neurodegeneration and motor dysfunction [25]–[27]. Mutant LRRK2 R1441G BAC transgenic and R1441C knock-in mice exhibit mild defects in dopaminergic neurotransmission and motor deficits [28],[29]. These observations are consistent with a toxic gain-of-function mechanism for disease-associated LRRK2 variants. The molecular mechanism(s) and/or pathway(s) by which LRRK2 variants induce neuronal toxicity are poorly understood and how alterations in GTPase or kinase activities regulate the toxic effects of LRRK2 are not well defined.
Model organisms including yeast, worms, flies and mice are commonly used to uncover the fundamental biology and pathobiology of proteins associated with neurodegenerative diseases, including poly-glutamine expansion disorders, Parkinson's disease, Alzheimer's disease, Prion diseases and Friedreich's ataxia. The baker's yeast Saccharomyces cerevisiae, a eukaryotic single-cell organism, provides a powerful experimental system in which to dissect complex biological pathways and processes. Major advantages of yeast include the high degree of conservation of pathways, processes and protein function with mammalian cells, and the accessibility of yeast cells to genetic manipulation and genome-wide screening approaches. For Parkinson's disease (PD), yeast have provided unique insight into the basic biology and pathobiology of the α-synuclein protein that is associated with autosomal dominant familial PD [30]–[33]. Here, we have employed yeast as a model to further understand the basic pathobiology of LRRK2. Expression of truncated human LRRK2 reduces yeast viability in a manner largely dependent on the GTPase domain of this protein. Reduced viability in this yeast LRRK2 model is independent of kinase activity and disease-associated mutations, but can be modulated instead by altering GTPase activity and is associated with defects in vesicular trafficking and autophagy. This yeast model provides insight into the basic pathobiology of LRRK2 and suggests that the GTPase domain may contribute to the cellular toxicity of LRRK2. These findings may guide future therapeutic strategies aimed at attenuating LRRK2-mediated neurodegeneration.
To gain novel insight into the pathobiology of LRRK2, we set out to develop a simple yeast LRRK2 model. Yeast cells were transformed with expression constructs that express at high copy V5-tagged full-length human LRRK2 under the control of the galactose-inducible GAL1 promoter. Expression of wild-type (WT) or G2019S LRRK2 variants fail to affect the viability of yeast cells, which is most likely due to the formation of large LRRK2-positive intracytoplasmic inclusions that are biochemically insoluble (Figure S1). The same results are observed with low copy expression constructs (Figure S1). Thus, we elected to examine the detrimental effects of various smaller protein fragments of human LRRK2 that contain different functional domains.
Following galactose induction of high copy expression constructs, LRRK2 fragments minimally containing the GTPase domain markedly reduce yeast viability relative to control cells, with the most toxic fragment containing the central GTPase, COR and kinase domains (GTP-COR-Kin) of LRRK2 (Figure 1A). A larger LRRK2 fragment additionally containing the C-terminus (GTP-COR-Kin-CT) reduces yeast viability to a similar extent. The GTPase domain alone is also sufficient to markedly reduce yeast viability (Figure 1A). LRRK2 fragments containing the kinase domain alone (Kin or Kin-CT) or a fragment lacking the N-terminal region (ΔN-LRRK2), which is poorly expressed, are much less toxic to yeast (Figure 1A). Western blot analysis confirms the expression of each LRRK2 fragment in yeast following galactose induction (Figure 1B). LRRK2 fragments exhibit similar diffuse cytoplasmic localization patterns in yeast as revealed by fluorescence microscopy (Figure S2). The loss of viability due to the expression of each LRRK2 fragment is confirmed by monitoring the growth rate of yeast cells in liquid media following galactose induction (Figure 1C).
We focused further on the GTP-COR-Kin fragment of LRRK2 throughout this study since its expression is most toxic to yeast cells and because it permits further analysis of the contribution of both enzymatic domains. To test if the toxicity is dose-dependent, we also examined the effects of low copy expression of the GTP-COR-Kin fragment. A similar phenotype is observed as with high copy expression of the GTP-COR-Kin fragment (Figure 1A and 1C). Thus, the protein length, expression levels or cellular localization of each LRRK2 fragment do not correlate with their effects on yeast viability suggesting that alterations in viability are dependent on the protein domain composition or activity of each LRRK2 fragment. Moreover, these data demonstrate that LRRK2 protein fragments that contain the GTPase domain, but not full-length LRRK2, can reduce the viability of yeast cells.
Since expression of the GTPase domain of LRRK2 is sufficient to markedly reduce yeast viability, we sought to determine whether alterations in GTPase activity could influence this growth deficit. A number of missense mutations were introduced into the GTPase domain within the GTP-COR-Kin LRRK2 fragment that are predicted to functionally alter enzymatic activity (Figure 2A). Two mutations, K1347A and T1348N, disrupt the conserved guanine nucleotide phosphate-binding loop motif (P-loop, residues 1341–1348) and prevent GDP/GTP binding to the GTPase domain [15],[22]. Two other mutations, R1398L and R1398Q, were targeted at the R1398 residue, a highly conserved glutamine residue in most small GTPases (i.e. Q61 in H-Ras). LRRK2 contains a highly conserved DFAGR motif (residues 1394–1398) in the switch II region which is mainly responsible for GTP hydrolysis. The P-loop residue T1343 is a glycine residue (G12) in H-Ras. In H-Ras, the combined G12V and Q61L mutations create a GTPase-inactive form of this protein, which is constitutively GTP-bound and active. We introduced these two key H-Ras residues into LRRK2 via the analogous mutations T1343G and R1398Q (RQ/TG) to create a Ras-like GTPase that leads to increased GTP hydrolysis activity (Figure 2A) [15]. Moreover, a common R1441C pathogenic variant was also introduced into the GTPase domain of LRRK2. Expression of the GTP-COR-Kin fragment of LRRK2 containing each mutation was induced by spotting yeast cells onto galactose media. Remarkably, altering the GTPase activity of LRRK2 leads to marked changes in yeast viability (Figure 2B). Compared to WT LRRK2, the GTP binding-deficient mutants K1347A and T1348N cause a dramatic reduction in yeast viability whereas the mutant R1398L and Ras-like mutant RQ/TG partially improve viability (Figure 2B). The disease-associated R1441C variant reduces yeast viability similar to WT LRRK2 (Figure 2B). Western blot analysis reveals that each mutant LRRK2 fragment is expressed at similar levels, which excludes alterations in expression level as a cause of their differential effects on yeast viability (Figure 2C). Furthermore, fluorescence microscopic analysis fails to reveal obvious differences in the cellular localization of truncated LRRK2 GTPase variants with each variant adopting a similar diffuse cytoplasmic distribution in yeast cells (Figure S2). Growth impairments induced by expression of each mutant LRRK2 fragment in yeast are further confirmed in liquid media following galactose induction (Figure 2D).
To determine how alterations in the GTPase activity of LRRK2 due to each functional mutation correlate with changes in yeast viability, we examined both the GTP binding and GTP hydrolysis activities of each mutant LRRK2 fragment. GTP binding was measured using an established GTP-sepharose pull-down assay on total yeast proteins expressing each LRRK2 fragment (Figure 2E). WT and the disease-associated mutant R1441C LRRK2 bind to immobilized GTP to similar extents whereas surprisingly all other mutants exhibit significantly reduced GTP binding (Figure 2E). Consistent with prior reports of full-length LRRK2 [14], [15], [19]–[22], the P-loop mutations, T1348N and K1347A, impair the GTP binding of LRRK2 (Figure 2E). Importantly, the GTP binding capacity of each LRRK2 GTPase mutant does not correlate with its effects on yeast viability. It is not currently possible to measure the capacity of each mutant LRRK2 fragment to bind GDP. It is likely that certain mutations (i.e. the P-loop mutants K1347A and T1348N) impair GDP/GTP binding whereas other mutations (i.e. the Ras-like mutant RQ/TG and R1398L) may alter the affinity for binding to GDP and GTP.
The effects of each mutation on LRRK2-mediated GTP hydrolysis were also determined in vitro by measuring the release of the γ-phosphate moiety from GTP (Figure 2F). Truncated WT LRRK2 displays detectable GTP hydrolysis activity whereas the R1441C mutant exhibits a small reduction in activity, similar to previous reports [14],[19],[20]. As expected, the Ras-like RQ/TG mutant leads to a marked increase in GTP hydrolysis activity but unexpectedly the R1398L mutant produces a similar increase in activity. The P-loop mutants K1347A and T1348N essentially abolish the GTP hydrolysis activity of LRRK2 as expected (Figure 2F). Therefore, alterations in GTP hydrolysis activity of each truncated LRRK2 GTPase mutant correlate closely with their effects on yeast viability. In this case, increased GTP hydrolysis partially improves the viability of yeast compared to WT LRRK2 whereas impaired hydrolysis dramatically reduces yeast viability. Notably, alterations in kinase activity via introduction of kinase-impaired (i.e. K1906M or T2031A/S2032A/T3035A) or a kinase-hyperactive (i.e. G2019S) mutation fails to similarly influence LRRK2-induced toxicity in yeast (Figure S3).
To further examine if GTPase activity plays a key role in the toxic process, we investigated the GTPase activity of full-length human LRRK2 harboring the most frequent mutations causing PD. Importantly, the mutations R1441C/G in the GTPase domain and Y1699C in the adjacent COR domain, significantly decrease GTPase activity (Figure 2G) although the mutations, G2019S and I2020T, in the kinase domain do not have a significant effect, suggesting that impaired GTP hydrolysis of LRRK2 can contribute to PD.
In yeast cells expressing human α-synuclein (SNCA, PARK1/4, OMIM #163890, GenBank #BC108275), defects in vesicular trafficking have been shown to underlie the cytotoxic effects of this protein with the earliest defect being a block in ER-to-Golgi vesicular trafficking [30]–[32]. Since α-synuclein pathology is a common feature of patients with LRRK2 mutations [3],[7],[34], vesicular trafficking was examined to determine whether similar defects could also underlie LRRK2-induced toxicity in yeast. The lipophilic fluorescent dye, FM4–64, is useful for monitoring endocytosis in yeast. FM4–64 binds to the plasma membrane of yeast cells where it is internalized by endocytosis into vesicles that subsequently undergo trafficking to the vacuole via the early and late endosome compartments. Thus, FM4–64 dye selectively stains the yeast vacuolar membrane appearing as a large ring-like cytoplasmic structure.
Yeast cells expressing truncated LRRK2 variants following galactose induction were incubated with FM4–64 dye and live-cell imaging was conducted by confocal fluorescence microscopy. WT LRRK2 expression partly disrupts the normal trafficking of FM4–64 to the vacuolar membrane relative to control cells, which exhibit normal ring-like vacuolar staining (Figure 3A). WT LRRK2 expression results in the appearance of large cytoplasmic punctate structures in addition to normal vacuolar staining, suggesting a modest defect in trafficking of FM4–64-labeled vesicles to the vacuole leading to their accumulation in endosomes. Yeast cells expressing truncated LRRK2 containing the two most toxic GTPase mutations, K1347A and T1348N, which impair the GTP binding and hydrolysis activity of LRRK2, exhibit severe defects in the endocytic trafficking pathway with a dramatic increase in the appearance of labeled punctate structures and the complete absence of normal vacuolar membrane staining (Figure 3A). Truncated LRRK2 variants that partially improved yeast viability compared to WT protein (i.e. RQ/TG and R1398L) induce similar trafficking defects to WT LRRK2 (Figure 3A). Normal FM4–64 labeling of vacuolar membranes is observed when yeast cells are grown in glucose media (data not shown). DIC images show that cells expressing each of the LRRK2 fragments have normal vacuolar morphology (Figure 3D).
Quantitation of defective endocytic trafficking reveals that the toxic GTPase-inactive mutants, K1347A and T1348N, lead to a significant increase in the number and frequency of FM4–64-labeled punctate structures per cell compared to WT LRRK2, whereas the GTPase-active mutants, RQ/TG and R1398L, display a small non-significant reduction in the number of punctate structures relative to WT (Figure 3B and 3C). Punctate structures are not normally observed in control yeast cells (Figure 3B and 3C). The vesicular trafficking defects induced by expression of each truncated LRRK2 GTPase variant in yeast do not correlate with alterations in their cellular localization (Figure S2). In particular, there is no specific enrichment in the vacuole or endosomal compartments of each LRRK2 variant that would obviously account for their differential effects on endocytic vesicular trafficking (Figure S2). These results indicate that the endocytic vesicular trafficking defect in yeast is associated with alterations in LRRK2 GTPase activity and likely underlies toxicity in yeast induced by truncated LRRK2.
To verify that the observed defects induced by LRRK2 expression in yeast are due to vesicular trafficking pathways rather than simply by protein aggregation, yeast cells expressing truncated LRRK2 variants following galactose induction were examined by transmission electron microscopy (TEM) (Figure 4). Interestingly, yeast cells expressing truncated LRRK2 containing the two most toxic GTPase mutations, K1347A and T1348N, which impair GTPase activity exhibit a significant increase of autophagic vacuoles (AVs) (74.7% in K1347A cells and 86.2% in T1348N cells) compared to WT LRRK2 (19.4% AVs) (Figure 4A and 4B). In contrast, AVs were uncommon in yeast cells carrying empty vector (9.2% AVs) (Figure 4A and 4B). In accordance with fluorescence localization studies of truncated LRRK2 variants in yeast (Figure S2), protein aggregates or inclusions were not readily observed in the electron micrographs. Taken together these data indicate that LRRK2-induced trafficking defects are mediated at least in part by alterations in autophagy in addition to effects on the endocytic vesicular trafficking pathway.
To provide insight into the mechanism of LRRK2-induced toxicity in yeast, and to determine whether there are differences or similarities with α-synuclein-induced toxicity, a small candidate genetic screen was performed in yeast focused on modifiers of α-synuclein-induced toxicity. We elected to analyze potent modifiers of human α-synuclein-induced toxicity Ypt1 (GenBank #AAS56793) and Ykt6 (GenBank #AAB32050) [30],[31], as well as Hsp31 (Genbank #AAB64972), the yeast ortholog of human DJ-1 [33], a neuroprotective redox-responsive protein associated with familial PD (PARK7, OMIM #606324) [35],[36]. Yeast cells were transformed with constructs expressing truncated WT LRRK2, each candidate protein alone, or both proteins together under the control of the GAL1 promoter and viability was examined by spotting of yeast cells on to galactose media. Expression of WT LRRK2 alone reduces yeast viability, whereas co-expression with each of the three candidate proteins fails to suppress the LRRK2-induced growth deficit (Figure S4). The three candidate yeast proteins were also tested for their ability to suppress toxicity due to the expression of the truncated LRRK2 variants, K1347A and T1348N, which induce a more pronounced loss of viability in yeast than WT LRRK2. Co-expression with each of the three candidate proteins also fails to suppress the K1347A- or T1348N-induced growth deficit (Figure S4). Collectively, our data demonstrate that known potent suppressors of α-synuclein-induced toxicity in yeast (i.e. Ypt1 and Ykt6) do not specifically suppress LRRK2-induced toxicity in this model suggesting that α-synuclein and LRRK2 induce toxicity in yeast through distinct pathways.
Following expression of truncated LRRK2 variants, we also fail to observe defects in the normal trafficking of carboxypeptidase Y (CPY) and alkaline phosphatase (ALP) proteins from the endoplasmic reticulum (ER) to the vacuole by pulse-chase analysis (data not shown), which represent two distinct biosynthetic transport pathways that converge upon the vacuole in addition to the endocytic pathway. Notably, human α-synuclein expression in yeast manifests prominent defects in normal CPY and ALP trafficking consistent with derangements in ER-to-Golgi vesicular trafficking [30]. Accordingly, toxicity induced by LRRK2 and α-synuclein expression in yeast most likely occur via impairment of distinct vesicular trafficking pathways.
In order to validate the observations from this yeast model of LRRK2 toxicity and determine its wider applicability to mammalian cells, we examined the effects of human LRRK2 domain fragments and GTPase variants on neuronal viability. Expression constructs containing LRRK2 fragments identical to those employed in yeast including the GTPase domain (GTP), kinase domain (Kin) and the GTP-COR-Kin fragment as well as full-length WT or G2019S LRRK2 were individually co-transfected together with eGFP as a marker into mouse primary cortical neurons and their effects on neuronal viability were compared. A well-established assay was employed to examine the viability of eGFP-positive neurons containing LRRK2 based on neurite process length and fragmentation as a reliable indicator of neuronal viability [21],[22],[37],[38]. Using this method, LRRK2 expression was confirmed in >95% of eGFP-positive cortical neurons that were also positive for the neuronal marker, MAP2 (representative images in Figure S5A and S5C), and neuronal viability was also confirmed by TUNEL staining (representative images in Figure S5B). Expression of the GTPase domain, the GTP-COR-Kin fragment and full-length WT LRRK2 induces significant and equivalent neuronal toxicity relative to control neurons expressing eGFP alone, with a 10–20% loss of viability (Figure 5A and 5B). The kinase domain alone fails to significantly reduce neuronal viability. Full-length LRRK2 containing the common G2019S pathogenic variant serves as a positive control for toxicity and induces a ∼50% loss of neuronal viability compared to control neurons (Figure 5A and 5B), as previously reported [21]–[23],[37].
Full-length human LRRK2 was packaged into a Herpes Simplex Virus (HSV) amplicon that co-expresses eGFP to generate an HSV-WT-LRRK2/CMV-eGFP amplicon. Expression of LRRK2 by the HSV amplicon causes similar neuronal toxicity to that of full-length WT LRRK2 transiently co-transfected into neurons with eGFP (Figure 5A and 5B), indicating that transient transfection is a reliable and valid method by which to assess LRRK2-induced toxicity. Thus, truncated LRRK2 proteins containing the GTPase domain produce similar neuronal toxicity to that induced by full-length WT LRRK2 implying that the GTPase domain may underlie the toxic effects of LRRK2.
To determine and compare the effects of truncated LRRK2 GTPase variants on neuronal viability, similar experiments were conducted with the GTP-COR-Kin LRRK2 fragment containing each mutation that was previously examined in the yeast model. Expression of the GTPase-active WT, R1398L and RQ/TG variants of LRRK2 induces a significant yet equivalent level of neuronal toxicity relative to control neurons characterized by a 10–15% loss of viability (Figure 5C and 5D). Expression of the LRRK2 GTPase-inactive variants, K1347A and T1348N, enhances neuronal toxicity compared to other GTPase variants with a ∼18% loss of viability for the K1347A variant and ∼23% loss for the T1348N variant that is significantly increased relative to the WT protein (Figure 5C and 5D). Thus, GTPase variants in truncated LRRK2 induce toxicity in neurons that closely parallel their toxic effects in yeast. Collectively, these data demonstrate the validity of the yeast model for accurately predicting the detrimental effects of truncated LRRK2 variants on neuronal viability. Taken together, these data reveal that alterations in GTPase activity contribute to LRRK2-induced neuronal toxicity.
Since the LRRK2 yeast model indicates that truncated LRRK2 may function in vesicular trafficking pathways, including endocytosis, the effect of full-length human LRRK2 on endocytosis and exocytosis was monitored in primary neurons. Mouse hippocampal neurons at days in vitro (DIV) 12 were transduced with HSV-WT-LRRK2/CMV-eGFP or control virus and 48 hours later synaptic vesicle (SV) endocytosis and exocytosis were monitored by using the lipophilic fluorescent dye FM4–64. Neurons were first exposed to FM4–64 in the presence of 90 mM KCl, which depolarizes the nerve terminal and induces vesicular recycling and subsequent loading of FM4–64 by SV endocytosis. SV exocytosis was then monitored in real time by depolarizing the nerve terminals to unload the FM4–64 dye. Based on comparison of the mean fluorescence intensity values, the synaptic boutons of neurons carrying HSV-WT-LRRK2/CMV-eGFP display an approximate 1.34-fold decrease in loading of FM4–64 by endocytosis compared to the HSV-PrPUC/CMV-eGFP control (Figure 6A left panels, Figure 6B at time point ‘0’ sec, and Figure 6C: control, 133.99±5.897; WT LRRK2, 100.23±7.098). Following depolarization of the FM4–64-loaded SVs, the control boutons displayed about 99% unloading of FM4–64 after 8 mins, whereas the synaptic boutons overexpressing LRRK2 show delayed unloading with an approximate 72% decrease in FM4–64 signal (Figure 6A right panels, Figure 6B at time point ‘480’ secs, and Figure 6C: control, 0.886±0.851; LRRK2, 28.3±0.804). These data indicate that overexpression of full-length LRRK2 causes defects in both synaptic vesicle endocytosis and exocytosis in neurons consistent with the observation that overexpression of truncated LRRK2 variants in yeast perturbs vesicular trafficking pathways.
To define mechanisms underlying LRRK2-induced cytotoxicity in yeast, we performed an unbiased genome-wide genetic screen to identify yeast genes that could suppress or enhance toxicity. A similar approach has been effective at identifying modifiers of α-synuclein or mutant huntingtin toxicity in yeast [39]. We mated a haploid query strain, harboring the galactose-inducible WT LRRK2 GTP-COR-Kin fragment, to a collection of ∼4,850 viable yeast deletion mutants. Following sporulation and haploid mutant selection, we isolated deletion mutants that suppressed or enhanced LRRK2 toxicity. Of 4,850 mutants screened, we identified 2 gene deletions that enhanced LRRK2 toxicity (Figure 7A, Table 1) and 7 deletions that suppressed toxicity (Figure 7B, Table 1). Furthermore, these 7 LRRK2 toxicity suppressors also suppressed toxicity induced by the LRRK2 mutants, K1347A and T1348N, in the GTP-COR-Kin fragment (Figure 7C). This set of yeast genes that modify LRRK2 cytotoxicty function in a number of diverse pathways including transcriptional regulation (AHC1 (GenBank #CAA99213) and GCN4 (GenBank #AAA34640)), MAP kinase signaling (SLT2 (GenBank #AAB68912), small GTPase signaling (GCS1 GenBank #CAA98805) and mitochondrial function (CCE1 GenBank #AAB24906) (Table 1).
To further determine if these genetic modifiers enhance or suppress LRRK2 toxicity by modifying trafficking defects in yeast, we performed the FM4–64 assay in the two enhancer deletion mutants carrying WT GTP-COR-Kin fragment and the 7 suppressor deletion mutants carrying the most toxic LRRK2 mutant, T1348N. Interestingly, both enhancer mutants promote the endocytic trafficking defect induced by WT LRRK2 with an increase in the appearance of labeled punctate structures (Figure 7D) while the 7 suppressor mutants at least partially rescue the T1348N LRRK2-induced endocytic trafficking defect with the appearance of normal vacuolar membrane staining and a decrease in punctate structures (Figure 7E). These data suggest that the genetic modifiers can at least partially modulate vesicular trafficking pathways and genetically interact with LRRK2 to modify LRRK2-induced toxicity. Accordingly, these data suggest that vesicular trafficking defects in yeast underlie, in part, LRRK2-induced toxicity.
Here, we employ yeast cells to provide insight into the pathobiology of human LRRK2, a protein that is associated with autosomal dominant PD. A number of important conclusions can be derived from this yeast model. First, expression of LRRK2 fragments containing the GTPase domain markedly reduces the viability of yeast cells relative to other protein domains of LRRK2. The expression of full-length LRRK2 in yeast is problematic since it is highly insoluble and is sequestered into large cytoplasmic inclusions, which prevents its potential for inducing toxicity. Thus, it is only possible to develop a yeast model of LRRK2 pathobiology based upon protein domain fragments rather than the full-length protein. Second, consistent with a prominent role for the GTPase domain in mediating the toxic effects of LRRK2 in yeast, the viability of yeast cells can be modulated by alterations in GTPase activity due to several functional mutations. Notably, interfering with GTPase activity (i.e. GTP hydrolysis) but not GTP binding or kinase activity is sufficient to modify LRRK2-induced toxicity in yeast. The pathogenic mutants R1441C/G and Y1699C in full-length LRRK2 have significantly decreased GTPase activity consistent with the notion that reduced GTPase activity is toxic to cells. Importantly, however, pathogenic mutations associated with familial PD (i.e. R1441C and G2019S) do not influence the toxicity induced by truncated human LRRK2 in yeast which perhaps suggests that these mutations may only exert their deleterious effects in the context of full-length LRRK2 or in mammalian cells. Third, the expression of functional LRRK2 GTPase variants induce defects in the endocytic vesicular trafficking and autophagy pathways. Vesicular trafficking and autophagic defects closely correlate with the level of toxicity induced by each truncated GTPase variant suggesting that defects in trafficking may underlie LRRK2-induced toxicity in this model. Accordingly, genetic modifiers that suppress LRRK2 toxicity in yeast also suppress trafficking defects. Fourth, known suppressors of α-synuclein-induced cytotoxicity in yeast do not suppress LRRK2 toxicity suggesting that both proteins mediate their toxic effects through distinct trafficking pathways yet with the common outcome of impairing vesicular transport to the vacuole, the yeast equivalent of the mammalian lysosome. Thus, defects in vacuolar or lysosomal transport may commonly underlie the pathogenic effects of α-synuclein and LRRK2. Fifth, the toxic effects of truncated LRRK2 GTPase variants are similar between yeast and neuronal models of LRRK2 pathobiology and truncated or full-length LRRK2 cause similar endocytic trafficking defects in both yeast cells and neurons, respectively, suggesting that the yeast LRRK2 model is predictive of mammalian cells. Finally, a genome-wide genetic screen identified potent modifiers of LRRK2 toxicity in yeast, which may provide novel clues to the underlying mechanism of LRRK2-induced toxicity.
Neuronal toxicity induced by WT and pathogenic variants of full-length human LRRK2 critically requires intact GTP binding and kinase activity [21]–[23]. However, it has not yet been possible to distinguish, which, if any, of these activities actually mediates the downstream toxic effects of LRRK2 or whether they serve to auto-regulate an alternative function or effector domain of this protein. In yeast cells, the detrimental effects of expressing truncated LRRK2 variants are independent of kinase activity and are not influenced by two common pathogenic variants located either in the GTPase domain (i.e. R1441C) or the kinase domain (i.e. G2019S). Instead, toxicity is dependent on GTP hydrolysis activity, but not GTP binding activity. In the context of the central GTP-COR-Kin fragment of LRRK2 that is used here to explore the effects of GTPase variants, mutations that impair GDP/GTP binding and are thus GTPase-inactive promote toxicity, whereas mutations that produce a hyperactive GTPase partially reduce the toxic effects of LRRK2 (Figure 8). The lack of effect of kinase activity or pathogenic mutations on yeast toxicity induced by truncated human LRRK2, might suggest that they require the full-length protein or a mammalian cellular context to exert their effects on LRRK2-induced toxicity.
In the context of full-length LRRK2, the K1347A and T1348N mutations prevent GTP binding and are GTPase-inactive but also impair kinase activity, which partially prevents LRRK2-induced neuronal toxicity [14], [15], [19]–[23]. The RQ/TG mutation produces a Ras-like GTPase that also has impaired kinase activity owing to its increased turnover of GTP [15],[40], a feature reflected in our yeast model. The R1398L mutation also promotes GTP hydrolysis and accordingly we observe that introduction of this mutation into full-length LRRK2 produces a kinase-inactive variant (data not shown). The effects of the hyperactive GTPase mutants, RQ/TG and R1398L, on neuronal toxicity induced by full-length LRRK2 have not been defined, but they are likely to be protective due to their impairment of kinase activity and enhancement of GTPase activity. Both R1398L and RQ/TG mutants are capable of hydrolyzing GTP but their affinity for binding to GTP is reduced suggesting that they most likely predominate in a GDP-bound inactive state. It is likely that GTPase-inactive variants of LRRK2 induce greater toxicity in yeast through a novel gain-of-function mechanism by interfering with a pathway or process, or sequestering one or more proteins, critical for yeast survival or growth. A dominant-negative mechanism for LRRK2-induced toxicity is unlikely since yeast do not contain an obvious ortholog of human LRRK2. While the truncated LRRK2 protein used herein does not behave in a manner identical to full-length protein with regards to the regulation of cytotoxicity in yeast or neurons, it instead reveals a fundamental contribution of the GTPase domain and particularly GTP hydrolysis activity in mediating the toxic effects of LRRK2. A major challenge in future experiments will therefore involve dissecting the precise contribution of GTPase activity, vesicular trafficking pathways and genetic modifiers to neuronal toxicity induced by full-length LRRK2 variants.
The fact that LRRK2 kinase activity plays no role in yeast toxicity allowed us to reveal instead a major role for the GTPase domain in toxicity induced by truncated LRRK2 in both yeast and neurons. Fragments of other disease-causing gene products, such as in Huntington's disease or other poly-glutamine repeat disorders [41]–[44], TDP-43opathies [45],[46] and α-synucleinopathies [47],[48] play prominent roles in neurodegeneration due to the pathogenic generation of these truncated proteins. Interestingly, putative truncation fragments containing the LRRK2 GTPase domain have been identified in PD brains [3],[49]. In addition, E1874stop is a LRRK2 pathogenic mutation in which the protein lacks the kinase and WD40 domains [50]. Thus, understanding whether GTPase domain-containing truncated LRRK2 proteins are important for disease pathogenesis and how the GTPase domain modulates full-length LRRK2 activity are important avenues of investigation. Moreover, since the truncated GTPase domain-containing LRRK2 constructs are toxic in the absence of kinase activity, caution may be warranted by solely focusing on kinase inhibition as a therapeutic target for preventing LRRK2-induced neurodegeneration. Indeed, the GTPase-inactive K1347A mutant in the context of the full-length G2019S LRRK2 protein only partially rescues LRRK2 toxicity despite completely inhibiting kinase activity suggesting that perturbations in the GTPase domain may have deleterious consequences in the setting of full-length LRRK2 independent of kinase activity [21].
The mechanism by which truncated human LRRK2 is toxic to yeast is unclear. The GTPase domain would appear to play a key role in mediating toxicity but other protein domains may also contribute. LRRK2-induced defects in endocytic vesicular trafficking and autophagy may underlie toxicity in yeast, an observation supported by the actions of genetic modifiers of toxicity on vesicular trafficking. Consistent with the yeast LRRK2 model, full-length LRRK2 causes defects in synaptic vesicle endocytosis and exocytosis in neurons. Many other observations suggest that full-length LRRK2 may play a role in vesicular trafficking in mammalian neurons. LRRK2 is localized exclusively to a wide range of vesicular and membranous structures in neurons, including lysosomes, endosomes, multivesicular bodies, the ER, Golgi, mitochondria and microtubule transport vesicles [51],[52]. The G2019S variant promotes the formation of LRRK2-positive axonal inclusions in neurons that are membrane-bound and contain swollen lysosomes, distended mitochondria associated with vacuoles, multivesicular bodies and disrupted cytoskeletal components [24], perhaps suggestive of disruption of normal vesicular trafficking.
Consistent with our studies, a potential role for LRRK2 in endocytosis has recently been described [53]. LRRK2 interacts and co-localizes with Rab5B on synaptic vesicles. Knockdown or over-expression of LRRK2 in rodent primary neurons impairs synaptic vesicle endocytosis that can be rescued by over-expression of Rab5B [53], a GTPase involved in the early endocytic pathway from plasma membrane to early endosome. Studies in C.elegans with the human LRRK2 homolog, LRK-1, reveal a role for this protein in regulating the proper transport of synaptic vesicles to axonal regions possibly by acting at the trans-Golgi network to sort vesicles away from an alternative dendrite-specific transport mechanism [54]. Thus, in yeast it is likely that truncated LRRK2 interferes with the endocytic trafficking and autophagic pathways through functionally interacting or competing with key proteins involved in as yet unspecified steps during the transport of vesicles or their protein cargo from the plasma membrane and/or autophagosomes to the vacuole.
LRRK2-associated neurite shortening induced by the G2019S variant may be mediated at least in part by autophagy, since it is associated with the development of autophagic vacuoles and can be reversed by impairing autophagy and potentiated by activating autophagy [55]. In yeast, macroautophagy constitutes an additional pathway for vacuolar transport involving the formation and delivery of large double-membrane vesicles termed autophagosomes containing cytoplasmic constituents and organelles to the vacuole for degradation and recycling. The macroautophagy pathway is also perturbed in our yeast LRRK2 model in addition to the endocytic vesicular trafficking pathway. Consistent with our studies, a potential role for LRRK2 in the endosomal-autophagic pathway has recently been described [56]. Collectively, the observations from neuronal and yeast models tend to support a role for LRRK2 in regulating the sorting or transport of vesicles via endocytosis or autophagic pathways that possibly converge on the vacuole/lysosome (Figure 8). Further study of the biology and pathobiology of LRRK2 in regulating vacuolar/lysosomal function and dynamics may prove particularly insightful. In particular, it will be important to clarify whether derangements in endocytic and autophagic trafficking pathways critically underlie the neuronal toxicity induced by disease-associated full-length LRRK2 variants and the mechanism(s) involved in this pathologic process.
The observation that GTPase activity plays a key role in LRRK2 toxicity may prove highly useful in dissecting the molecular mechanism(s) underlying LRRK2-induced cytotoxicity and in the identification of genes or small molecules that can directly or indirectly modulate the GTPase activity of LRRK2. The relevance of such an approach would be to identify modifiers of GTPase activity that would additionally prevent kinase activation as an alternative novel strategy to inhibit the pathogenic effects of LRRK2. The key demonstration that truncated LRRK2 variants have similar effects on the viability of both yeast and neuronal cells suggests that this yeast LRRK2 model could be predictive for identifying genetic and chemical modifiers of conserved pathways, processes or proteins that are relevant for LRRK2-induced toxicity in neuronal models including human neuronal models derived from iPS cells.
Our genome-wide genetic screen to identify suppressors and/or synthetic sick or lethal interactions of LRRK2-induced toxicity in yeast identified modifiers in a number of diverse pathways including genes that are involved in transcriptional regulation, MAP kinase signaling, small GTPase signaling and mitochondrial function. These genes may play important roles in the pathobiology of LRRK2-linked PD. Notably, two of the deletion suppressors have human homologs. SLT2 has four human homologs, which are serine/threonine MAP kinases MAPK1, 3, 11 and 14 involved in the initiation of translation, meiosis, mitosis, and postmitotic functions in differentiated cells. In addition they mediate their response via activation by environmental stress, pro-inflammatory cytokines and lipopolysaccharide by phosphorylating a number of substrates. The human homolog of GCS1 is ADP-ribosylation factor GTPase activating protein 1 (ARFGAP1) which plays a role in membrane trafficking and/or vesicle transport. These deletion suppressors may prove to be attractive drug targets and they may provide important insight into the function of LRRK2.
In summary, our results provide evidence that the GTPase domain may contribute to LRRK2-induced toxicity, with enhanced GTP hydrolysis leading to reduced LRRK2 toxicity and impaired GTP hydrolysis leading to enhanced LRRK2 toxicity. In addition, our identification of genetic modifiers of LRRK2-induced toxicity in yeast provides important clues to proteins or pathways that may play key roles in mediating LRRK2-induced toxicity in higher organisms.
All procedures involving animals were approved by and conformed to the guidelines of the Institutional Animal Care Committee of Johns Hopkins University.
Yeast haploid strain BY4741 (MATa, his3Δ1, leu2Δ0, met15Δ0, ura3Δ0) obtained from Open Biosystems (Huntsville, AL) was used throughout this study. For the yeast genetic screen, the LRRK2 query strain was constructed in Y7092 (MATα, can1Δ::STE2pr-Sp_his5, lyp1Δ, his3Δ1, leu2Δ0, ura3Δ0, met15Δ0). Yeast manipulations were performed and media were prepared using standard procedures. Transformations of yeast were performed using a standard high efficiency lithium acetate procedure [57]. Yeast cells carrying galactose-inducible expression constructs were routinely grown in YPD or synthetic complete media lacking uracil (SC-URA) containing glucose (2% dextrose) to repress the GAL1 promoter. Yeast cells were pre-grown in SC-URA containing 2% raffinose (no repression of the GAL1 promoter) prior to growth in medium containing 2% galactose (to induce the GAL1 promoter), to allow rapid, synchronous induction of expression. Yeast cells that were co-transformed with two galactose-inducible constructs (with URA3 or LEU2 markers) were grown in SC-URA/-LEU to select for both plasmids.
Human LRRK2 fragment cDNAs were amplified from a pcDNA3.1-LRRK2-Myc-His vector [18] by PCR with primer pairs specific for the different domains of LRRK2 (refer to Figure 1A) with incorporation of an optimal yeast Kozak sequence (AAAAATGTCT) surrounding an ATG start codon (underlined). PCR products were first cloned into the pCR2.1-TOPO TA cloning vector (Invitrogen, Carlsbad, CA) before subcloning into the GAL1 promoter-based yeast expression vector pYES2/CT (2 µ ori, URA3; Invitrogen) or p416GAL (CEN ori, URA3; kindly provided by Martin Funk [58]) containing a C-terminal V5 tag, or into mammalian expression vector pcDNA3.1-Myc-His (Invitrogen) via BamHI and XhoI restriction sites. Missense mutations were introduced into the GTP-COR-Kin fragment of LRRK2 by PCR-mediated, site-directed mutagenesis, using the QuickChange XL kit (Stratagene), followed by sequencing of the entire cDNA to confirm their correct incorporation. Candidate genes (YPT1, YKT6 and HSP31) were amplified from yeast genomic DNA by PCR to also introduce a C-terminal V5 tag and stop codon, and resulting cDNAs were cloned into the GAL1 promoter-based yeast expression vector p425GAL (2 µ ori, LEU2, kindly provided by Martin Funk [58]. All cDNAs were subjected to DNA sequencing to confirm their integrity.
Mouse monoclonal antibodies to yeast 3-phosphoglycerate kinase (PGK, clone 22C5), anti-V5 and anti-V5-HRP were obtained from Invitrogen. Mouse monoclonal anti-myc antibody (clone 9E10) was purchased from Roche Biochemicals. Rabbit polyclonal anti-GFP antibody (NB 600-303) was obtained from Novus Biologicals. HRP-linked anti-rabbit or anti-mouse IgG antibodies were obtained from Jackson ImmunoResearch Labs (West Grove, PA). AlexaFluor-488 anti-mouse IgG and AlexaFluor-594 anti-rabbit IgG antibodies were from Molecular Probes. Human LRRK2-specific antibody JH5517 has been described previously [18],[59].
Yeast cells carrying galactose-inducible LRRK2 constructs were grown and induced as described for spotting experiments. Total RNA was isolated from yeast cells by hot phenol extraction [60] and further purified using the Qiagen RNEasy Mini kit (Qiagen). Total RNA concentrations were determined with a Nanodrop spectrophotometer (Nanodrop Technologies) prior to RT-PCR. cDNAs were generated from total RNA using the OneStep RT-PCR kit (Qiagen) and oligo-d(T). PCR was conducted on equal quantities of mRNA-derived cDNAs for 25 cycles with LRRK2-specific primers located within the kinase domain (Forward: 5′-CCAGATCAACCAAGGCTCAC-3′, Reverse: 5′-CCTGCTGTTGTGATGTGTAG-3′) or yeast actin (ACT1) primers (Forward: 5′-TCGATTTGGCCGGTAGAGATT-3′, Reverse: 5′-AAGATGGAGCCAAAGCGGTGATT-3′) as a loading control.
Yeast cells carrying galactose-inducible LRRK2 constructs were grown and induced as described for spotting experiments. Total proteins were extracted from yeast by a standard method using glass bead lysis. Briefly, yeast cells were pelleted and lysed in 1 ml lysis buffer (1 X PBS, pH 7.4, 1% NP-40, 1 x phosphatase inhibitor cocktail 1 and 2 [Sigma-Aldrich], 1 x Complete mini protease inhibitor cocktail [Roche]) by vigorous shaking with glass beads at 4°C for 15 min and lysates were clarified by centrifugation at 17,500×g for 10 min at 4°C. Supernatants were incubated with 50 µl γ-aminohexyl-GTP-sepharose bead suspension (Jena Bioscience, Jena, Germany) by rotating at 4°C for 2 hr. The sepharose beads were sequentially pelleted and washed twice in wash buffer (1 X PBS, pH 7.4, 1% Triton X-100) and twice with PBS alone. GTP-bound proteins were eluted into 50 µl Laemmli sample buffer (BioRad) containing 5% 2-mercaptoethanol by heating for 10 min at 95°C. GTP-bound proteins or input controls (0.1% total lysate) were resolved by SDS–PAGE and subjected to Western blot analysis with anti-V5 antibody. Bands were visualized by enhanced chemiluminescence (Amersham). Quantification of protein expression was performed using densitometry analysis software (AlphaImager, Alpha Innotech Corp.).
GTP hydrolysis activity was measured by monitoring the release of free γ-phosphate (Pi) from GTP. Briefly, total proteins were prepared from yeast cells carrying galactose-inducible LRRK2 constructs as described for GTP binding assays. Soluble lysates were subjected to immunoprecipitation with anti-V5 antibody (1 µg) pre-incubated with 50 µl Protein G Dynabeads (Invitrogen) by rotating at 4°C overnight. Dynabeads were stringently washed 5x with lysis buffer before being subjected to GTPase activity assay in 96-well plates using the colorimetric GTPase assay kit (Innova Biosciences, Cambridge, UK) as per manufacturers instructions to measure the concentration of free Pi with absorbance measured at 590–660 nm. LRRK2 immunoprecipitates (anti-V5) were also analyzed by Western blot analysis with anti-V5 antibody to quantify the input levels of each LRRK2 variant for normalization purposes. Densitometric analysis was conducted on protein bands using appropriate software (AlphaImager, Alpha Innotech Corp.). A similar procedure was employed for myc-tagged full-length human LRRK2 variants derived from HEK-293T cells to measure LRRK2 GTPase activity.
The HSV amplicon platform was utilized to generate HSV-LRRK2 expression vectors containing full-length human LRRK2 [61].
TEM was performed on yeast cells expressing truncated LRRK2 variants as previously described [45],[62] at the Integrated Imaging Center, Johns Hopkins University.
Yeast cells carrying galactose-inducible LRRK2 constructs were grown and induced as described for spotting experiments. Following galactose induction for 6 hrs, 1 ml of culture was harvested by brief centrifugation, resuspended in SC-URA media containing galactose and 40 µM FM4–64 red fluorescent dye (Molecular Probes) and incubated at 30°C for 20 minutes to allow dye internalization by endocytosis. Cells were washed once in SC-URA media containing galactose before being dispersed and mounted onto microscope slides. Imaging of red fluorescence was conducted on a Zeiss LSM510 live confocal system.
Mouse primary hippocampal neurons (E15–16) were transduced by HSV-WT-LRRK2/CMV-eGFP and HSVPrPUC/CMV-eGFP virus at DIV 12. After 48 hour transduction, cells were mounted in a laminar-flow perfusion chamber on the stage of a custom-built laser scanning confocal microscope using a calcium containing buffer (Solution B: 119 mM NaCl, 2.5 mM KCL, 4 mM MgCl2, 30 mM Glucose, 25 mM HEPES, 2 mM CaCl2). After gently removing Solution B cells were then continuously perfused with Solution A (Solution B without CaCl2). The first stimulus was then applied with FM dye containing Solution D (90 mM KCl, 29 mM NaCl, 2 mM CaCl2, 2 mM MgCl2, 30 mM Glucose, 25 mM HEPES and 15 µM FM4–64, Molecular Probes) for 2 min. This step leads to presynaptic release, vesicle fusion and dye incorporation by synaptic vesicle endocytosis. Cells were then washed by perfusion with Solution A for up to 10 min to minimize background staining. After gently aspirating Solution A, Solution C (Solution D without FM dye) is applied to cause release of the FM dye by synaptic vesicle exocytosis. Images were acquired every 10 sec with a CCD camera. The fluorescence intensity of manually designated pre-synaptic regions was quantified.
Primary cortical neuronal cultures were prepared and transiently transfected with LRRK2 or eGFP expression constructs as described previously [22],[37]. Briefly, cortices were dissected from embryonic day 15–16 fetal mice (CD1 strain), dissociated by a 12 min digestion in TrypLE (Invitrogen), and neurons were seeded into 24-well plates coated with poly-L-ornithine. Neurons were routinely maintained in Neurobasal media (Invitrogen) containing 2 mM L-glutamine and 2% B27 supplement at 37°C in a 7% CO2 humidified incubator. Glial cell growth was inhibited by addition of 5-fluoro-20-deoxyuridine (5F2DU, 30 µM, Sigma) to the media on days in vitro (DIV) 4. Media was replaced once every third day. At DIV 10, neurons represented >90% of total cells in the culture. To assess LRRK2-induced toxicity, neurons at DIV 10 were transiently co-transfected with LRRK2 and eGFP expression constructs at a molar ratio 10∶1, respectively, using Lipofectamine 2000 reagent (Invitrogen) according to the manufacturer recommendations. At 48 hrs post-transfection (DIV 12), live fluorescent images were collected on a Zeiss Automatic stage microscope with Axiovision 6.0 software. Neurons with obvious neurite process and/or nuclear fragmentation were counted as non-viable cells by investigators blinded to the identity of the experiment. For each independent experiment, the percent viability of eGFP-positive neurons (n = 200) was determined and data are presented as a percent of control neurons transfected with eGFP alone. Neurons were subsequently fixed with 4% paraformaldehyde and immunocytochemistry was conducted with anti-myc (Roche) and anti-GFP (Novus Biologicals) antibodies and appropriate fluorescent secondary antibodies. LRRK2 expression was confirmed in >95% of eGFP-positive neurons (Figure S5A).
The above transfected neurons were fixed in 4% paraformaldehyde (PFA) after 48 hrs transfection. TUNEL staining was performed using the In Situ Cell Death Detection Kit (Roche) as per the manufacturer's instructions.
Yeast cells carrying galactose-inducible LRRK2 constructs were grown and induced as described for spotting experiments. Following galactose induction for 6 hrs, 1 ml culture was harvested by brief centrifugation, and fixed in 4% formaldehyde/PBS for 1 hr. Cell walls were digested by incubation with Zymolyase 20T solution (ICN Biochemicals), as recommended. Following permeabilization, cells were gently washed twice in KS solution (100 mM potassium phosphate pH 7.0, 1 M sorbitol), and then resuspended in KS solution. Immunostaining with mouse monoclonal anti-V5 antibody (Invitrogen) and AlexaFluor-488 anti-mouse IgG (Molecular Probes) was conducted as previously described [63] Cells were dispersed onto microscope slides and mounted using Vectashield mounting medium containing DAPI (Vector Laboratories) for nuclear visualization. Fluorescent images were collected on a Zeiss Automatic stage microscope with Axiovision 6.0 software.
The yeast LRRK2 toxicity modifier screen was performed using synthetic genetic array (SGA) analysis [64] as previously described [65]. We used a Singer RoToR HAD yeast pinning robot for manipulating yeast colonies at high density. A MATα yeast haploid query strain, Y7092, carrying WT LRRK2 GTP-COR-Kin fragment was mated with a haploid yeast gene deletion collection of 4850 viable mutants, sporulated and then underwent selection for haploid mutants that also harbored the LRRK2 plasmid on solid media containing G418 and lacking uracil. Haploid deletion mutants that also carried the LRRK2 plasmid were identified on selectable media containing glucose and then the expression of LRRK2 was induced by growth on galactose media. After comparing colony sizes on galactose plates to those on glucose plates and normalizing for differences in the growth of deletion mutants between carbon sources, genes that suppressed or enhanced LRRK2 toxicity were identified. Initial hits from the screen were independently verified by fresh transformations and spotting assays.
|
10.1371/journal.pcbi.1004499 | The Elementary Operations of Human Vision Are Not Reducible to Template-Matching | It is generally acknowledged that biological vision presents nonlinear characteristics, yet linear filtering accounts of visual processing are ubiquitous. The template-matching operation implemented by the linear-nonlinear cascade (linear filter followed by static nonlinearity) is the most widely adopted computational tool in systems neuroscience. This simple model achieves remarkable explanatory power while retaining analytical tractability, potentially extending its reach to a wide range of systems and levels in sensory processing. The extent of its applicability to human behaviour, however, remains unclear. Because sensory stimuli possess multiple attributes (e.g. position, orientation, size), the issue of applicability may be asked by considering each attribute one at a time in relation to a family of linear-nonlinear models, or by considering all attributes collectively in relation to a specified implementation of the linear-nonlinear cascade. We demonstrate that human visual processing can operate under conditions that are indistinguishable from linear-nonlinear transduction with respect to substantially different stimulus attributes of a uniquely specified target signal with associated behavioural task. However, no specific implementation of a linear-nonlinear cascade is able to account for the entire collection of results across attributes; a satisfactory account at this level requires the introduction of a small gain-control circuit, resulting in a model that no longer belongs to the linear-nonlinear family. Our results inform and constrain efforts at obtaining and interpreting comprehensive characterizations of the human sensory process by demonstrating its inescapably nonlinear nature, even under conditions that have been painstakingly fine-tuned to facilitate template-matching behaviour and to produce results that, at some level of inspection, do conform to linear filtering predictions. They also suggest that compliance with linear transduction may be the targeted outcome of carefully crafted nonlinear circuits, rather than default behaviour exhibited by basic components.
| Any attempt to model human vision must first ask: can it be approximated by a process that linearly matches the visual stimulus with an internal template? We often take this approximation for granted without properly checking its validity. Even if we assume that the approximation is valid under specific conditions, does this mean the system operates template matching across the board? We would not know exactly in what sense and to what extent the approximation may be viable. Our results address both issues. We find that template matchers are locally applicable in relation to a wide range of conditions, providing much-needed justification for several relevant computational tools. We also find, however, that there is no sense in which the system is globally a linear template: it remains inescapably nonlinear. Our findings suggest that linear transduction is not cost-free: it is not a default building block that is used for constructing expensive nonlinear processes. Rather, linear sensory representations arise from carefully constructed nonlinear processes that strike a balanced act between the necessity to retain other important computations, and the desirability of transducing and representing the visual world on a linear scale.
| Animals constantly submit environmental signals to neural operations designed to extract useful information for guiding behaviour. Whether their sensory apparatus is considered in its entirety as a behavioural machine or in relation to hardware components like individual nerve cells, it can be described as an input-output transformation that maps external stimuli onto neural representations. The simplest way to characterize the operation of such a sensory device is to assign a set of weights to different elements of the incoming stimulation, then sum across all elements, and finally convert this weighted sum into a number compatible with the scale and units of the output variable [1, 2]. For input stimulus s, this simple operation can be written as g(⟨s,w⟩) (w is the weighting function, ⟨, ⟩ inner product, and g a static nonlinearity). To provide an example, w may be the receptive field of a simple cell, and g the nonlinearity that maps membrane voltage onto average spike rate [3]. For another example, more relevant to the present study, we can think of w as the perceptual impact associated with different portions of a visual display presented to a human observer, and g the decisional transducer that maps aggregate perceptual impact onto a binary decision of the kind ‘I saw the stimulus’ or ‘I did not see it’ [4].
The above linear-nonlinear cascade model has been applied to innumerable phenomena in neuroscience [5, 6], to the extent that it would be impossible to summarize them here. Particularly when referring to perceptual processes, it is often termed ‘template-matching’ [7] to indicate that an internal template (the filter) is matched against the incoming stimulus (via linear weighting) before a decision is made as to whether the stimulus does or does not contain the template signal [8, 9]. We will use the terms ‘linear-nonlinear’ (abbreviated LN) and ‘template matcher’ interchangeably. We will also occasionally refer to this process as ‘linear filtering’ or ‘linear transduction’ and contrapose it to a ‘nonlinear’ process, with the understanding that in these instances we are specifically referring to the processing stage that precedes the decisional nonlinearity (g in the previous paragraph). The latter element is an integral part of all psychophysical models (and those considered here are no exception), but can be largely bypassed to access the preceding layers using the methods employed in this study [4, 10].
Qualitative thinking about sensory processing almost invariably refers back to the LN model [11], not least because its explanatory scope can be greatly extended by adopting arbitrary descriptors for s, effectively remapping the stimulus onto a space that is available for inner-product treatment [12]. The well-known dipper function for contrast discrimination, for example, is often accounted for by a specific choice of g for a human observer modelled as a LN cascade [13], and the link to neural activity can also be inferred via this simple framework [14]. Apparently counter-intuitive phenomena such as stochastic resonance can be accommodated via the LN model [15]. Furthermore, certain methodological approaches (notably reverse correlation) often rely on the assumption that the system of interest is well approximated by the LN operator [4, 16, 17].
Notwithstanding such widespread applicability, there are well-known instances when the LN model is unable to provide a satisfactory account of relevant phenomena. The operation of a complex cell, for example, cannot be described by the LN cascade acting directly on the stimulus image [18]. Several neural systems exhibit pronounced gain control properties [19], and these too fall outside the explanatory reach of barebone LN operators. In human vision, detection under uncertainty represents a classic example of the inapplicability of simple template-matching models belonging to the LN family [20, 21]. Adaptive phenomena, e.g. learning-mediated plasticity, can only be partially approximated by LN descriptors [10]. It is therefore uncontroversial that LN models are sometimes inadequate.
The critical issue is to recognize when they are adequate or not and, whenever they appear adequate for a specific application of the system under interest, how far their applicability can be generalized to other applications of that same system. We can illustrate this issue with the following example. In experiment 1, we characterize the response of a neuron, or a whole observer, to visual orientation by selectively manipulating the orientation content of a simple visual stimulus (e.g. a textured object; see [22] for an example from literature). In experiment 2, we characterize the response of the same system to pattern size by selectively manipulating the spatial frequency (SF) content of the same stimulus (see [23] for examples). First, we ask whether the manner in which the system operates under the conditions of experiment 1 can be approximated by the LN operator applied to the input stimulus defined with respect to orientation content: s is a vector specifying orientation energy in the stimulus for different orientations, and w is the orientation tuning function of the system; is g(⟨s,w⟩) adequate? We can ask the same question with reference to experiment 2, except s is now the vector specifying stimulus energy across SF, and w is the SF tuning function of the system.
An altogether different question is to ask whether the operation of the system with relation to both orientation and SF can be accounted for by the same LN cascade [24, 25]. For this purpose, the visual stimulus must be projected onto a space s that encompasses both orientation and SF, because the LN cascade must be applied to one common input space; at the same time, it must be able to make predictions for the two different spaces probed in the two separate experiments. The natural space of choice in this case is that of the image itself, i.e. the 2D pixel array detailing stimulus intensity at each spatial location on the display. If we call this image 2Ds, the question is whether we can identify linear filter 2Dw and nonlinear transducer g so that g ( ⟨ 2D s , 2D w ⟩ ) will capture the results of experiments 1 and 2. As we demonstrate in this study, a positive answer to the question posed in the previous paragraph (i.e. both experiments falling within the explanatory power of the LN family) does not guarantee a positive answer to this latter question: the system may appear to operate in the manner of the LN cascade with relation to a number of different probes defined within substantially different spaces (orientation, SF, 2D space), yet its behaviour may not be collectively captured by a single LN cascade. Our results have important implications for the applicability of LN cascades to visual perception, and establish some general notions/tools relating to both the potential and limitation of this modelling family for capturing human sensory processing.
Ethics approval was obtained from the College Ethics Review Board (CERB) at Aberdeen University (http://www.abdn.ac.uk/clsm/working-here/cerb.php). All participants gave written informed consent.
Stimuli lasted 80 ms and consisted of 3 regions (∼3×3 deg each): a central ‘probe’ region at fixation (the fixation marker consisted of a dark pixel in the centre measuring ∼3×3 arcmin and never disappeared); above and below it, two identical ‘reference’ regions containing the template (see S1 Video). The template signal consisted of a cosine-phase (peaking at centre) vertical Gabor wavelet (standard deviation (SD) of Gaussian envelope 0.5 deg, spatial frequency 1 cycles/deg), and was presented at 17% contrast within the reference regions (background luminance 30 cd/m2). On each trial, observers saw two instances of the stimulus separated by a 500-ms gap. Reference regions were identical on both instances (and across all trials), thus providing no useful information for performing the task; their purpose was to remind observers of the target signal shape, so as to facilitate a template-matching strategy [21, 26]. The probe region contained target signal plus noise mask on one instance, and non-target signal plus noise mask on the other instance. Observers were asked to select the instance (first or second) that contained the target signal by pressing one of two buttons, after which they received immediate trial-by-trial feedback (correct/incorrect). The target signal was simply the template Gabor wavelet described above (see also Fig 1A), presented at 8% (alternatively 4%) contrast in the detection (alternatively discrimination) task. The non-target signal was blank for the detection task (Fig 1F), and a horizontal variant of the target signal for the discrimination task (see icons to the left of Fig 2E). Except for taking on a different orientation, the non-target signal in the discrimination task was identical to the target-signal. Data for the two tasks were collected in separate blocks of 100 trials each. We also collected separate data for a ‘symmetric’ variant of the discrimination task. In this additional experiment, two identical reference regions containing non-target templates were presented to the left and to the right of the central probe region.
The noise mask could be 1 of 4 different types in the detection task, and 1 of 2 different types in the discrimination task (thus explaining why the detection data in Figs 2B and 2D and 3C and 3E are not matched by equivalent data for discrimination). Mask type was randomly selected on every trial with equal probability for each mask. In the detection task, the noise mask could be ‘2D’ (Fig 1B and 1G): each pixel (within a 65×65 array) was separately assigned a random luminance value from a zero-mean Gaussian distribution with SD ∼16% contrast (we use the ∼ symbol because this value was tailored to each observer to target threshold performance of d′ ∼1, see abscissa values in Fig 4C); ‘1D’ (Fig 1C and 1H): each column of pixels spanning the probe region was separately assigned a random value from a zero-mean Gaussian distribution with SD ∼9% contrast, and the vertical profile of each column was modulated by the envelope (Gaussian window) of the Gabor target signal; ‘Θ’ (Fig 1D and 1I): a set of 12 Gabor wavelets spanning the 0-π orientation range (Fig 1K), all identical to the target signal except for rotation, were assigned a random contrast value from a Gaussian distribution with mean ∼3% (alternatively ∼5%) contrast and SD ∼0.7% (alternatively ∼1.2%) contrast in the detection (alternatively discrimination) task; ‘SF’ (Fig 1E and J): a set of 12 Gabor wavelets ranging in spatial frequency (SF) from 0.2 to 3.5 cycles/deg in logarithmic steps (Fig 1L), all identical to the target signal except for SF, were assigned a random contrast value from a Gaussian distribution with mean ∼3% contrast and SD ∼0.7% contrast. Noise masks were specified as detailed above so that each mask type was associated with a non-zero probability of realizing the target signal. Because only 2D and Θ masks present a non-zero probability of realizing the non-target signal in the discrimination task, only these two masks were adopted for the discrimination condition. Furthermore, due to their vertical characteristic and the limited contrast range afforded by a combination of design and hardware constraints, 1D and SF noise probes did not effectively mask the horizontal non-target signal and introduced spurious cues (e.g. cross-oriented regions) for performing the task. 2D/1D noise was fully orthogonal (each pixel was modulated independently and did not overlap with other pixels); Θ/SF noise was not fully orthogonal because the underlying wavelets (Fig 1K–1L) were not themselves orthogonal except for specific instances. In relation to the logic of this study (see Results) orthogonality was important only between vertical and horizontal components of orientation noise (indicated by peak and trough respectively in Fig 2G–2H); these two components were very nearly orthogonal (within hardware precision). All aspects of the study were validated via explicit implementation of fully specified computational models (see below).
We tested 10 observers in the detection task (∼3.3K trials per noise type per observer, total of ∼130K trials); 5 of those 10 observers also participated in the discrimination task (∼4.4K trials per noise type per observer, total of ∼44K trials), and 1 of those 10 observers also participated in the symmetric variant of the discrimination task together with an additional 5 observers who did not belong to the original pool of 10 observers (∼4K trials per noise type per observer, total of ∼49K trials). All observers were naive to the purpose and methodology of the study; they were paid 7GBP/9EUR per hour for their participation. The total number of trials collected for this study (single-pass and double-pass, see below for description of the latter type) was 269700.
Following completion of data collection and acquisition of observer responses, we can classify each stimulus type z [ q , r ] as being of type type presented in the target-present (q = 1) or target-absent (q = 0) interval on a specific trial, to which observers responded correctly (r = 1) or incorrectly (r = 0). It is a matrix of dimension 65×65 for type 2D, a 65-element vector for type 1D, a 12-element vector for type Θ and SF. It was constructed by summing a signal s to a noisy sample n: type z [ q , r ] = type s [ q ] + type n [ q , r ]. s and n were specified as detailed above. When projected onto 2D space s is the same across type (see Fig 1), however its vector representation with respect to each type differs: it is the Gabor wavelet specified above for type 2D, a horizontal slice through said wavelet for type 1D (green trace in Fig 2B), a non-zero entry for the 7th element of a 12-element vector for type Θ and SF. The first-order target-present PF (i.e. computed only from noise fields containing the target) was type p 1 [ 1 ] = avg ( type n [ 1 , 1 ] ) − avg ( type n [ 1 , 0 ] ) where avg ( . ) indicates average across the subset of trials indexed by the assigned type and [q, r] values [4]; the target-absent PF was type p 1 [ 0 ] = avg ( type n [ 0 , 0 ] ) − avg ( type n [ 0 , 1 ] ). The full PF was simply the sum of target-present and target-absent PF’s: type p 1 = type p 1 [ 1 ] + type p 1 [ 0 ] [29]. The second-order PF is similarly computed as typep 2 = cov ( type n [ 1 , 1 ] ) + cov ( type n [ 0 , 0 ] ) − cov ( type n [ 1 , 0 ] ) − cov ( type n [ 0 , 1 ] ) where cov ( . ) indicates covariance across trials of the specified classification [10].
Because we found an inevitable degree of variability across observers, it is difficult to draw conclusions from simply inspecting individual PF’s. We therefore performed additional analyses that captured relevant aspects of filter structure, and quantified each aspect using a single value (scalar metric) for each PF. This approach made it then possible to perform simple population statistics and confirm or reject specific hypotheses (against unambiguously defined null benchmarks) about the overall shape of the filters. Our conclusions are therefore based on individual observer data, not on the aggregate observer; aggregate descriptors (e.g. Fig 2A–2E and 2G) are only presented for visualization purposes. Some previous studies using classified noise have relied on qualitative inspection of aggregate data, but this approach is inadequate to draw robust conclusions primarily for two reasons: 1) there is no generally accepted procedure for generating an average PF from individual images for different observers [30]; 2) we have shown in previous work that effects observed via qualitative inspection of aggregate filters may not survive quantitative inspection using metric analysis, and vice versa [22].
We performed a series of additional experiments specifically designed to measure human-human agreement [39]. During these experimental blocks, observers saw the same stimuli presented under typical data collection conditions, with the only difference that the second half of each 100-trial block (from trial #51 to trial #100) consisted of a repetition of the first half (from trial #1 to trial #50) in randomly permuted order. Human-human agreement (plotted on the x axis in Fig 4A) is simply the % of repeated trials on which observers gave the same response. By adopting a minimal signal detection theory (SDT) framework, internal noise (plotted on the y axis in Fig 4C) can be estimated (in units of external noise) from human-human agreement and the % of correct-response trials. Details of this routine procedure have been extensively documented in previous publications [27, 39, 40]. Human-model agreement (plotted on the y axis in Fig 4B) is the % of trials on which the human response matches the response generated by a computational model to the same stimulus set (see below for details of models implemented here). For a given value of human-human agreement x in a 2AFC task, upper and lower bounds on the maximum achievable human-model agreement are given by ( 1 + 2 x − 1 ) / 2 and x itself [41]; the corresponding region is indicated by green shading in Fig 4B. Human-model agreement may exceed chance for models that are decoupled from the trial-by-trial perturbation delivered by the external noise source; to identify instances where this may be the case, we have developed a ‘decoupled’ baseline (see S1 Text). We collected double-pass data from 9 of the 10 observers in the detection task (∼650 trials per noise type per observer, total of ∼23K trials), 2 of the 5 observers in the discrimination task (∼1050 trials per noise type per observer, total of ∼4.2K trials), and all 6 observers in the symmetric variant of the discrimination task (∼1650 trials per noise type per observer, total of ∼20K trials). The first half/pass of this dataset was combined with the original dataset for the purpose of PF estimation and associated analysis.
We asked observers to detect the most common target stimulus in contemporary vision science [43]: the Gabor wavelet (Fig 1A). On each trial, one interval contained this target signal, while the other interval did not (Fig 1F); observers were asked to select the target interval. We then added four different types of visual noise to both target and non-target stimuli: 2D pixel noise (Fig 1B and 1G), 1D ‘line’ noise (Fig 1C and 1H), orientation (Fig 1D and 1I) and spatial frequency (SF) noise (Fig 1E and 1J). We applied psychophysical reverse correlation [10, 17] to retrieve the perceptual filters (PF) associated with each noise probe separately (see Methods). The PF can be thought of as the psychophysical equivalent of the physiological receptive field [11, 41]: it provides an overall picture of the weighting function applied by the observer to the incoming stimulus for the purpose of identifying the assigned target signal [4, 17]. This description is useful for intuitive purposes, but is inaccurate and possibly misleading upon closer inspection due to important differences between the two processes instantiated by single neurons on the one hand, and human observers on the other [10, 41].
The different noise probes were randomly mixed within the same block, so that observers did not know which noise type would appear on the next trial (see S1 Video); furthermore, their task was identical throughout all blocks, regardless of the noise type applied on any given trial. Data from different noise types therefore enable different vantage points on the same underlying elementary visual operation (extraction of a localized oriented wavelet). As expected, the PF’s associated with different noise types resemble the target signal: the 2D spatial PF presents a Gabor-like modulation not dissimilar from the target (compare Fig 2A with Fig 1A), the 1D spatial PF takes on a Mexican-hat shape that largely overlaps with a horizontal slice through the target (compare black data with green trace in Fig 2B), the orientation-tuned PF peaks at target orientation (indicated by green line in Fig 2C), and the SF-tuned PF peaks at target SF (indicated by green line in Fig 2D).
In a series of additional experiments, we determined that similar results were obtained for a discrimination variant of the same task where the non-target signal was orthogonal to the target signal (icons to the left of Fig 2E). The PF’s associated with the discrimination task were virtually identical to those returned by the detection task (compare E with A and G with C in Fig 2; only 2D and Θ masks are applicable to this task, see Methods), contrary to the ideal observer prediction that the 2D PF should be an image of the target minus the non-target signal [1]: the estimated PF for discrimination (Fig 2E) appears to contain exclusively vertically oriented structure (across observers and detection/discrimination tasks, PF match with target (vertically-oriented) on y axis in Fig 2F is significant at p < 10−3 (Bonferroni-corrected for multiple comparison) by two-tailed Wilcoxon signed rank (WSR) test for match>0, while match with rotated target (horizontally oriented) on x axis is not significant at p>0.05). A feature of specific interest for the purposes of later modelling efforts is that orientation-tuned PF’s (Fig 2C and 2G) present clear troughs at the non-target orientation (±π/2 on x axis) of magnitude comparable to their peaks (across observers/tasks, peak/trough amplitudes on y/x axes in Fig 2H modulate significantly at p < 10−3 (same test as above) and are not different (except for sign) at p>0.05 by paired two-tailed WSR test). This result is not trivially expected as other outcomes are possible (see [22] for examples).
The lack of any discernible difference between discrimination and detection experiments indicates that the same mechanism supports both operations, prompting us to seek a single model able to account for the entire dataset in Fig 2.
We carried out 4 tests, based on established literature [10, 34, 36, 38, 39, 41, 44], designed to gauge the applicability of LN transduction. They converge to indicate that, for any given representation of the visual stimulus, the LN framework provides an adequate description of the manner in which the human sensory system operates with respect to that representation.
The first test capitalizes on the prediction by the LN model that filter estimates returned by noise fields associated with the target stimulus (target-present) must match estimates returned by target-absent noise fields [21, 31–38]. We observed virtually no difference between the two estimates (compare left versus right surface plots in Fig 3B and 3F and blue versus red traces in C-E,H), consistent with the prediction of the LN model (see below for individual observer analysis and quantitative corroboration of the above-detailed qualitative observations).
An additional test relies on nonlinear (second-order) operators that capture system properties not conveyed by first-order estimates [33] (see Methods); the LN model predicts structureless second-order perceptual filters [10]. Fig 3G and 3I plot an index of such second-order nonlinear structure (y axis) versus a similar index of structural differences between target-present and target-absent PF’s (relating to the ‘first-order nonlinear’ test described in the previous paragraph) across observers (different data points); neither was significantly different from 0 (p>0.05 whether corrected or not for multiple comparison, one-tailed WSR test; see also overall black/gray distributions in Fig 3A, alongside orange distribution of same structural index applied to overall first-order PF’s to demonstrate that this index can adequately expose filter structure when present). Although the two tests are not equivalent in that second-order nonlinearities do not necessarily impact target-present filter estimates [10], they probe related aspects of the underlying process [26]; they may therefore be expected to correlate in the event of departures from template matching (we provide one such example from data in Fig 8I). Contrary to this expectation, there was no detectable correlation (p>0.05, robust correlation toolbox [45]) between the two tests (see lack of substantial tilt for gray ovals in Fig 3G and 3I), further supporting the notion that if any departure from template matching was present in our data, it was too small to measure.
A third test compares human absolute efficiency (plotted on y axis in Fig 4A) against corresponding predictions for LN models incorporating the empirically estimated PF [44]. There was reasonable agreement between measurements and predictions (points scatter around diagonal unity line), however we report an appreciable tendency for measured efficiency values to exceed corresponding predictions (this effect is visually demonstrated by the tendency for data points to fall above the unity line in Fig 4A; see also histogram within inset). More specifically, when different noise conditions are tested separately, only the 1D condition approaches statistical significance (two-tailed paired WSR test for measured versus predicted estimates returns p < 0.005 but this value does not survive Bonferroni correction for multiple comparison; the remaining three noise conditions return p>0.05 whether corrected or not). When tested collectively across conditions and tasks (all data points in Fig 4A collapsed onto one dataset), measured values exceed predicted values at p < 10−3 (two-tailed paired WSR test, survives correction for multiple comparison). Similar deviations have been previously documented for a variety of stimuli and discrimination tasks [44]. Our current understanding of relevant phenomena does not allow us to confidently identify the source of this small discrepancy, and it should be noted that the efficiency predictions are based on specific assumptions about the nature of the internal noise source [44]. The additional test detailed below makes no such assumptions.
The last test compares the output generated by the LN model on specific trials against the human output on those same specific trials [46] (this comparison was also independently assessed with respect to decoupled baseline, see Methods and symbol size in Fig 4B). The resulting human-model agreement (proportion of trials on which the two outputs match) is plotted in Fig 4B (y axis) against the proportion of trials on which human observers replicate their own response for two passes of the same stimulus (human-human agreement, see Methods). Collectively, estimates fall significantly above the lower bound of the maximum predictability region [41] (green shaded area) on a targeted comparison (two-tailed paired WSR test for y>x returns p <10−5), they do not differ from the midpoint between upper and lower bound (indicated by the green dotted line in Fig 4B) at p = 0.26, but they are smaller than the upper bound (p < 10−8). These findings indicate that the level of trial-by-trial predictability achieved by the LN model is compatible with optimality, although they do not guarantee this result: it remains possible that the internal noise process operating within observers is such that maximum predictability should be assigned to the upper bound of the green area in Fig 4B [41], in which case the trial-by-trial predictability achieved by the LN model would be suboptimal. We currently lack effective tools for characterizing the detailed structure of internal noise within human observers, however based on recent attempts [26] it appears reasonable that the typical region of maximum predictability for human vision should lie between the two extremes; this region is compatible with the values generated by the LN model (see above). We also converted these measurements into estimates of internal noise [39] to confirm that its properties are compatible with the notion of a late additive source often assumed by LN models [4] (Fig 4C–4D, see caption).
It may appear surprising that human visual processing displayed such compliance with the LN model, particularly with relation to the target-present/target-absent comparison (Fig 3): a representative survey of relevant literature indicates that these two estimates differ at least as often as they do not [4, 21, 36], and sometimes substantially so [34, 37, 38, 47]. It may therefore be argued on the basis of pevious studies that the lack of any difference, rather than its presence, should be viewed as unexpected and atypical. Our results, however, must be interpreted in light of the consideration that every aspect of the adopted experimental design was optimized to achieve template matching on the part of human observers (e.g. the stimulus was presented centrally [26], we placed two noiseless target signals above and below the central probe [48], we explicitly instructed observers to match the probe against those target replicas [21, 26], observers were given trial-by-trial feedback). Our objective was to test the applicability of LN modelling under conditions that favoured this processing mode, so that we could gauge the full extent of its explanatory power. Our analyses support the applicability of the LN cascade with respect to each dimension we probed. This result does not imply that the same LN model operating within a common representation also accounts for all results obtained across different representations. We turn to the latter issue below.
In attempting to identify a single specific implementation of the LN cascade that may account for our entire dataset, the natural starting point is a LN model that applies a 2D template similar to the estimated 2D PF (followed by the static decisional nonlinearity, see Methods). We implemented the 2D LN model using the target signal as template (Fig 5A) rather than the estimated PF (Fig 2A). This choice is motivated by the following considerations: 1) it allows us to exploit the full resolving power of the entire dataset by avoiding the need for cross-validation [49]; 2) our results will serve future investigations even where data mass is insufficient to support PF estimation [50]; 3) model evaluation will be based on ‘structural failures’ of the associated simulations (i.e. qualitative departures from data that cannot be ameliorated by tweaking model parameters), so that fine details of template structure are irrelevant.
Unsurprisingly, the 2D LN model generates good PF predictions for 2D/1D probes (Fig 5B–5C), however it is unable to simulate the negative modulations orthogonal to target orientation observed for orientation-tuned PF’s (compare blue with black traces in Fig 5D and 5I); as we have demonstrated in Fig 2H via quantitative analysis, those modulations reflect genuine structure in the data. No degree of model tweaking would allow the 2D LN model to generate those negative modulations: the underlying human template contains no structure along the orientation orthogonal to the target (Fig 2F), which in turn implies orientation-tuned PF’s lacking modulations within that region of orientation space. The LN model specified above must therefore be rejected as a viable account of our dataset. Because this conclusion is solely based on characteristics associated with the linear filtering stage (L) in the LN model (those characteristics are evaluated via the corresponding PF estimates which, under the LN model, return an image of the linear filtering stage), it is valid regardless of the characteristics associated with the static nonlinearity (N in the LN model) and therefore generalizes to any such nonlinearity supporting a sensible discrimination model. In other words, even if one were to allow for different characteristics of the static nonlinearity to be associated with different noise types, the 2D LN model just considered would not be compatible with our results.
The structural failure detailed above can only be addressed by inserting a model component orthogonal to the target signal; because peak and trough amplitudes were comparable (Fig 2H), it would appear that the orthogonal component should be assigned equal gain to the component aligned with target orientation. We must rule out a push-pull model where the template is obtained by simply taking the difference between target and non-target signals (we further consider this model later in the article), because the associated 2D filter estimate should be itself an image of that difference, contrary to the observed PF (Fig 2A and 2E–2F). The next minimal incremental modification of the push-pull model involves squaring the output of the two templates before they are subtracted [51] (Fig 5F). This variant, which no longer belongs to the LN family, generates orientation-tuned filters fully overlapping with those observed experimentally (compare red and black traces in Fig 5D and 5I), however it fails to generate suitable 2D filter descriptors (Fig 5B and 5G). This failure is once again structural: there is complete symmetry between vertical and horizontal orientations at the level of stimuli, task and model, so that all PF estimates must be similarly symmetric; we observed symmetry for orientation-tuned PF’s (Fig 2H) but not for 2D filters (Fig 2F), ruling out the nonlinear push-pull model. To summarize, the push-pull model improves on the LN model in its superior ability to capture orientation-tuned PF’s (blue data points in Fig 5J fall below the unity line at p < 0.0002 by two-tailed paired WSR test for x different than y values), however it is poorer at accounting for 2D/1D PF’s (red/orange data points in Fig 5H fall above the unity line at p < 10−4).
An additional canonical operation in the construction of small-scale circuits for neural computing is divisive normalization [19, 52, 53]. We implemented the most basic version of this operation: the linear drive is supported by the target-like matched template, and the gain-control operator pools from only parallel and orthogonal filters (Fig 6A). This minimal version of gain-control is sufficient to account, at least qualitatively, for all empirical PF estimates with no identifiable structural failure (Fig 6). On average, the model accounts for 94% of the variance in the aggregate data across 1D, Θ and SF conditions (variance accounted for in the 2D condition is inevitably low at 0.27 due to the high density of the noise probe, which results in high measurement noise and a sparsely modulated PF; for this condition we rely on the reasonable qualitative match between fitted surfaces indicated by ovals in Fig 6B and 6F). This value is more than satisfactory when one considers that the relevant implementation involves no free parameters and can be deployed without prior estimate of the underlying 2D filter. To summarize, the gain-control model is able to rectify the inadequacy of the LN model in capturing orientation-tuned PF’s (blue data points in Fig 6I fall below the unity line at p < 0.0002), while at the same time retaining its ability to account for 2D/1D PF’s (red/orange data points in Fig 5G scatter around the unity line at p = 0.43) and characteristic features of the SF-tuning data (magenta data in Fig 6I scatter around the unity line at p = 0.9; see also Fig 6E).
Despite its highly nonlinear nature, the gain-control circuit in Fig 6A operates in a manner well approximated by linear templates when projected onto and defined across each of the four dimensions probed by our experiments: Fig 7A plots indices returned by the two LN tests previously applied to the human data (see Fig 3A); there was no detectable sign of departure from template matching for any condition under either test (black/gray histograms centred around 0). This result is not due to lack of resolving power: when applied to the push-pull model (Fig 7B) both tests return positive distributions for the 1D condition, because this model generates mismatched target-present and target-absent estimates (compare red with blue traces within left inset to 1D condition in Fig 7B; see also clear modulations of its second-order kernel, right inset). Neither test detects departures from LN transduction for the 2D condition, however this is not because the PF’s associated with the push-pull model were consistent with template matching, but rather because they were nearly featureless and therefore inconsistent with the human data (2D orange distribution in Fig 7B, reflecting first-order PF sructure, is centred around 0; see also Fig 5B and 5G).
As a final step in cross-checking the applicability of the gain-control model, we computed its trial-by-trial predictive power and confirmed that it falls within the optimal range (inset to Fig 4B): collectively across all estimates, human-model agreement falls significantly above/below the lower/upper bounds of the maximum predictability region (p < 10−6/10−8) and does not differ from the midpoint between the two bounds (p < 0.42).
The above detailed experiments were specifically designed with the objective of isolating a single elementary operator for cortical processing of visual signals. One stimulus feature associated with this effort involved placing two vertically oriented signal templates above and below the probe region (see Methods and S1 Video). As explained previously, these templates served the purpose of prompting a matching strategy on the part of the human observers [48]; if observers could not implement a pure matching strategy (as supported by our results), the target templates would at least prompt reliance on the read-out mechanism associated with the target signal, whether the horizontally oriented non-target signal was absent (detection) or present (discrimination). In this manner, we hoped to facilitate experimental conditions where observers relied on the same elementary operation across different tasks, so that we could inspect the properties of said elementary operation in the presence of different stimulus conditions. The similarity between PF’s derived from the two tasks (detection versus discrimination) provides compelling evidence that we succeeded in this targeted effort.
In the discrimination task, the optimal strategy for observers is to engage the difference between the elementary operator associated with extracting the vertically oriented target signal and the elementary operator associated with extracting the horizontally oriented non-target signal. However, because we deliberately steered observers towards relying only on the former operator (while ignoring the latter) via placement of the target templates, it is conceivable that the observed departure from the optimal strategy (Fig 2E–2F) may be a direct consequence of our targeted effort to bias their read-out machinery towards the just detailed suboptimal strategy. It may therefore be of interest to characterize the system under conditions in which the combined output of the two elementary operators is facilitated by the additional placement of non-target templates to the sides of the central probe region (Fig 8A, green).
Although both 2D and orientation PF’s associated with this ‘symmetric’ variant of the discrimination task largely resemble those obtained in the absence of the non-target templates (compare Fig 8B and 8C with Fig 2E and 2G), there are important differences. First, the symmetric variant is associated with a measurable presence of the horizontal component. Although this effect is only mildly visible at the level of the aggregate 2D PF (Fig 8B, left), it is more clearly exposed by the non-target match values obtained from individual observer PF’s. More specifically, all 5 observers who had not participated in the original version of the discrimination task (that is in the absence of non-target templates) returned negative match values (all black data points in Fig 8D fall to the left of the vertical dashed line), indicating the presence of the negative non-target image expected under conditions where both vertical and horizontal elementary operators are engaged. This result was not observed for the original variant of the discrimination task (open symbols in Fig 2F). The only observer who did not follow the pattern prompted by the new variant of the discrimination task (grey data point in Fig 8D) was also the only observer from this pool who had participated in the original variant; it is reasonable to interpret this finding as reflecting the possibility that this observer retained the read-out strategy she had previously developed and failed to readjust.
A second important feature that characterizes the symmetric variant of the discrimination task is the appreciable difference between target-present and target-absent PF’s. Qualitative inspection of the target-present PF (Fig 8B, top right) suggests that it contains primarily if not exclusively vertical structure, while the target-absent PF (Fig 8B, bottom right) resembles more closely the difference between vertically-oriented and horizontally-oriented operators. These qualitative impressions are quantitatively confirmed by the lack of significant non-target content within target-present PF’s from individual observers (red data points in Fig 8D scatter around the vertical dashed line) and by the roughly equivalent content of target and non-target match within target-absent PF’s (blue data points in Fig 8D scatter around the diagonal negative unity line). Orientation-tuned PF’s are suggestive of potentially related differences: although these effects appear slight upon qualitative inspection of aggregate PF’s (compare red versus blue traces in Fig 8C), target-absent PF’s from individual observers present less marked peak values (corresponding to the target orientation) and more marked trough values (corresponding to the non-target orientation) than target-present PF’s (blue data points are shifted down and to the left of red data points in Fig 8E).
These differences between target-present and target-absent PF’s are indicative of a departure from the LN model [21, 31–38]. We therefore expect that this departure should be measurable using the first-order/second-order nonlinear tests we have applied to previous data. Indeed, we not only find that taken together the two tests return significantly positive values (at p < 0.002), but also that they strongly correlate (green data points in Fig 8I extend into the upper-right quadrant and display correlated scatter at r = 0.77, p < 0.004). No such effects are visible for the original variant of the discrimination task (black data points in Fig 8I).
We attempt to model these results by combining the modules developed in the previous section. The primary goal of this exercise is to exclude a role for template matching, not to simulate all details of the dataset. For this reason, we do not attempt to capture the observed differences between target-present and target-absent PF’s. The LN model involves subtracting the output of the horizontal template matcher from the output of the vertical template matcher. Because the empirical 2D PF contains more vertical than horizontal structure (Fig 8D), we halved the output of the horizontal template matcher before applying the subtraction in order to improve its ability to simulate the 2D condition (see Methods). For this model, we know that target-present and target-absent PF’s do not differ [21, 31–38] (not shown). The output reduction applied to the horizontal component translates into a reduced trough within the orientation-tuned PF (magenta in Fig 8G). If we do not apply output reduction, the LN model successfully returns trough and peak of equal amplitude for the orientation-tuned PF as observed experimentally, but fails to capture the unbalanced structure of the 2D PF. In other words, the LN model is able to account for either condition (2D versus orientation) in isolation, but not both concomitantly (similar to what we found previously).
The gain-control model fared substantially better. Again, we subtracted a horizontally-tuned gain-control circuit from the vertically-tuned gain-control circuit developed previously (Fig 6A). As with the LN model, the output of the horizontally-tuned circuit was halved before subtraction. This model captures both 2D and orientation-tuned PF’s, in that it produces trough and peak of equal amplitude and overlaps fully with the human data (Fig 8F–8G). Although it is therefore superior to the LN model (Fig 8H), we find that it does not generate appreciably different target-present versus target-absent PF’s (top/bottom right in Fig 8F), failing to account for this specific feature of the human data. It is conceivable that this failure may be ameliorated by more elaborate variants of the gain-control model, however this is not our goal. As we have explained above, due to its mixed nature, the read-out mechanism engaged by observers in the symmetric variant of the discrimination task is not ideal for achieving the goal of excluding/supporting the LN model using a non-parametric approach. Although it is certainly interesting to consider this variant and how it may lead to partially different results, our core conclusions rely on the main detection/discrimination tasks; under those conditions, a single elementary operator can be characterized in isolation.
It is uncontroversial that human vision often displays highly nonlinear characteristics [54], yet the linear-nonlinear model retains a paramount role in shaping past and current accounts of this fundamental sensory process. There are at least two reasons for its popularity.
First, the presence of an output static nonlinearity combined with a judicious choice of input space for stimulus projection often allow for effective and compact accounts of apparently more complex phenomena. A fitting example is motion detection, an inherently nonlinear process [55]. In the retina, this phenomenon is typically modelled using Reichardt cross-correlation [56], a nonlinear scheme that combines the output of multiple (at least two) elementary units [57]. Although this model does not conform to the LN scheme with reference to its original structure, it can be recast in the form of a linear oriented spatiotemporal filter followed by a static nonlinearity [58]; indeed, the latter scheme is more commonly adopted to account for cortical and behavioural processes [59, 60].
Second, LN models are often adequate for qualitative thinking and descriptive purposes, particularly when the system is challenged under a limited range of experimental conditions. This approach is exemplified by popular accounts of contrast-based illusions like the Hermann grid phenomenon [61], where the associated phenomenology is referred back to LN models incorporating front-end linear filters with an inhibitory surround [11]. Under some conditions, these accounts can be exploited to some degree of quantitative interpretation, providing for example a rationale for the broad agreement between filter size estimated from the above model of the Hermann grid illusion, and corresponding neuronal receptive field size measured electrophysiologically at ranging eccentricities [62].
Practically speaking, the LN model is attractive thanks to its analytical tractability. Combined with controlled input stimulation, this model makes simple predictions that have been extensively exploited to support characterization of the front-end linear filter [4, 16, 63]. Indeed, transparent interpretability of most measurements presented in this study is largely compromised wherever the LN model becomes inapplicable, although we have demonstrated here and in previous work that adequate tools exist for tackling such situations [10].
It is therefore clear that LN models are both useful and desirable, but it is also clear that they may fail under a range of conditions for quantitative purposes. We currently have no clear indication of when such failures may occur, how widespread they may be, and to what extent they may impact quantitative conclusions regarding human sensory processing. Only a few studies have examined this issue in sufficient detail [26, 36, 37, 44], and none has carried out an extensive characterization that would deliver a comprehensive picture across substantially different aspects of the same stimulus/task. As we discuss below, an integrated approach of the latter kind does not merely represent a quantitative extension of previous efforts, but rather enables a qualitatively different level of dissection of the relevant mechanisms and supports novel conclusions not available to previous studies.
The present study represents an attempt to determine whether there is at least one limited set of identifiable conditions under which human vision engages exclusively LN circuits. It is not intended as an attempt to determine whether the entirety of human visual processing can be reduced to LN transduction: as indicated above, an attempt of this kind would be fool-hearted, because it is inconceivable that the whole of vision would involve no more than template matching. Rather than asking whether all visual operations are template matchers, we ask whether there is at least one visual operation that involves template matching. We reasoned that the strongest and most relevant test of the latter possibility would involve task and stimulus specifications that are not only representative of core interests in vision science [43], but also probe elementary operations supported by visual cortex [64] using experimental protocols specifically designed to facilitate LN transduction (see below for further discussion of these points).
There is a sense, of considerable practical significance, in which our results provide encouraging evidence to support the applicability of LN models for understanding and characterizing human pattern vision: under all the experimental conditions we tested, the human process could be adequately described in the form of a simple LN model applied to the dimension probed by the perturbation associated with a given noise mask. This result enables a wide range of tools that have been tailored to the LN family [4, 16, 65]. It must be emphasized that the conditions of the detection/discrimination experiments presented here were carefully adjusted to prompt a template-matching strategy on the part of human participants. As we have demonstrated with the symmetric variant of the discrimination task (Fig 8), apparently irrelevant methodological details (e.g. inclusion of target/non-target replicas adjacent to the probe) may impact the extent to which these experiments are representative of a wider range of specifications: it goes without saying that, as more elementary operators and processing layers are loaded onto the read-out stage, the collective process (which in the limit will encompass the whole of vision) will inevitably manifest departures from LN transduction.
There is however a different sense, arguably of greater theoretical significance, in which our results are not equally supportive of LN modelling schemes: it is the sense of understanding the deeper structure of the system beyond its superficial compliance with LN transduction in the manner discussed above (see [63] for related pursuits in neuronal modelling). In this sense, we were unable to identify a single implementation of a specific LN model that would capture all aspects of our complex dataset (Fig 5; see related results from electrophysiological recordings of simple cells [24, 53]). This failure was structural in that it involved qualitative departures from the human data that could not be ameliorated via further exploration of parameter space. Outside the LN family, we successfully identified a viable candidate by incorporating a canonical computation for cortical circuits: gain control via divisive normalization [19]. The applicability of this operation to neural processes is extensively documented [52, 66]. The success of this model in capturing our own data therefore conforms to current trends in the computational literature. The non-parametric logical/analytical process by which we mustered support fo the gain-control model differs in its outlook from previous attempts based on fitting multiparametric models [67–70]; nevertheless, it is notable that gain control circuits feature prominently in those studies too.
Is it conceivable that a comparable set of experiments might have been identified that did not require any modelling tools outside the LN family? This seems unlikely. As mentioned earlier, extensive piloting went into ensuring that template matching would be encouraged as exhaustively as feasible on the part of human participants, leaving little room for further tailoring of stimulus specifications to favour LN strategies. Furthermore, the function probed by our protocol is elementary: Gabor wavelets represent some of the most efficiently detectable visual patterns [64], consistent with the notion that their structure resembles neuronal preference in cortex [71]. Any detection/discrimination task relevant to human visual function would presumably involve combinations of analogous operations [72, 73], leading to the expectation that it would display at least an equivalent, if not more pronounced, degree of departure from the LN model. This is indeed the result we observed when we altered the discrimination protocol to prompt a strategy whereby observers would combine two elementary operations (Fig 8). Based on the above considerations, we believe that our experiments are ideally positioned to draw conclusions about the applicability of LN models to human vision: if there exist any conditions, no matter how limited, under which such models are applicable, those conditions would include the specifications probed by our protocols.
The outcome of our experiments enables one characterization for the underlying mechanism while excluding a number of alternative scenarios, all nonetheless viable and plausible. In other words, it would have been entirely reasonable to expect and observe a substantially different outcome. Under one scenario, the orientation and SF tuning characteristics returned by our PF measurements may have conformed to those predicted by the specified 2D version of the LN model (blue traces in Fig 5D and 5E), leaving open the possibility that the system was exclusively engaging LN circuits across the board. Under a different scenario, the system may have operated in the manner of the push-pull model outlined in Fig 5F (see [51] for a concrete example of the applicability of this model to data from a discrimination task closely related to the one used here): lack of compliance with LN transduction would have then become apparent from applying the linearity tests to one condition alone (1D, see Fig 7B), yielding the conclusion (opposite to the one immediately above) that the system was engaging more elaborate circuitry than LN components, and furthermore that the design of such circuitry did not support LN behaviour even when restricted to individual probes (as we observed for the symmetric variant of the discrimination task, see Fig 8D–8E and 8I). Both our findings, i.e. that system structure does not conform to LN circuitry and yet complies with LN transduction under varying conditions, are therefore independent contributions that place important constraints on the range of plausible scenarios potentially associated with the visual processes examined here. It is relevant in this respect that our ability to dissect the underlying mechanisms with adequate discriminatory power specifically relied on evaluating different features of our combined dataset: had we considered each condition in isolation (e.g. only the 2D results or only the orientation-tuning results), it would have been impossible to constrain our conclusions to the extent that was enabled by the integrated analysis presented here.
A potentially productive way of summarizing our results may be obtained via reference to the simple concept of locally linear approximation for nonlinear functions. As illustrated by the cartoon in Fig 9, we can think of the visual process as a manifold spanning a space that encompasses all possible dimensions across which the stimulus may be usefully represented (clearly such high dimensional spaces cannot be adequately represented in a 2D cartoon, so Fig 9 is only intended as an intuitive tool and not an accurate description of the process). When projected onto a specific subspace (e.g. orientation or spatial frequency), and when inspected with respect to that restricted subspace, the behaviour exhibited by the process may be satisfactorily approximated by the LN framework, even though this framework may not be adequate to describe the process as a whole, i.e. with respect to its collective characteristics across multiple projections. Our results indicate that the operations of human vision, no matter how elementary and limited in their immediate scope (e.g. detection of a Gabor wavelet), cannot be reduced to a straight pipe through Fig 9: they retain an irreducible level of nonlinear structure possibly reflecting the minimal functional characteristics implemented by cortical circuits [5, 19, 52, 53].
Fig 9 may misleadingly suggest that the introduction of different noise masks in our study is equivalent to the expansion of stimulus range afforded by previous investigations that manipulated e.g. stimulus uncertainty [20, 21, 37] and/or pedestal contrast [67, 74] (these two factors being potentially intertwined [20, 74–76]). From the perspective of those and related studies, it may seem trivial that linearity breaks down as stimulus range is expanded. There is however a critical difference with respect to our approach, in that most previous studies manipulated the range/specification spanned by the signal to be detected, thereby potentially prompting the system to engage different modules/regimes to perform the assigned task [36]. The target signal to be detected in our experiments was fixed and uniquely specified, regardless of the noise perturbation that was added to it. Under these conditions, observers were prompted to engage a single elementary perceptual operator. Furthermore, the effects of expanding stimulus range have often been modelled via LN cascades with a sigmoidal static nonlinearity and/or one that changes exponent [13, 74, 77], but which nevertheless perform LN transduction; our results exclude LN models regardless of the specific characteristics associated with the static nonlinearity, and regardless of whether those characteristics may differ for different noise types. Finally, the notion that linearity should break down as stimulus range is expanded implicitly relies on the assumption that linearity does apply in the first place within a restricted input range; this assumption has never been adequately checked, at least not to the extent afforded by the experiments/analyses presented here. With these caveats in mind, Fig 9 is best interpreted not as the trajectory of a perceptual system that traverses different stimulus regimes and potentially modifies its characteristics along the way, but rather as the intrinsic structure of an elementary process operating within a minimally defined input range. The combined application of different noise probes delivers a multifaceted view of this process that is not afforded by each individual probe in isolation.
It may seem counterintuitive that an inherently nonlinear architecture would be in place for it to behave linearly with respect to substantially different visual dimensions, such as 1D space or orientation as probed by our stimuli. If linear transduction is the goal, why not implement it using the template matcher in Fig 5A? If conversely nonlinearity is the goal, why build a nonlinear system that retains such degree of linear transduction as we measured here, and not adopt the push-pull circuit in Fig 5F regardless of its highly nonlinear transduction properties (Fig 7B)? Our results suggest that, at least under the conditions of our experiments, the system strives to achieve linear encoding: it seems otherwise difficult to explain why we observed such extensive compliance with template matching for processing 4 different noise probes across 2 separate tasks, when prior studies have exposed clear deviations under more limited conditions [4, 21, 34, 36–38].
Although elementary template matchers like the mechanism in Fig 5A support linear transduction, they may not be adequate for the purpose of versatile stimulus encoding in ways that are both linear and useful; in this context, utility may involve the necessity to represent orientation in a balanced push-pull fashion as we observed in our experiments, a goal that cannot be achieved by the circuit in Fig 5A (see blue traces in Fig 5D and 5I). An alternative strategy, supported by our findings, involves assembling small nonlinear circuits that support effective stimulus encoding (e.g. push-pull orientation selectivity as in Fig 2C and 2G and sharp bandpass SF tuning as in Fig 2D) while at the same time retaining linearity across a wide range of tasks and stimulus perturbations [24, 52]. Divisive normalization has proven an effective tool for efficient transduction while maintaining gain within near-linear regimes [19]; Fig 9 elaborates on this property to encompass the collective space of extended stimulus projection for multi-feature encoding (with the caveats outlined above). Although this interpretation is highly speculative, its proposed mode of operation is known to underlie other aspects of sensory processing [66, 78], in particular retinal encoding of ON/OFF signals: in the retina, linearity is a luxury that comes at the cost of carefully assembled nonlinear subunits [79]. Our results suggest that similar principles may apply to some cortical computations [80].
|
10.1371/journal.ppat.0040028 | Methamphetamine Inhibits Antigen Processing, Presentation, and Phagocytosis | Methamphetamine (Meth) is abused by over 35 million people worldwide. Chronic Meth abuse may be particularly devastating in individuals who engage in unprotected sex with multiple partners because it is associated with a 2-fold higher risk for obtaining HIV and associated secondary infections. We report the first specific evidence that Meth at pharmacological concentrations exerts a direct immunosuppressive effect on dendritic cells and macrophages. As a weak base, Meth collapses the pH gradient across acidic organelles, including lysosomes and associated autophagic organelles. This in turn inhibits receptor-mediated phagocytosis of antibody-coated particles, MHC class II antigen processing by the endosomal–lysosomal pathway, and antigen presentation to splenic T cells by dendritic cells. More importantly Meth facilitates intracellular replication and inhibits intracellular killing of Candida albicans and Cryptococcus neoformans, two major AIDS-related pathogens. Meth exerts previously unreported direct immunosuppressive effects that contribute to increased risk of infection and exacerbate AIDS pathology.
| There is a new population of HIV+ men who are developing AIDS over months instead of years as typical. It has recently become popular among gay and bisexual men to consume very high levels of Meth. Unsafe sex together with Meth abuse has been suspected to lead to rapid disease progression. While studies show exacerbated AIDS symptoms and disease progression in HIV+ Meth abusers, the molecular mechanism is yet unknown. It was postulated, yet unproven, that the rapid disease progression might be due to a mutant “superstrain” of HIV that was extremely virulent. It was also assumed that the effects of the drug on behavior may lead to unsafe sex, although this would not explain the more rapid time course of the disease. We now demonstrate the first direct evidence that Meth is an immunosuppressive agent, and that the molecular mechanism of this immunosuppression is due to the collapse of acidic organelle pH in cells of the immune system, inhibiting the functions of antigen presentation, as well as phagocytosis. These effects compromise the immune response to opportunistic infections and HIV. These findings could have a major impact on public health, as there are over 35 million Meth abusers worldwide
| Chronic methamphetamine (Meth) abuse has reached epidemic proportion throughout the United States, where a 2003 survey indicated that approximately 5% of the population over 12 years of age has tried Meth and the rate of treatment admissions for primary Meth abuse increased over 3-fold (The DASIS Report, http://www.oas.samhsa.gov/2k6/methTX/methTX.htm) [1,2]. In particular, among gay and bisexual men [3] Meth it is associated with high-risk sexual behavior, HIV viral infection, and a high incidence of AIDS [4]. Meth exacerbates AIDS pathology, including cognitive deficits [5,6], and is strongly suspected to inhibit normal immunological response to secondary infections such as hepatitis C, which is prevalent in those who smoke or take Meth intranasally [7]. It has recently been suggested that Meth could contribute to a particularly rapid progression of AIDS in individuals exposed to a strain of HIV that is resistant to drug treatment [8,9]. Indeed, animal studies clearly demonstrate that Meth suppresses both innate and adaptive immunity [10,11], enhances cytokine production in combination with HIV TAT protein [12], and alters gene expression in cells of the immune system [13]. However, the molecular basis for Meth's immune suppression is unknown. Here we have examined the relationship between Meth and the impairment of specific immune cell functions.
In the clinical setting, Meth abusing individuals who present with opportunistic infections possess high blood and tissue levels of the drug. Meth is generally self-administered in this population in binges of 3–4 grams ingested over a six day interval [14], with an initial ingestion of ∼0.5 gram (Judith Rabkin, Columbia University, personal communication) and a total mean level of 2.2 grams ingested during the first day [15]. Such high levels of administration result in a blood concentration of ∼10–50 μM Meth, and levels in the hundreds of micromolar range in organs including brain and the spleen [16] (see Results).
Pathogens are processed and displayed by antigen presenting cells (APCs) for T cell recognition. Antigen presentation in tissue resident macrophages, as well as dendritic cells, involves fluid phase or receptor-mediated endocytosis followed by fusion of the phagosome with specialized lysosomes known as MHC class II compartments (MIIC) [17]. The foreign antigens are partially degraded by lysosomal hydrolases and the resulting peptides are loaded on MHC class II molecules and transported to the cell membrane to be presented to T cells. Both pathogen transport to the MIIC and its degradation into immunogenic peptides are functions that require an acidic endosomal pH. Endogenous antigens and viruses can be delivered to the MIIC upon autophagosome fusion [18]. Autophagy also mediates single-stranded RNA virus detection and consequent interferon-α release in plasmacytoid dendritic cells activating anti-viral cellular defense mechanisms in uninfected cells [19]. Further evidence shows the direct role of autophagosomal-lysosomal degradation in elimination of intracellular HSV-1 viral particles [20] and that the HSV-1 has alternative methods to counteract cellular autophagy [21].
In this study, we have identified a novel molecular mechanism that explains the immunosuppressive effects of Meth via the alkalization of acidic organelles in dendritic cells and macrophages that are critical for the immunological function of these APCs.
To model effects of Meth on the immune system, we estimated Meth levels in drug abusers. Meth is self-administered intravenously, by nasal inhalation, anally, and orally, in typical doses of 250–500 mg by occasional users to levels as high as 1 g by chronic abusers (personal communication, Perry N. Halkitis, New York University). Meth blood levels measured in individuals detained by police in California were 2.0 μM on average but as high as 11.1 μM [22]. Controlled studies indicate that a single 260 mg dose reaches a level of 7.5 μM [22]. Thus, a single dose of 260 mg – 1 g would be expected to produce 7.5 – 28.8 μM blood Meth levels.
The abusers however tend to self-administer METH in binges, and as the drug exhibits a half-life of 11.4 - 12 h [23,24], this leads to higher levels. Recently published studies modeling binge pattern of use in individuals show that after the fourth administration of 260 mg during a single day produces maximum blood levels of 17 μM, reaching 20 μM on the second day of such a binge [22]. Thus, binge doses of 260 mg – 1 g would produce 17 - 80 μM blood Meth levels. The estimates appear consistent with blood levels detected after fatalities [25–27], ranging as high as 84 μM for an individual for whom Meth intoxication was determined as the cause of death ∼16 hours after ingestion.
It is also important to estimate how Meth is distributed from blood to other tissues involved in immune response, particularly the spleen, which houses high numbers of dendritic cells. Tissue-to-serum Meth ratios in rat are: brain, 9.7; kidney, 35.3; spleen, 14.3 [16]. Thus, relevant levels in spleen, the organ critical for immune response, after administration of 250 mg – 1 g as a single dose is 100 – 400 μM, and during binges between 240 – 1144 μM.
Dendritic cell MIIC processing organelles are characterized by acidic pH [28], a limiting membrane enclosing internal vesicles or lamellae [17], the presence of proteases [29], internal expression of LAMP [30] and MHC II proteins [31]. The morphological and functional integrity of the MIIC depends on the maintenance of an acidic pH [32] due to the action of the V-ATPase, which may also regulate transport from early to late endosomes and lysosomes [33]. Meth and its metabolite, amphetamine, are membranophilic weak bases that collapse intracellular organelle pH gradients in neurons [34,35]. We therefore tested whether Meth collapsed intracellular organelle pH gradients in dendritic cells by monitoring quenching of acridine orange, a weak base vital dye that accumulates in acidic organelles including endosomes and lysosomes [35]. We found that Meth concentrations of 50 μM or higher rapidly collapsed acidic organelle pH gradients in dendritic cells (Figure 1A–1B). LysoSensor Yellow/Blue ratiometry is a generally accepted method to measure organellar pH in live cells, and was used to measure average pH in acidic organelles using LysoSensor Yellow/Blue fluorescent dye. After 10 min treatment with Meth (50 or 100 μM) or chloroquine (10 μM) acidic organellar pH was elevated (pH 6.5, 6.9 and 6.4 respectively) significantly above the levels in untreated control cells (pH 4.8) (Figure 1C). As predicted with alkalizing agents, Meth also disrupted MIIC structure [36], producing large organelles (> 1 μm diameter) devoid of internal vesicles (right panels) with LAMP-1 and MHC II staining confined to the limiting membrane (Figure 1D and 1E). Chloroquine (Clq) [10 or 20 μM], another weak base was used as a control showed similar effects to Meth as described above (Figure 1A–1E). At the used concentrations (20 and 100 μM) Meth did not affect cell viability as determined by flow cytometric analysis (Figure S1).
To investigate whether Meth-induced endosomal alkalization blocks antigen processing by impairing dendritic cell lysosomal proteolytic degradation of foreign proteins, we exposed cells to the fluorescently labeled MHC II antigens [bovine serum albumin (BSA), casein, and ovalbumin (OVA)], and measured the degradation of each protein by western blot. In untreated cells, each antigen was proteolytically degraded, while antigen degradation was blocked in Meth or Clq treated cells (Figure 2A and 2B). Meth (10, 50, 100 μM) and Clq (10, 20 μM) effectively inhibited processing of antigens previously taken up by dendritic cells (Figure 2A). Control experiments showed that Meth did not block endocytosis of 2 μm fluorescent dextran beads by dendritic cells (Figure S2), indicating that the drug does not inhibit nonspecific phagocytosis [37] .
To determine the stage at which Meth compromised post-endocytic proteolytic antigen processing, we prepared fractions of early and late endosomes and lysosomes from dendritic cells exposed to BSA and casein antigens. Each fraction was examined for β−hexosaminidase to identify lysosomes and late endosomes, the transferrin receptor (TrfR) to identify early endosomes, and LAMP-1 to identify late endosomes and lysosomes (Figure 2C and 2D). No detectable antigen remained in early endosomes under any of these conditions. There were, however, much higher levels of BSA and casein in lysosomes and late endosomes of Meth- and Clq-treated cells than in controls. In particular, casein was completely degraded in untreated cells but relatively unprocessed in dendritic cells treated with Meth or Clq (Figure 2E). Thus, Meth inhibited antigen proteolysis within late endosomal/lysosomal compartments. Similarly, processing of invariant chain was compromized after Meth treatment as indicated by increased levels of p25/28 and p10 fragments during the chase time-point (Figure 2F).
MHC II antigen presentation stimulates T cell proliferation, providing a means to measure effects of Meth on antigen presentation. We prepared cultures of immature bone marrow-derived dendritic cells and splenic purified T cells, both from OTII transgenic mice [38]. Using OVA as an antigen, we assayed cellular proliferation of T cells by radiolabeled thymidine uptake and incorporation into DNA. At all levels tested, Meth decreased the T cell proliferative response to the intact antigen (Figure 2G) but not the pre-processed OVA-323–339 peptide (Figure 2H).
In addition to blocking lysosomal antigen degradation, Meth could disrupt processing and presentation of antigens by inhibiting autophagosome formation, thereby halting antigen delivery to MIIC. To test this possibility, we prepared dendritic cells from a transgenic mouse expressing a GFP-fused autophagosome-associated protein LC3 (GFP-LC3) which has been used as an in vivo autophagosomal marker [39]. Consistent with previous reports in neurons [40], 50–500 μM Meth induced autophagosome accumulation in dendritic cells (Figure 3A–3C). Our data show that Meth does not block autophagosome formation, but rather impairs lysosomal-autophagosomal degradation, resulting in the accumulation of autophagosomes, and impaired intracellular antigen proteolysis. Together with the evidence from endosomal/lysosomal fractions (Figure 2C–2E), these results strongly suggest that Meth inhibits degradative antigen processing by disrupting pH gradients.
As with dendritic cells, we also found that Meth and Clq effectively collapsed the intracellular pH gradients within macrophages (Figure S3A–S3C) and blocked autophagosome degradation, resulting in the accumulation of GFP-LC3 labeled autophagosomes (Figure S4A–S4D). Similar results were obtained with the fluorescent dye monodansylcadaverine, a lipophilic weak base that accumulates in lysosomes and autophagosomes (Figure S4E).
Since Meth users can also present with bacterial infections [41], we examined whether Meth disrupts macrophage phagocytosis, a primary mechanism for clearance of these extracellular pathogens. Murine peritoneal-derived macrophages were incubated in the presence or absence of Meth. After 2 h, IgG-antibody coated erythrocytes [E(IgG)] were added to each of the conditions and the number of ingested erythrocytes counted. We found that Meth (50 and 250 μM) inhibited E(IgG) phagocytosis (Figure 4A and 4B) by 20 and 45%, respectively. Meth did not inhibit the receptor-independent endocytosis of Lucifer Yellow (Table S1). Similar effects were observed with Clq (Figure 4A and 4B) These results demonstrate that Meth inhibits Fcγ-mediated phagocytosis in macrophages and are consistent with the observations that the macrophage Fcγ receptors are continuously recycled between phagosomes and the plasma membrane [42], a process that requires appropriate acidification of secretory vesicles and tubules for trafficking [43]. These results demonstrate that Meth inhibits Fcγ-mediated phagocytosis in macrophages.
As a next step we analyzed the effects of Meth on Candida albicans (Ca) and Cryptococcus neoformans (Cn) phagocytosis and killing by murine macrophages since these two organisms are the most commonly isolated fungi in individuals infected with HIV [44]. Clq and Meth (10 and 50 μM) inhibited phagocytosis of Ca and Cn by macrophages by 40% (Figure 4C). Moreover, Meth enhanced the proliferation of fungi within macrophages (Figure 4D), indicating that intracellular replication of both yeast was facilitated by Meth. In contrast, Clq had no significant, or slightly reducing, effect on Ca (p= 0.056) and Cn (p = 0.060) CFU numbers in macrophages. Control experiments showed that in the absence of macrophages, Ca and Cn proliferation was unaffected by the addition of Clq or Meth to the BHI medium (data not shown).
To further examine why chronic Meth abuse has been recently associated with the rapid development of immune deficiency among gay and bisexual men [3], we studied the effect of Meth on HIV proliferation in macrophages from HIV-transgenic mice. These JR-CSF/huCycT1 double transgenic mice express HIV-1 JR-CSF, which is a full length R5 HIV-1 provirus regulated by the endogenous HIV-1 LTR, as well as the human cyclin T1 controlled by a murine CD4 expression cassette [45]. These mice have constitutive HIV production in CD4 T lymphocytes and monocytes. GM-CSF differentiated bone marrow cells were either left untreated or treated with a range of Meth levels (10, 50, 150 μM) or NH4Cl (10 mM), another weak base, for 7 days. Cell supernatants were then collected and p24, a secreted HIV-specific protein, was quantified by Elisa. Cells from JR-CSF/huCyc T1 mice treated with Meth for 7 or 9 days exhibited a 30–60% increase (Figure S5A) of p24 antigen production. Clq does not provide a positive control, since its well-established inhibition of HIV production is probably due to inhibition of viral capsid protein glycosylation in the Golgi [46]. Thus, in this experiment NH4Cl was used as positive control (Figure S5A).
To determine whether HIV replication was affected by Meth in vivo, we studied the effect of Meth on HIV virus proliferation in the JR-CSF/huCycT1 double transgenic mice [45]. Mice were treated with increasing concentration of Meth over a 7 day-period. One group of animals received 5 mg/kg of Meth at day 0, 2 and 4 and was sacrificed at day 6 (low Meth). Another group received 6 mg/kg at day 0; 7 mg/kg at day 2 and 7.5 mg/kg at day 4 (high Meth) and was also sacrificed at day 6. Copy number of HIV-1 RNA was quantified in the serum of each mouse by RT-PCR using primers spanning the highly conserved region of the HIV-1 gag gene. No statistically significant differences were observed between the untreated and the Meth-treated mice (Figure S5B). These data suggest that Meth treated HIV-1 transgenic mice do not exhibit an increase in HIV viral load.
We find that the widely abused addictive psychostimulant, Meth, at pharmacologically relevant levels acts as an immunosuppressive agent, due to its inhibition of endosomal acidification. These actions result in Meth's inhibition of antigen presentation and phagocytosis. Maintenance of low endosomal and lysosomal pH serves many functions, including regulation of protein degradation, pathogen inactivation, and regulation of the amount of several surface receptors. All of these functions require active transport via the endocytic pathway and fusion with lysosomal compartments.
First, via alkalization, METH-inhibited lysosomal-autophagosomal degradative function for both exogenously and endogenously internalized antigens resulting in accumulation of proteins entering the endocytic pathway through phagocytosis, as well as autophagic vacuoles. Chaperone-mediated autophagy and macroautophagy in lysosomes have been described as major pathways for endogenous antigen processing in MHC class II compartments [18,47,48] and as a means to directly degrade intracellular virus particles [20,21]. Thus, through its alkalizing effect, Meth blocks normal antigen processing and presentation. Progression of internalized antigens along the endocytic pathway rely on the progressive maturation of early endosomes into late endosomes and ultimately to lysosomes. Maintenance of an acidic internal pH and a pH gradient in these compartments is important for the cargo progression. Endosomal acidification is accomplished by H+ transport across the endosomal limiting membrane by the proton pump vacuolar ATPase (V-ATPase). The recent discovery that V-ATPase interacts with components of the endocytic transport machinery indicates that V-ATPase is also a pH sensor that regulates early to late endosomal transport [33]. This would explain why endosomal alkalization by Meth not only disrupts antigen processing but phagocytosis and cargo progression along the endosomal pathway.
Second, endosomal alkalization by Meth inhibited Ca and Cn phagocytosis and killing by macrophages. These effects are expected to be particularly devastating in AIDS-related disorders, since Ca and Cn are the two most commonly isolated fungi from sterile body fluids obtained from HIV infected individuals [44]. Also, these results could explain recent reports of rapid insurgence of AIDS in Meth-addicted individuals soon after infection with HIV. Thus, Meth also blocks pathogen killing by macrophages. Exposure to Clq and other weak bases was already been shown to inhibit growth of Cn in macrophages [49,50] even though Clq has no direct toxicity to Cn [50]. Not all basic compounds are equivalent in their fungicidal activity; for example, ammonium chloride also enhances the anti-cryptococcal activity of macrophages, yet the potency of ammonium chloride is less than that of Clq, which may be a result of the ability of ammonium chloride to inhibit phagolysosomal fusion [51]. Hence, Meth apparently has complex effects on macrophages that result in an intracellular milieu that enables the replication of the examined pathogenic fungi.
Third, even though an increase in p24 secretion has been observed in Meth-treated dendritic cells and macrophages, a direct effect of Meth on HIV viral load could not be demonstrated. It is likely that viral proteins enter late endosomal and lysosomal compartments independently of viral assembly, and due to inhibition of processing, more viral proteins are secreted in the extracellular milieu. Within the infected macrophages, HIV has been previously shown to assemble in compartments with characteristics of multivesicular late endosomes (CD63, Lamp-1, CD81 and CD82 positive) [52]. It now appears that the HIV particles present in multivesicular endosomes are the results of endocytosis, and the site of viral assembly is at the invagination of the plasma membrane particularly enriched in tetraspannin proteins [53]. These results would explain why, by compromising endosomal pH, an increase in HIV viral load has not been observed.
In conclusion, the immunosuppressive effects of Meth are consistent with reports that Meth-treated mice demonstrate decreased immunity [10,11]. To be noted is that even though there is a linear concentration-dependent response between Meth concentration and pH disruption such an effect is less evident at the biological level. In all assays (antigen processing, Ca or Cn killing and p24 production) different Meth dosages behave very similarly. Likely a small disruption in the endosomal pH is sufficient to alter the microenvironment and endosomal-related functions. This is consistent with the effects of low concentrations of Meth tested on lysosomal pH in Figure 1C, with alkalinization of > 1 pH unit, which would effectively inhibit lysosomal protease activity.
The collapse of endosomal pH by Meth and the resulting decrease of normal immune response provide an explanation for the compromised immunity and exacerbate infections occurring in Meth abusers. In fact, Meth is strongly suspected to more dramatically inhibit normal immune responses than other drugs of abuse since Meth users often present with skin lesion and “Meth mouth”, a devastating periodontal disease (http://www.drugfree.org/) [54] and [1,2]. A similar immunosuppressive activity has also been shown for chloroquine, also a well known inhibitor of endosomal acidification [36].
In particular, Meth immunosuppression may underlie the mechanism of the recently reported extremely rapid development of immune deficiency, with devastating effects in AIDS-related disorders in Meth abusers that had contracted HIV. In particular, there is evidence suggesting the presence of a new population of HIV+ positive men who are developing AIDS over months rather than over 10 or more years as is typical. The most widely reported individual was documented by Dr. Martin Markowitz of the Aaron Diamond AIDS Research Center in New York in 2004. In this individual, a gay man who was also a Methabuser tested HIV negative in May, 2003, then likely contracted HIV during unprotected sex in mid-October 2004, displayed acute retroviral syndrome in November, 2004 and 3-drug-class-resistant HIV-1 (3DCR HIV) with apparently rapid progression to AIDS by December, 2004 (CDC, MMWR July 28, 2006/55(29);793–796). Meth self-administration by HIV+ individuals during the acquisition of sexually transmitted pathogens appears likely to interfere with immunological resistance and lead to AIDS progression.
Additional detailed methods used in this paper can be found in Protocol S1.
Mouse femur hematopoietic stem cells from bone marrow were harvested from the hind legs of 8 to 12-week old male wild-type C57BL/6J (The Jackson Laboratory, Bar Harbor, Maine) or GFP-LC3 transgenic mice [39], and plated at 2 × 106 cells/ml density in DMEM supplemented with 10% FBS, 1x non-essential amino acids (Gibco, Carlsbad, California), 2 mM L-glutamine, 1 mM sodium pyruvate and 20 mM HEPES. For differentiation to dendritic cells or macrophages, 10 ng/ml of recombinant mouse GM-CSF (Biosource, Carlsbad, California) or 10 ng/ml of recombinant mouse M-CSF (R&D Systems, Minneapolis, MN) was added to media, respectively. Cells were fed every 2 days with fresh DMEM containing the appropriate macrophage colony stimulating factor. Cells were trypsinized after 8 days and, unless otherwise noted, plated at 4 × 105 cells/cm2 density to be used for experiments the following day.
Resident mouse peritoneal macrophages were isolated from 8 to 12 week-old female wild-type C57BL/6J as described [55] and plated on 12 mm-diameter glass coverslips in 24-well tissue culture plates at 3 × 105 cells/well density in RPMI media supplemented with 10% FBS with streptomycin and penicillin. Non-adherent cells were removed by washing 2 h after plating. The remaining adherent cells were over 95% macrophages as assessed by esterase staining [55]. Cells were incubated overnight prior to experiments.
Cells were stained with 10 mM acridine orange in phenol red-free media for 1 h. Images of stained cells were acquired using fluorescence microscopy as described above. Images of 10 fields with 5 to 10 cells for each image were taken using multiple stage positions in Multidimensional Acquisition mode under conditions of no photobleaching (ND2 filter, 1000 msec exposure) that enabled the acquisition of multiple images of the same cells. Stage position for each image was stored so images of the same fields and cells could be taken before and after treatment. Phenol red-free media containing 10 mM acridine orange in the presence or absence of Meth or Clq was added to the cells. At the end of the incubation time, images of treated cells were taken and used for morphometric analysis. Change in acridine orange intensity was measured as a change in average pixel intensity in the cytoplasmic area using MetaMorph Version 6.1r6 image analysis software (Molecular Devices, Sunnyvale, CA). Background mean pixel intensity was measured in nuclear area and subtracted.
Dendritic cells were stained with 5 μM LysoSensor Yellow/Blue (Invitrogen, Carlsbad, CA) for 5 min before Meth (10, 50, 100 μM) or Clq (10μM) was added for an additional 10 min incubation followed by washing with phosphate-buffered saline (PBS) (pH 7.4). Fluorescent images were taken of the same cells using Olympus IX81 microscope with Photometrics CoolSNAP HQ cooled camera, MetaMorph Version 6.1r6 imaging software (Molecular Devices, Sunnyvale, CA), Olympus PlanApo 40x/1.4 Oil objective, equipped with fluorescent yellow customized Chroma (D350/50 excitation, 400DCLP dichroic splitter and D535/40m emission) (Chroma technology Corp., Rockingham, VT) and blue Chroma 31000v2 (D350/50 excitation, 400DCLP dichroic splitter and D460/50m emission) filter sets. Average pixel intensity was measured in the cytoplasm of the cells excluding the nucleus using MetaMorph software, and the ratio of yellow to blue intensity was compared to a pH calibration curve to determine pH values. For the calibration curve, cells were stained with 5 μM LysoSensor Yellow/Blue for 20 min, washed with PBS and incubated in buffer of known pH (4.0 to 7.4) containing 10 μM monensin and 10 μM nigericin [56] before images were taken and processed as above.
Dendritic cells were derived and incubated for 4 h with or without Clq (20 μM) or Meth (100 μM). Cells were fixed in 2% paraformaldehyde and 2.5% glutaraldehyde in 0.1 M sodium cacodylate buffer, postfixed with 1% osmium tetroxide followed by 1% uranyl acetate, dehydrated through a graded series of ethanol and embedded in LX112 resin (LADD Research Industries, Williston, VT). Ultrathin (80 nm) sections were cut on a Reichert Ultracut UCT, stained with uranyl acetate followed by lead citrate and viewed on a JEOL 1200EX transmission electron microscope at 80 kV.
For immunogold labeling, cells treated as described above were fixed in 2% paraformaldehyde and 4% polyvinylpyrolodone in phosphate buffer 0.2 M (pH 7.4) at 4 °C. Fixed cells were processed for ultrathin cryosectioning as previously described [14]. Immunogold labeling was performed using LAMP-1 antibody (clone 1D4B, BD Pharmingen, San Diego, CA) followed by anti rat Ig-G coupled with 10 nm gold particles and biotinylated anti-MHC II (clone M5/114.15.2, BD Pharmingen, San Diego, CA) followed by streptavidin gold (15 nm). Contrast was obtained with a mixture of 2% methylcellulose (Sigma, St. Louis, MO) and 0.4% uranyl acetate pH 4 (EMS, Hatfield, PA).
FITC-coupled BSA, ovalbumin or casein were fed to immature bone marrow-derived dendritic cells (between 1 to 3 × 107 cells for each condition) at a concentration of 100 μg/ml. Cells were untreated or treated with Meth or Clq. After overnight incubation, cells were washed twice in PBS and lysed in 150 mM NaCl, 50 mM Tris-HCl and 1% NP40 supplemented with protease inhibitor cocktail. Post-nuclear supernatants were normalized for protein content and 80 μg of total protein was run on SDS-PAGE gel. Membrane blots were probed with the anti FITC mAb or β-tubulin mAb (Sigma, St. Louis, MO).
The immature bone marrow-derived dendritic cell line JAWS (ATCC) was used for sub-cellular fractionation. One hundred million cells were used for each experimental condition. Cells (4 × 106 cells/ml) were incubated overnight with 80 μg/ml of FITC-labeled BSA or casein in presence or absence of Clq (20 μM) or Meth (50–100 μM). Cells were then lysed in 250 mM sucrose, 1 mM EDTA pH 7.4. Early and late endosomes and lysosomes were prepared over consecutive Percoll gradients (27% and 10%) from cells treated as reported above [57]. Each fraction was tested for β−hexosaminidase to locate lysosomes and late endosomes. The late endocytic marker Lamp-1 (clone 1D4B, BD Pharmingen, San Diego, CA) and the early endosomes/plasma membrane marker transferrin receptor (TrfR) (clone M-A712, BD Pharmingen, San Diego, CA) were also used to assess the purity of the endosomal preparations. Pulled fractions 3–6 from the 27% Percoll gradient (lysosomes), 2–5 from the 10% Percoll gradient (late endosomes) and 7–10 from the 10% Percoll gradient (early endosomes) were run on SDS-PAGE and blotted membranes analyzed for FITC-labeled antigens as reported above.
GM-CSF differentiated bone marrow dendritic cells were cultured in methionine- and cysteine-free medium complete DMEM media containing 5% dialyzed serum for 1 h. Cells were then labeled with 0.2 mCi/ml [35S]-methionine (Perkin Elmer, Waltham, MA) for 30 minutes (pulse). Cells were then washed three times and incubated in complete DMEM media supplemented with 10X cold methionine for 4 hours (chase) in the presence or absence of 50 μM Meth. Cells were subsequently lysed in 1% NP40, 150 mM NaCl, 50 mM Tris containing a cocktail of protease inhibitors (Complete Mini, Roche Diagnostics, Indianapolis, IN) for 30 min on ice, spun at 14000 rpm for 30 minutes to remove cell nuclei and debris. The amount of incorporated radioactivity in each sample was determined by precipitating 10 μl of the post-nuclear supernatants with 10% trichloroacetic acid (TCA). Equivalent amounts of radioactive lysates were pre-cleared with rat serum adsorbed to Prot G beads followed by Protein G beads alone, for 2 hours at 4°C. Immunoprecipitation was performed using 10 μg of anti CD74 (clone In-1, Pharmingen, San Diego, CA) bound to Protein G beads. The beads were washed 3 times with lysis buffer and eluted with sample buffer. The elute was boiled and resolved by SDS-PAGE. The gel was subsequently dried and exposed in autoradiography.
Bone marrow-derived dendritic cells from OT II transgenic mice (Jackson Laboratory, Bar Harbor, MN) were grown in 10 ng/ml of mouse GM-CSF for 10–12 days. Splenic T cells were purified using the pan-T cell isolation kit (Miltenyi Biotec, Auburn, CA) according to the manufacturer's suggestions. One hundred thousand dendritic cells were cultured with 4 × 105 T cells in the presence or absence of 0, 3, 10, and 30 μM OVA protein for 3 days at 37°C. In some experiments Meth or Clq was added on the first day of culture until the end of the proliferative response. In other experiments, dendritic cells were pretreated with Meth or Clq for 4 h at 37°C, washed and fixed in 1% paraformaldehyde before adding OVA 323–339 peptide and the T cells. In all experiments, [3H]-thymidine (1 μCi/well) was added during the last 18 h of incubation to assay T cell proliferation. Plates were harvested and the DNA [3H]-thymidine incorporation was monitored using a Wallac liquid scintillation counter (Perkin Elmer, Waltham, MA).
Bone marrow-derived dendritic cells and macrophages from GFP-LC3 animals were plated at 4 × 105 cells/cm2 density in glass bottom dishes and incubated overnight. Meth or Clq was added to the media, and the cells incubated for 2 h and 24 h before fluorescence microscopy using an Olympus IX81 microscope with Photometrics CoolSNAP HQ cooled camera, MetaMorph Version 6.1r6 imaging software (Molecular Devices, Sunnyvale, CA), TC-324B Automatic Temperature Controller (Warner Instrument Corporation, Hamden, CT), Olympus PlanApo 60x/1.4 Oil objective, and a Chroma FITC 41001 (HQ480/40x excitation, Q505LP dichroic splitter and HQ535/50m emission) (Olympus, Center Valley, PA) to determine GFP-LC3 puncta formation in the cells. At least 100 cell profiles per dish were assayed in triplicate for GFP-LC3 puncta.
Mouse peritoneal macrophages at 3.5 × 105 cells/cm2 density on cover slips were incubated at 37°C with 20 μl of IgG opsinized sheep erythrocytes in 520 μl volume for 90 min. Uningested erythrocytes were lysed by sequential washing with PBS, water, and PBS. The phagocytic index was quantified by measuring the number of erythrocytes phagocytosed per 100 macrophages using bright field microscopy and 20x magnification, and inhibition is identified as percent of control. Data were collected from 4–7 independent experiments. The average control phagocytic index was 433.
To determine the phagocytic index for Candida albicans (Ca) and Cryptococcus neoformans (Cn) the macrophage like-cell line J774.16 (cultured in DMEM with 10% heat-inactivated FCS, 10% NCTC-109 medium, and 1% nonessential amino acids) was treated with 10 or 50 μM Methfor 2 h and then washed three times in media. As controls, J774.16 cells were incubated in medium alone or in the presence of 20 μM chloroquine. Ca strain SC5314 yeast cells were grown in brain heart infusion medium (BHI) at 37°C for 24 h then washed three times in PBS prior to application to macrophage. Cn yeast strain H99 was grown for 24 h in BHI at 37°C, incubated with capsule-specific monoclonal antibody (mAb) 18B7 [58] as an opsonin at 10 μg/ml for 2 h and then washed three times in PBS prior to incubation with macrophage. Ca and Cn cells were added to the macrophage monolayer at an effector to target ratio 1:1, and the suspension was incubated at 37°C for 30 min with Ca or 1 h with Cn. After incubation, remaining extracellular yeast cells were removed with three washes of PBS. The phagocytic index was determined by microscopic examination. For each experiment, five fields in each well were counted, and at least 100 macrophages were analyzed in each well. Wells were performed in triplicate for each condition examined.
Colony counts were made to determine the number of viable Ca and Cn yeast cells after phagocytosis. For the colony forming unit (CFU) determination, J774.16 macrophages were treated with or without Methor chloroquine and infected with Ca or Cn as described. The cultures were washed after 2 h to remove extracellular yeast and then incubated for an additional 22 h. After the 24 h total incubation, macrophage cells were lysed by forcibly pulling the culture through a 27-gauge needle 5 times. The lysates were serially diluted, and plated on Sabouraud dextrose agar at 37°C. CFU determinations were made after 72 h. Controls also consisted of yeast grown without macrophage, but in the presence of chloroquine or methamphetamine. All tests were preformed in triplicate. |
10.1371/journal.pntd.0000808 | A Simple Colorimetric Assay for Specific Detection of Glutathione-S Transferase Activity Associated with DDT Resistance in Mosquitoes | Insecticide-based methods represent the most effective means of blocking the transmission of vector borne diseases. However, insecticide resistance poses a serious threat and there is a need for tools, such as diagnostic tests for resistance detection, that will improve the sustainability of control interventions. The development of such tools for metabolism-based resistance in mosquito vectors lags behind those for target site resistance mutations.
We have developed and validated a simple colorimetric assay for the detection of Epsilon class Glutathione transferases (GST)-based DDT resistance in mosquito species, such as Aedes aegypti, the major vector of dengue and yellow fever worldwide. The colorimetric assay is based on the specific alkyl transferase activity of Epsilon GSTs for the haloalkene substrate iodoethane, which produces a dark blue colour highly correlated with AaGSTE2-2-overexpression in individual mosquitoes. The colour can be measured visually and spectrophotometrically.
The novel assay is substantially more sensitive compared to the gold standard CDNB assay and allows the discrimination of moderate resistance phenotypes. We anticipate that it will have direct application in routine vector monitoring as a resistance indicator and possibly an important impact on disease vector control.
| Aedes mosquitoes transmit many human viral pathogens including dengue, yellow fever and chikungunya. Most of these pathogens have no specific treatment or vaccine and hence their control is reliant on controlling the mosquito vectors, which usually involves the use of insecticides. In order to prevent the alarming prospect of mosquito control failure due to the rapid selection and spread of insecticide resistance in several mosquito populations worldwide, it is essential that effective resistance management strategies are implemented and adhered to. The development of simple diagnostic tests for the early identification and monitoring of resistance is an important prerequisite for this task. Here, we describe the development of a simple colorimetric test for the detection of GSTE2-2/DDTase-based resistance in individual mosquitoes. The novel assay combines the most desirable features of specificity and sensitivity with the low cost and ease of use required for a routine test in endemic countries. It can have direct application in routine vector monitoring as a resistance indicator and help improve the sustainability of insecticide based control strategies.
| Prevention of mosquito-borne diseases depends in large part on vector control and usually involves the use of insecticides. Insecticide-based methods include insecticide-impregnated bed nets, indoor or aerial sprays and water treatments. Pyrethroids and the organochlorinated insecticide DDT (1,1,1-dichloro-2,2-bis(p-chlorophenyl)ethylene) are the preferred choice for Indoor Residual Spraying (IRS) and have been used extensively for many decades for the control of disease vectors. Despite environmental concerns, DDT remains one of the cheapest and most effective long-term weapons against vector borne diseases in several stable endemic areas [1]. Although wide scale insecticide implementation has led to impressive decreases in vector borne disease transmission, the emergence and spread of insecticide resistance poses a serious threat and there is a need for new tools that will improve the sustainability of current control interventions [2]. Understanding resistance mechanisms and developing simple diagnostic tests for the early detection of insecticide resistance is an important prerequisite for the application of resistance management strategies.
Insecticide resistance in disease vectors has been attributed to increased rates of insecticide detoxification or mutations in the target sites [3]. Increased rates of glutathione transferase (GST) - mediated DDT dehydrochlorination confers resistance to DDT in several mosquito species, such as Aedes aegypti, the major vector of dengue and yellow fever worldwide, and Anopheles gambiae, the major malaria vector in sub-saharan Africa [4], [5]. This DDT detoxification reaction is catalysed by the Epsilon class GST, GSTE2-2 in An. gambiae, An. cracens and Ae. aegypti mosquitoes from different geographical origins [4], [5], [6].
Detection of metabolism – based insecticide resistance is more complex than screening for specific mutations known to cause target site resistance. Current techniques for measuring elevated GSTE2-2 levels in mosquitoes, such as real time PCR or specific ELISA based on antibodies are elaborate or require the use of expensive equipment and consumables, and are therefore not accessible to laboratories on a limited budget [4], [5].
Biochemical assays for detecting metabolic resistance generally employ generic substrates that are recognised by most or all members of the enzyme families. For example, GST activity is usually measured using 1-chloro-2,4-dinitrobenzene (CDNB), 1,2-dichloro-4-nitrobenzene (DCNB), and, more recently, monoclorobimane [7], [8]. Unlike assays to detect elevated esterase activity, which can be read by eye, the current GST assays require a spectrophotometer that can measure absorbance in the UV range, or fluorimeter with multiple emission/excitation channels [8], [9], limiting their applicability in the field. Potentially greater sensitivity and specificity could be achieved if substrates that were specifically recognised by the enzyme(s) responsible for insecticide metabolism were employed.
A colorimetric assay for GSTs with alkyl transferase activity, capable of catalysing the release of iodine from haloalkene substrates, has been recently described [10]. Using a modified version of this assay, Dowd et al. [11] screened a large number of recombinant mosquito GSTs for alkyltransferase activity with several haloalkene substrates, to identify potential enzyme biosensors for detecting insecticides. Recombinant epsilon GSTs, but not the delta or sigma GSTs, which are the most abundant in insects [12], showed a remarkable ability to utilise iodoethane as a substrate and produce a dark blue colour, which can be measured spectrophotometrically or visually [11].
Here, we have adapted the alkyl transferase/iodoethane -based colorimetric assay to measure GST activity associated with DDT resistance in individual Ae. aegypti mosquitoes.
Six Ae. aegypti mosquito strains were used in this study: The standard laboratory reference strain (New Orleans) was kindly provided by the Center for Disease Control and Prevention (CDC), Atlanta, USA, the susceptible Ivory Coast strain was collected from Cote d'Ivoire, the Iquitos strain originating from Peru and the Solidaridad, Isla Mujeres and Merida strains, from Mexico, were kindly provided by Prof. William Black (Colorado State University, USA). All strains were reared under standard conditions (28°C±2°C, 80% RH) at the Liverpool School of Tropical Medicine. Bioassays were performed on 1–3 day old adults using the World Health Organization (WHO) adult susceptibility test papers – DDT 4% [7]. The time causing 50% mortality (LT50) was obtained 24h after the exposure.
Cloning into a pET3a vector, expression in Escherichia coli BL21(DE3) plysS, and purification of Ae. aegypti recombinant Epsilon GSTs were conducted as described previously [5], [13]. The eluted enzyme was concentrated using a Vivaspin 15R concentrator and exchanged using a PD-10 column into 50 mM sodium potassium phosphate (pH 7.4), 10 mM dithiothreitol, and 40% glycerol according to the manufacturer's instructions and samples were stored at −80°C, until used.
Mosquitoes were homogenised in 0.1M Tris-HCl, pH 8.2 (20 µl per individual), the mixture was centrifuged at 14,000×g for 10 min at 4°C, and the supernatant was used as the enzyme source for the biochemical assays. Standard GST spectrophotometric assays were performed by monitoring the formation of the conjugate of CDNB or 1, 2-dichloro-4-nitrobenzene (DCNB), and reduced glutathione (GSH) [9]. The iodide-releasing reaction was carried out as previously described [10] and optimised by Dowd et al. [11], with GSH (2.5 mM) and iodoethane (2.5 mM) in 0.1M phosphate buffer pH 8.2 and enzyme source in a total volume of 100 µl at 25°C. The reaction was incubated at 30 min, or for different periods of time depending on the reaction rate studied during optimisation stages. Blue colour developed immediately after addition of 50 µl starch solution (0.25 g partially hydrolysed potato starch in 25 ml of Milli-Qwater and boiled in a microwave oven until all starch has dissolved) and 100 µl acidified peroxide solution (2% H2O2 in 2 mM HCl). The blue colour was quantified spectrophotometrically at 610 nm using a VERSAmaxTM microplate spectrophotometer (Molecular Devices, Sunnyvale, CA, USA), or estimated visually by eye. A standard curve was prepared from different concentrations of KI in 0.1M Tris–HCl buffer, pH 8.2. Specific activities towards iodoethane were calculated from the linear range of the enzymatic reaction, and a plot of absorbance at 610 nm against potassium iodide concentration. They are expressed as µmole of iodide released /min/mg. All measurements were made in triplicate. Protein concentrations were measured using Bio-Rad protein assay reagent with bovine serum albumin as the protein standard [14].
Mosquito extracts (0.060 mg total protein) were analysed with SDS-polyacrylamide gel electrophoresis (10% acrylamide running gel and 4% acrylamide stacking gel) and electroblotted onto polyvinylidene difluoride membrane. The membrane was probed for 2 hours with an anti-AaGSTE2-2 antibody at 1∶1000 dilution in 3% milk-PBS-Tween solution and for 1 hour with a peroxidase-labelled anti-rabbit antibody at 1∶10000 dilution. Immunoreactive proteins were visualised using a horseradish peroxidase sensitive ECL chemiluminescent Western blotting kit (GE Healthcare).
We recently showed that, unlike other GST members tested, epsilon GSTs can very efficiently utilise the haloalkene iodoethane as a substrate [11]. In order to determine the amount of mosquito protein required to measure GST activity in the visual range (colour change) and set the linear limits of the colorimetric assay, we tested different amounts of mosquito extracts (0.010–0.120 mg of total protein) at time points between 5 and 60 min. The minimum amount of protein extract that gave a visible colour range in any mosquito strain, after 30min incubation period, was 0.030 mg (Figure 1A). No visible colour change was observed for the reference susceptible strain New Orleans, even when much higher amounts of protein (and longer incubation times up to 60 min, data not shown) were included in the assay (Figure 1A).
The product/colour formation is linear for at least 30 min, when approximately 0.060 mg mosquito homogenate (equivalent to ¾ of an individual Ae. aegypti female) was assayed (Figure 1B). The linear range of the reaction was not affected by temperature fluctuations between 25 and 35°C (data not shown).
The LT50 values of six Ae. aegypti mosquito strains following exposure in 4% DDT were determined (Figure 2A). The susceptible New Orleans and Ivory Coast strains showed LT50 values of 20 min or less, whilst the Iquitos and Solidaridad strains exhibited LT50 values of 76 min and 100 min, respectively; accurate LT50 values could not be determined for Merida and Isla Mujeres strains, due to the very high levels of resistance (LT50>300 min). To confirm the association of the AaGSTE2-2 enzyme with the resistance phenotype, we performed Western blot analysis, using crude mosquito homogenates probed with anti-AaGSTE2-2 antiserum. A single band of approximately 25 kDa was detected in all strains, with intensity levels highly correlated with the LT50 values/DDT resistance data (Figure 2B).
Using the optimised colorimetric assay, we determined the specific GST activity in adult females from several Ae. aegypti strains. As shown in Table 1, there is a >15-fold difference in alkyl transferase activity between the highly DDT resistant Merida and Isla Mujeres strains, and the susceptible Ivory Coast strain. The alkyl transferase activity of the Iquitos and Solidaridad strains, which showed moderate resistance levels, was 4- and 12-fold higher, respectively, compared with the Ivory Coast strain (Table 1). A highly significant correlation was observed between the LT50s and the enzymatic activities obtained by the iodoethane/colorimetric assay (R2 = 0.97, P<0.01). The difference in alkyltransferase activity between the different strains can be easily visualised by eye (Figure 2C), via the effort of multiple individuals.
This correlation between resistance phenotype and specific activity does not hold for the model substrates CDNB and DCNB. For CDNB there was significantly higher activity in the four resistant strains compared to the two susceptible stains (Table 1) but no difference in activity between the moderately and highly resistant groups. For DCNB, the relationship was even less clear. For example no significant difference was observed between the moderate resistant strain Solidaridad and the New Orleans susceptible strain (Table 1).
By screening a large number of recombinant mosquito GSTs for alkyltransferase activity with several substrates, Dowd et al. [11] showed that mosquito epsilon GSTs, AaGSTE2-2 and AaGSTE4-4, can utilise the haloalkene iodoethane as substrate but that this substrate was not recognised by delta or sigma class GSTs. To determine whether the ability to catalyse the release of iodine from iodoethane was a general property of epsilon GSTs, we expressed six family members and measured their specific activity against this substrate. As shown in Table 2, the highest activity was obtained with the DDTase AaGSTE2-2 (10.3 µmole iodide/min/mg). Other members of the Epsilon class also recognised this substrate but their specific activities were lower (0.03 to 4.3 µmole iodide/min/mg). AaGSTE8-8 exhibited the lowest activity, possibly due to the low amino acid identity (approximately 30%) of this gene with other members of this class [15]. The respective CDNB activities of the recombinant epsilon GSTs are also shown in Table 2 for comparison. The DDTase AaGSTE2-2 has lower or similar specific activity with CDNB compared to other members of the family and hence this substrate cannot specifically recognise GSTs implicated in insecticide resistance.
We have developed a simple colorimetric assay for the specific detection of GST activity associated with DDT resistance in Ae. aegypti. The colorimetric assay is substantially more sensitive in detecting DDT resistance in Ae. aegypti, compared to the gold standard CDNB assay currently being used in routine mosquito resistance monitoring studies [7]. The differences in GST activities among strains with high, moderate or negligible resistance were over 15-fold for iodoethane, but only 1.5–4.3-fold for CDNB and DCNB substrates. In contrast to iodoethane, the latter general substrates failed to discriminate moderate resistance phenotypes (Table 1). This increased sensitivity of the novel colorimetric assay provides greater potential for the identification of resistance at early stages, a crucial pre-requisite for the implementation of evidence-based resistance management tactics.
Unlike UV/spectrophotometric CDNB and DCNB assays, the alkyltransferase/iodoethane assay produces a dark blue colour that is both highly correlated with AaGSTE2-2-overexpression-based DDT resistance and can be estimated by eye at least semi-quantitatively (Figure 2). This novel assay can be performed by non-qualified personnel, without sophisticated equipment. It is robust at temperatures between 25–35°C, with a wide linear range of quantification, and a sensitivity which allows the measurement of GST activity in a single mosquito. The cost of the assay is less than 0.05 USD per mosquito, while the shelf life of the substrate iodoethane is at least 1 year at 4°C.
Here, we have focused on Ae. aegypti, as DDT resistance is extremely high in many populations of this species in dengue endemic regions [16]. However, the assay can be adapted for measuring GSTE2-2/DDTase – based DDT resistance in other mosquito species, such as the major malaria vector An. gambiae. This was not tested here, as there were no suitable resistant strains available. Nevertheless, given that iodoethane is a very good substrate also for the orthologue enzyme AgGSTE2-2 (data not shown) and this enzyme is the key enzyme responsible for DDT resistance in this species [4], there is no reason to believe that this assay will not work for Anopheles mosquitoes too.
In conclusion, we describe a simple colorimetric test for the detection of the GSTE2-2/DDTase- based resistance in mosquitoes. It combines the most desirable features of specificity and sensitivity with the low cost and ease of use required for a routine test in endemic countries. We anticipate that the assay will have direct application in routine vector monitoring as a resistance indicator and help improve the sustainability of insecticide based control strategies.
|
10.1371/journal.pgen.1002502 | Computational Prediction and Molecular Characterization of an Oomycete Effector and the Cognate Arabidopsis Resistance Gene | Hyaloperonospora arabidopsidis (Hpa) is an obligate biotroph oomycete pathogen of the model plant Arabidopsis thaliana and contains a large set of effector proteins that are translocated to the host to exert virulence functions or trigger immune responses. These effectors are characterized by conserved amino-terminal translocation sequences and highly divergent carboxyl-terminal functional domains. The availability of the Hpa genome sequence allowed the computational prediction of effectors and the development of effector delivery systems enabled validation of the predicted effectors in Arabidopsis. In this study, we identified a novel effector ATR39-1 by computational methods, which was found to trigger a resistance response in the Arabidopsis ecotype Weiningen (Wei-0). The allelic variant of this effector, ATR39-2, is not recognized, and two amino acid residues were identified and shown to be critical for this loss of recognition. The resistance protein responsible for recognition of the ATR39-1 effector in Arabidopsis is RPP39 and was identified by map-based cloning. RPP39 is a member of the CC-NBS-LRR family of resistance proteins and requires the signaling gene NDR1 for full activity. Recognition of ATR39-1 in Wei-0 does not inhibit growth of Hpa strains expressing the effector, suggesting complex mechanisms of pathogen evasion of recognition, and is similar to what has been shown in several other cases of plant-oomycete interactions. Identification of this resistance gene/effector pair adds to our knowledge of plant resistance mechanisms and provides the basis for further functional analyses.
| Oomycete plant pathogens are among the most devastating agricultural pests and employ arsenals of effector proteins to manipulate their plant hosts. Some of these effectors, however, are recognized in the plant and trigger an immune response. Hyaloperonospora arabidopsidis (Hpa) causes downy mildew on the model plant Arabidopsis thaliana and this interaction has been developed as a model system for oomycete pathogenesis. Here, we employ computational predictions to identify a novel effector ATR39-1, which is highly conserved among different Hpa isolates. A two amino acid-insertion in the alternative allele ATR39-2 correlated with evasion of recognition. We identified the corresponding resistance gene RPP39 and found that the signaling gene NDR1 is required to establish full resistance. Recognition of ATR39-1 by RPP39 in the plant did not inhibit growth of the oomycete, suggesting that complex mechanisms exist to prevent effector recognition. Knowledge of such novel resistance interactions provides the backbone of our understanding of plant resistance mechanisms and will aid in the further dissection of plant immunity.
| Oomycetes comprise a number of agriculturally important plant pathogens, including Phytophthora infestans (potato and tomato late blight), P. ramorum (sudden oak death) and P. sojae (soybean root rot). Hyaloperonospora arabidopsidis (Hpa, downy mildew, formerly known as Peronospora parasitica) is a naturally occurring oomycete pathogen of the model plant Arabidopsis thaliana. The Hpa/Arabidopsis pathosystem allows the scientific community to take advantage of the genetic tools developed for Arabidopsis in the dissection of plant responses to oomycetes [1]. During their parasitic life stages, oomycetes deliver an arsenal of effector proteins to their plant host, which are hypothesized to target basal defense mechanisms and/or manipulate host metabolism to extract nutrients for the pathogen [2]. The first oomycete effector proteins were, however, identified based on their avirulence functions, i.e. their presence triggered an immune response in the host resulting in resistance to the pathogen. This so-called effector triggered immunity (ETI) is characterized by the specific recognition of pathogen avirulence effectors by plant resistance receptors, either directly or indirectly [3]. ETI is often accompanied by localized cell death at the site of infection, the hypersensitive response (HR), which limits the spread of biotrophic pathogens inside the plant host [4]. The Hpa effectors Arabidopsis thaliana recognized 1 (ATR1), ATR13 and ATR5 as well as Avr1b from P. sojae were identified using classic genetic crosses between virulent and avirulent pathogen strains [5]–[8], and have been shown to be under strong positive selection. P. infestans effector Avr3a, on the other hand, was isolated using association genetics [9]. Interestingly, despite having no sequence homology, Avr3a and ATR1 reside in conserved syntenic regions within the genome [9] and their three-dimensional structures reveal a similar fold between Avr3a and a sub-domain of ATR1 [10]–[13].
All currently described oomycete effectors were found to have a modular domain structure, containing amino-terminal domains involved in effector translocation and carboxyl-terminal effector domains. The translocation domains typically include a secretion signal sequence followed by the amino acid motif Arg-x-Leu-Arg (RxLR), in which x could be any amino acid. The RxLR motif was shown to be important in translocation of effectors to the host cytoplasm and it is also functionally interchangeable with the translocation motif of Plasmodium falciparum effectors [14], [15]. The absence of a canonical RxLR motif in the recently cloned effector ATR5 suggests that other sequences may also be involved in translocation of effectors [8]. The carboxyl-terminal effector domains are highly divergent and typically do not have strong sequence similarity to other proteins.
The genomes of several Phytophthora species as well as of Hpa strain Emoy2 have recently been sequenced and this has fueled bioinformatics efforts to elucidate the complete arsenal of effectors [16]–[18]. Effector predictions were based on the presence of the amino-terminal translocation domains. While bacterial pathogens such as Pseudomonas syringae contain around 30 to 40 effector genes [19], the oomycete genomes were found to contain expanded effector repertoires, ranging from around 350 in P. sojae and P. ramorum to more than 700 in P. infestans [20]. Initial effector predictions from the Hpa genome yielded 149 RxLR effector genes [21], however, the published genome sequence for Hpa strain Emoy2 contains only 134 annotated RxLR effectors [17]. Currently, major efforts are being undertaken in dissecting the effector complement of several oomycetes in order to identify novel avirulence determinants as well as to define effector virulence functions. Several recent large-scale effector screens in different oomycetes focused on the localization of effectors in the plant cell during infection and on their roles in facilitating oomycete infections [22], [23]. These transgenic approaches identified effectors that localize to the oomycete haustorial feeding structures and may be important in mediating intercellular communication. Another study employed mining of expressed sequence tags (ESTs) to identify genes highly expressed during Hpa infection, and investigated their potential functions during compatible interactions [24]. An in planta expression screen of P. infestans effectors was successful in identifying the cognate avirulent effectors ipiO/AVRblb1 and AVRblb2, recognized by the R proteins Rpi-blb1 and Rpi-blb2, respectively [25], [26].
Resistance to different strains of Hpa was mapped to a number of RPP (Resistance to Peronospora parasitica) loci in several Arabidopsis ecotypes [27], [28]. Six of these R genes were subsequently cloned, but the corresponding recognized effectors have only been identified for three of them, RPP1, RPP13 and recently RPP5, which recognize ATR1, ATR13 and ATR5 respectively [5], [7], [8]. Similar to the R genes that function against other microbial pathogens, RPP genes belong to the large Nucleotide Binding Site-Leucine Rich Repeat (NBS-LRR) gene family in Arabidopsis, which comprises a total of around 150 members, but only a few with an assigned function [29]. Characterized R genes confer resistance to various classes of pathogens including oomycetes, bacteria, fungi and viruses. Additionally, NBS-LRR genes have been implicated in non-self recognition in inter-accession hybrids [30].
Research on RPP1/ATR1 and RPP13/ATR13 has greatly advanced our understanding of effector recognition and resistance signaling. Recognized ATR1 alleles have been shown to associate in planta with the LRR-domain of RPP1 before triggering an immune response [31]. Intracellular recognition of ATR13 by the CC-NBS-LRR protein RPP13 was shown to signal independently of the known signaling genes EDS1 and NDR1, indicating the presence of additional signaling pathways activated upon effector recognition [32].
In order to gain a better understanding of the interactions between Hpa and its host A. thaliana, we set out to screen 83 Arabidopsis ecotypes for novel recognition specificities with a subset of predicted Hpa effectors. Here, we describe a successful approach to mine the Hpa genome for functional effector proteins based on domain structure similarity to known oomycete RxLR effectors. We identified a novel avirulent RxLR effector, ATR39, which is recognized by the Arabidopsis ecotype Weiningen. Comparison of ATR39 alleles identified two amino acids that are critical for recognition. We cloned the corresponding R gene, RPP39, and showed that it is a member of a small cluster of CC-NBS-LRR genes and requires NDR1 for downstream signaling of plant defense responses. Our ability to combine computational predictions with molecular and genetic techniques will facilitate the rapid identification of novel R genes as well as inform our understanding of the evolution of pathogenesis and resistance.
Based on characterized effectors from Phytophthora sp. and Hpa, intracellular oomycete effector proteins are predicted to contain several conserved domains: an N-terminal secretion signal peptide (SP), a central RxLR motif, and a C-terminal variable effector domain. Previously, Win et al. mined the Hpa genome (version 7.0) for predicted open reading frames of >70 amino acids, which contained the N-terminal SP and the RxLR motif between amino acids 30 and 60, and found 149 effectors fulfilling these criteria [21]. In order to refine this search we generated a Hidden Markov Model (HMM) from the N-terminal conserved domains (SP and RxLR) of previously identified effectors and their homologues (this set includes 43 proteins, [21], Figure 1A). HMMs are widely used to predict homologies with statistical significance, most prominently in the protein domain database Pfam [33]. This method allowed us to screen the 149 initially predicted effectors with the HMM model and prioritize them for experimental validation, as outlined in Figure 1B. We decided to focus downstream characterization on the 18 highest-scoring predicted effectors, with E-value scores <0.001 (Figure 1C). Amino acid and nucleotide sequences for these effectors are available in fasta format online as File S1 and S2, respectively. Interestingly, some of the predicted effectors scored even higher than two known effectors, ATR1 (Hp_Contig137.3_F55) and ATR13 (Hp_Contig1514.4_F2). We tested whether the predicted effectors were expressed during Hpa infection by RT-PCR and confirmed expression for 15 of them at seven days post-inoculation (Figure 1D).
Because of Hpa's obligate biotroph lifestyle, studies on Hpa have relied on surrogate systems, delivering oomycete effectors biolistically [34] or using bacterial or viral vectors [35], [36].
We PCR amplified the highest scoring expressed effectors from the Emoy2 isolate of Hpa, past the predicted signal peptide cleavage site, and cloned them into two Pseudomonas expression vectors. First, the effectors were shuttled into a Gateway-compatible Pseudomonas vector as C-terminal fusions with the AvrRpm1 type three secretion system (TTSS) signal peptide (pPsSP, [35]). However, we observed that ATR1 was not functional in this system, but was functional as a C-terminal fusion with the AvrRps4 TTSS signal peptide in the alternative pEDV3 system [36]. We thus decided to also subclone the predicted effectors into pEDV3 and test them in both expression systems. We conjugated these expression plasmids into Pseudomonas fluorescens (Pf0), a non-pathogenic Pseudomonas strain that lacks an endogenous TTSS and effectors and was engineered to express the Pseudomonas syringae pv. tomato (Pst) DC3000 hrp cluster and TTSS [37]. Using this strain allowed us to deliver individual effectors to the plant host, circumventing considerable background we often observed when inoculating a variety of ecotypes with Pst DC3000 (data not shown). We did not observe a reaction to Pf0 carrying an empty vector control in most ecotypes, even after 48 hours post-inoculation (hpi). Effector recognition on the other hand resulted in a visible hypersensitive response within 24 hpi (Figure 2A). We then screened 83 Arabidopsis ecotypes from the Nordborg collection (1, [38]) with Pf0 expressing each of the predicted effectors.
One of the predicted effectors, Hp_Contig399.11_F1, triggered a visible HR in the ecotype Weiningen (Wei-0), when delivered by Pf0 (Figure 2A). Hp_Contig399.11_F1 has an HMM score of 9.3 and ranks number 16 of the predicted effectors (Figure 1C). None of the other 82 available ecotypes from the Nordborg collection displayed an HR upon delivery of Hp_Contig399.11_F1. According to nomenclature previously applied to Hpa effectors, we renamed this effector ATR39-1 (for A. thaliana recognized 39-1, accession number JQ045572).
In order to assess whether ATR39-1 conferred avirulence to pathogenic bacteria, we conducted bacterial growth assays but found that Wei-0 showed natural resistance towards several pathogenic Pseudomonas strains, including Pst DC3000 (Figure 2B) and P. syringae pv. maculicola (Psm) ES4326 (Figure S1). We therefore generated an introgression line in which the Wei-0 recognition locus was introduced into the Col-0 background by repeated backcrossing. Using this line (WC-5BX), we showed that ATR39-1 is indeed recognized by the Wei-0 locus and that this recognition results in decreased growth of Pst DC3000 (Figure 2B).
Unlike ATR1, ATR39-1 does not contain the acidic DEER (Asp-Glu-Glu-Arg) motif following the RxLR translocation motif. We also tested whether the effector-truncation past the RxLR motif triggers the hypersensitive response and found that the effector domain (ATR39-1 Δ48) is also able to trigger an HR when delivered by Pf0 (Figure S2).
Using the same primers we amplified two alleles of ATR39 from Hpa strain Emoy2, but only one of them, ATR39-1, is recognized by Wei-0 (Figure 2). Both alleles are predicted in our HMM search with scores of 9.3 and 8.8 for ATR39-1 and ATR39-2, respectively. The two alleles differ by 10 nucleotides, resulting in 9 amino acid substitutions, and a 2 amino acid insertion in ATR39-2 relative to ATR39-1 (Figure 3A). In the published Hpa Emoy2 genome assembly ATR39-1 is not annotated, however, its allelic variant ATR39-2 is annotated as HaRxL48 [17].
We have generated Illumina paired-end sequencing data for the Emwa1 isolate of Hpa and could identify both alleles in the assembly (data not shown). In this assembly we do not see signatures of duplication events, suggesting that Hpa Emwa1 and possibly also Emoy2 and Noco2 are heterozygous at this locus.
Using allele-specific primers we amplified ATR39 alleles from seven isolates of Hpa and found that ATR39-1 is only present and expressed in Emoy2, Emwa1 and Noco2 isolates (Figure 3B). ATR39-2, on the other hand, is more prevalent in this set of Hpa isolates. Surprisingly, we could not amplify ATR39-2 from the Emoy2 isolate currently grown in our lab. Since we initially amplified the effectors from a DNA sample obtained from a different source than the Hpa Emoy2 strain, and there is evidence for heterozygosity in the various lab strains, it is possible that our current lab strain is now homozygous for ATR39-1. In order to determine whether this strain is Emoy2, we amplified and sequenced the ATR1 effector and verified that our lab strain is indeed Emoy2 (data not shown).
We sequenced the ATR39 amplification products and found no nucleotide polymorphisms among ATR39-1 or ATR39-2 alleles from the different Hpa isolates. This conservation is in stark contrast with other characterized Hpa effectors ATR1, ATR5 and ATR13, which are highly divergent [5], [8], [39]. These findings indicate that the two ATR39 alleles may have an important function in Hpa and may be maintained under strong balancing selection. Sequence comparisons and pattern searches yielded no obvious homologs or putative function for ATR39. Taken together, we have identified a novel effector from Hpa with unknown function that is able to trigger a resistance response in Arabidopsis.
The Wei-0 ecotype was previously shown to be susceptible to multiple Hpa isolates [40]. We confirmed that despite being able to recognize ATR39-1 present in Emoy2, Emwa1 and Noco2, Wei-0 still supports growth of these isolates (Figure 3C). This lack of resistance is not due to the lack of ATR39-1 transcript since we detected ATR39-1 expression in infected tissue using RT-PCR (Figure 3B and Figure S3). A similar suppressed recognition phenotype was observed for one of the ATR1 alleles. The ATR1Emco5 allele is recognized in several Arabidopsis ecotypes, which remain susceptible to infection by Hpa Emco5 [5], [41].
ATR39-1 and ATR39-2 differ by 9 non-synonimous substitutions, and a two amino acid insertion in ATR39-2. In order to define the region in ATR39 responsible for differential recognition of the two alleles, we generated ATR39-1in by inserting E168/V169 into ATR39-1 using site-directed mutagenesis. Similarly, we generated ATR39-2del, in which E168/V169 were deleted (Figure 4A). ATR39-2del, but not ATR39-1in triggered an HR in Arabidopsis Wei-0 suggesting that the presence of amino acids E168/V169 blocks recognition of ATR39 (Figure 4B). Additionally, we performed Pst DC3000 growth assays and found that ATR39-2del restricted bacterial growth, whereas Pst DC3000 expressing ATR39-1in grew to similar levels as the non-recognized allele ATR39-2 or empty vector control (Figure 4C). These findings suggest that amino acids E168/V169 in ATR39-2 are critical in evading recognition by the cognate R protein.
Resistance to Hpa strains is mediated by a number of RPP (Resistance to Peronospora parasitica) loci. In order to identify the RPP39 gene responsible for recognition of ATR39-1 we generated a cross between Wei-0 and Col-0 and found that recognition segregated as a single dominant locus in the F2 progeny. Using 886 F2 plants, we delineated the RPP39 locus to a 150 kilobase region on the bottom of chromosome 1. In Col-0 this region contains two homologous CC-NB-LRR genes with 91% identity arranged in a tandem repeat (At1g61180 and At1g61190). Because there is no sequence information available for Wei-0, we generated a fosmid library and identified six overlapping clones that span this region (Figure 5B). We sequenced the fosmids using Illumina next generation sequencing and found several rearrangements and a transposon insertion relative to the Col-0 sequence (Figure S4). Wei-0 also contains two CC-NBS-LRR genes at this locus, which were the most promising candidates for RPP39.
We amplified genomic regions containing the two candidate R genes, R_180-Wei-0 (accession number JQ045574) and R_190-Wei-0 (accession number JQ045573) and transformed them into Arabidopsis Col-0 for complementation. Only transgenic plants containing the gene corresponding to the At1g61190 locus (R_190-Wei-0) developed an HR in response to ATR39-1 delivery, indicating that this gene is RPP39 (Figure 5D). Growth assays with Pst DC3000 expressing ATR39-1 performed on plants in the T2 generation confirmed these results as we observed reduction in bacterial growth only in plants containing the R_190-Wei-0/RPP39 transgene (Figure 5E).
The predicted RPP39 coding region contains an intron close to the C-terminus, connecting a large N-terminal exon with a short C-terminal exon encoding the last 15 amino acids (Figure 6A). We amplified RPP39 from Wei-0 cDNA and confirmed the presence of this intron. In Agrobacterium-mediated transient expression experiments in Nicotiana benthamiana we showed that the RPP39 cDNA driven by the CaMV35S promoter is able to trigger ATR39-1 dependent HR (Figure 6B). Interestingly, expression of the genomic RPP39 clone in N. benthamiana was not sufficient to trigger HR, yet it was functional in transgenic Arabidopsis. Because a 35S driven clone of the genomic sequence of RPP39 is able to trigger HR (Figure 6B), we believe that the lack of responsiveness of the genomic RPP39 clone is probably due to low expression off the native promoter in the transient assay.
RPP39 is very similar to its Col-0 paralogs, At1g61190 (NM_104800.1) and At1g61180 (NM_104799.3), with 86% identity at the nucleotide level and 81% to 87% at the amino acid level (Figure S5). The homologs are most divergent in the C-terminal LRR domain (Figure S6). The most closely related R proteins outside the RPP39 cluster are RPS5 (NP_172686.1), RFL1 (AAL65608.1) and RPS2 (AAA21874.1) with 50%, 49% and 27% identity to RPP39 at the amino acid level, respectively (Figure S5). Taken together, we identified a functional resistance gene as a member of small R gene cluster.
Non-specific disease resistance 1 (NDR1) is a common signaling gene required for CC-NBS-LRR mediated resistance functions [42]. Since RPP39 is similar to the resistance genes RPS2 and RPS5, both of which require NDR1, we investigated the involvement of NDR1 in RPP39 mediated resistance. We generated stable transgenic plants containing RPP39 or its non-functional paralog R180_Wei-0 in the ndr1 mutant background. RPP39 transgenic plants displayed no difference in their ability to trigger ATR39-1-dependent HR as compared to transgenics in Col-0 wild type background (Figure 7A). However, in Pst DC3000 growth assays we did not see a similar growth reduction in the ndr1 transgenics (Figure 7B). These results indicate that the ability of RPP39 to trigger HR is separable from complete disease resistance, and is reminiscent of RPM1, which displays similar NDR1 dependency [43].
We have used a combination of computational prediction methods and phenotypic screening to identify a novel recognized effector from Hpa. Being an obligate biotroph, Hpa is currently not amenable to in depth genetic analysis, and cloning of the previously identified RxLR effectors ATR1 and ATR13 was a lengthy and cumbersome process [5], [7]. In our approach, we mined the Hpa genome for putative effector sequences and ranked them in a comparison with a Hidden Markov Model (HMM) generated based on the N-terminal conserved translocation domains of previously identified oomycete effectors. When we generated the HMM, all characterized effectors contained an RxLR and in our initial ranking we screened the HMM only against predicted RxLR effectors. Recently, Bailey et al. cloned and characterized ATR5 from Hpa strain Emoy2, the first recognized effector without a canonical RxLR sequence, suggesting that this motif can be modified [8]. In accordance with this, when screening the HMM against all annotated proteins in the final Hpa genome release [17], we also identified several non-RxLR variants within the top-scoring effector candidates, including ATR5 (data not shown). These findings suggest that our screen is by no means exhaustive and that more recognized effectors await characterization. When screening the HMM against the published annotation of the Hpa protein database, we also noticed that a few of our predicted effectors did not appear in the results because their annotated genes do not include the N-terminal translocation domains due to a difference in annotation of the translational start site.
Our choice of using the non-pathogenic P. fluorescens that does not normally trigger a response in Arabidopsis as a surrogate expression and delivery system allowed us to rapidly screen a large number of Arabidopsis ecotypes for novel recognition specificities. Interestingly, despite the fact that Hpa strains are predicted to contain between 10 and 20 avirulent effectors, according to association studies in the 1990s [44], in our set we only identified one novel effector that was able to trigger resistance in an Arabidopsis ecotype. Under our assay conditions and using our expression vectors we did not detect consistent phenotypes (either virulence or avirulence) for most of the tested effectors from the Emoy2 isolate of Hpa. Notably, in this study we compared the two delivery systems pEDV3 and pPsSP, which fuse the TTSS signal peptides from the bacterial effectors AvrRps4 and AvrRpm1, respectively, upstream of the Hpa effector [35], [36]. Intriguingly, ATR39-1 was functional when expressed as AvrRpm1 fusion in the pPsSP vector, but not as AvrRps4 fusion in the pEDV3 vector. ATR1, on the other hand, did not trigger responses when delivered as AvrRpm1 fusion protein in the pPsSP system and is only functional in the pEDV3 system. These results suggest that the choice of expression system can greatly influence the observed phenotypes and should be taken into account. Therefore, we cannot dismiss the possibility that several of the predicted effectors might be functional in different assay conditions or recognized by different Arabidopsis ecotypes not tested in this study. In a set of Arabidopsis ecotypes from the United Kingdom, where all currently available Hpa strains originate from, several ATR13 alleles were tested and were found to be recognized by RPP13 alleles in different ecotypes [45]. The fact that we did not observe prevalence of effector recognition in our subset of Arabidopsis ecotypes suggests that the co-evolution between pathogen and host may play an important role in this obligate biotroph interaction. We also screened various alleles of ATR1 and ATR13 on the Nordborg collection of Arabidopsis ecotypes [38] and found six ecotypes capable of recognizing ATR1Emoy2, including Ws-0 and Nd-1, but none that recognized ATR13Emoy2 [41]. Compared with the prevalence of recognition specificities for bacterial effectors such as AvrRpt2 or AvrPphB, which are recognized by RPS2 and RPS5 in more than half of the tested ecotypes [38], these results indicate a more dynamic effector repertoire in Hpa.
We identified two amino acids, E168 and V169, which abrogate recognition of ATR39-2. Deletion of these amino acids in ATR39-2 resulted in a gain-of-recognition phenotype while insertion of E168/V169 in ATR39-1 lead to loss-of-recognition by RPP39. It is possible that this insertion/deletion polymorphism alters the three dimensional structure of ATR39, thus disrupting the interaction with potential target proteins. Another possibility would be that the polymorphism alters putative enzymatic properties of ATR39. However, since the primary amino acid sequence of ATR39 is not homologous to any known protein, we can only speculate about the consequences of the insertion on the function of this protein. Experiments to determine the structure and to identify interacting proteins of ATR39 alleles will help elucidate the function of this novel effector.
The Arabidopsis ecotype Wei-0 exhibits an interesting resistance pattern: it is resistant against several tested bacterial pathogens but highly susceptible towards Hpa isolates ([40], this study). Intriguingly, despite being able to recognize ATR39-1 when delivered by a surrogate system, Wei-0 is not able to restrict growth of Hpa expressing ATR39-1. This finding is reminiscent of ATR1Emco5, which is recognized in the Arabidopsis ecotype Ws-0 by RPP1-WsB, but this recognition does not abrogate growth of Hpa in a natural infection. Additional Arabidopsis ecotypes recognizing ATR1Emco5 were recently identified, and these were also shown to support growth of Hpa isolate Emco5, indicating similar mechanisms in virulence may act in these interactions [41]. These data suggest that Hpa may contain a pathogenicity factor, perhaps another secreted effector, which prevents recognition of ATR1 in Emco5 or ATR39-1 in Emwa1 and Emoy2. Support for this hypothesis comes from a study on different alleles of the P. infestans effector ipiO/Avrblb1, where expression of one allele, ipiO4, was found to suppress resistance mediated by the ipiO1/Rblb1 interaction [46]. Alternative explanations for the lack of recognition in these instances could be inhibition of effector translocation to the host, mistimed expression of either effector or R gene or incomplete resistance mediated by the R gene that is not strong enough to contain the pathogen. Experiments aimed at investigating this interesting phenotype should yield important insight into Hpa virulence.
RPP39 is a member of a small cluster of CC-NBS-LRR genes, which is rapidly evolving through duplication and inversion events. A dominant mutation in the LRR domain of an RPP39 homolog in Ws-0, uni-1D (named after the Japanese word for sea urchin because of its morphological phenotype) was found to display several defects in growth and hormone signaling which are often seen with gain of function R proteins such as SNC1 [47], [48]. Interestingly, the uni-1D phenotype was not dependent on NDR1 [47]. We found that RPP39 requires NDR1 to fully suppress P. syringae growth, but activates hypersensitive cell death independently of NDR1. In future, it will be interesting to more completely dissect the RPP39 signaling pathway.
Taken together, our results show the feasibility of employing computational predictions in the identification of functional pathogen effectors. In combination with classical genetic methods it was possible to determine function for one member of the large family of predicted R proteins in Arabidopsis. Preliminary data on several polymorphic effectors from our HMM priority list indicate that Hpa isolates other than Emoy2 contain functional/recognized effector alleles, which will be further pursued and may lead to the identification of additional R genes, the analysis of which will further advance our understanding of R protein signaling.
A set of confirmed oomycete effectors and their close homologs used for the bioinformatic analyses included 43 sequences from Phytophthora species and Hyaloperonospora arabidopsidis and has been previously published [21]. The Signal Peptide and RxLR portion of these sequences (positions 1 to 90) were aligned using the muscle algorithm [49]. The HMM building, calibration, and searches were performed using the HMMER software package with hmmbuild, hmmcalibrate, and hmmsearch algorithms, respectively (http://hmmer.org/). The HMM search included only one iteration.
The predicted effectors were PCR amplified without signal peptide from Hpa strain Emoy2 DNA (obtained from Jonathan Jones) using primers listed in Table S2 and cloned into the pENTR/D-Topo vector (Invitrogen). The insert sequences were verified and the effectors recombined with LR clonase (Invitrogen) into the binary Pseudomonas expression vector pPsSP [35]. The pEDV3 clones of effectors were generated using restriction digests with SalI/BamHI or compatible restriction enzymes. Site-directed mutants ATR39-1in and ATR39-2del were generated in pENTR using the QuikChange Lightning kit (Agilent) and primers listed in Table S2, and recombined with binary vectors as above. For transient expression in Nicotiana benthamiana, ATR39 alleles were recombined into pEarleygate201 containing a 35S promoter and N-terminal HA-tag [50]. The fasta files containing the amino acid and nucleotide sequences of the predicted effectors are available online as Text S1 and Text S2.
A. thaliana plants were grown on soil in controlled growth chambers at short days (8 hrs light/16 hrs dark cycle) and 24°C. Transgenic Arabidopsis were surface sterilized and selected on MS medium with the appropriate antibiotics. For Hpa growth assays the plants were transferred to a growth chamber with 18°C and 100% humidity.
Hpa isolates were obtained from E. Holub (Maks9), X. Dong (Emwa1), J. McDowell (Emoy2, Emco5), X. Li (Cala2) and J. Jones (Noco2) and maintained on susceptible Arabidopsis plants as previously described [27]. For disease assays, conidiospores were harvested by vortexing in water, adjusted to 5*104 spores/mL and sprayed on 2-week-old Arabidopsis seedlings. The infected plants were kept in growth chambers at 18°C and 100% humidity for 7 days before being stained with lactophenol-trypan blue [40] to assess Hpa growth or resistance.
Pseudomonas fluorescens strains were grown on Pseudomonas agar with glycerol (PAG) supplemented with the appropriate antibiotics. For HR assays, Pf0 strains were grown on PAG plates for 2 days and resuspended in 10 mM MgCl2 to OD600 nm = 1 (corresponding to 109 cfu/mL). Bacteria were inoculated into halves of pierced Arabidopsis leaves using a blunt syringe. Visible HR symptoms were scored 24 hours post-inoculation. Pseudomonas growth assays were performed as described previously [51].
Markers and probes used in map-based cloning of RPP39 are summarized in Table S3. A fosmid library of Wei-0 was generated following the instructions in the copy control fosmid kit (Epicentre) and screened with Digoxigenin-labeled probes using the DIG DNA labeling and detection kit (Roche). To sequence the RPP39 region, 350 bp-sized fragments of overlapping fosmids were sequenced in a 60 bp paired-end sequencing run on an Illumina G2. Reads were cleaned for vector sequences and bacterial contamination using MAQ (http://maq.sourceforge.net/) and assembled using CLC genomics workbench (http://www.clcbio.com/). Gaps and misassembled regions were filled in using Sanger sequencing data.
Genomic fragments of about 6 kb length, containing RPP39 (R190_Wei-0) or R180_Wei-0 were amplified using primers specified in Table S2 and introduced into pENTR/D-Topo (Invitrogen). Sequences were confirmed and the fragments recombined into pEarleygate301 [50] for expression in Agrobacterium. The RPP39 cDNA clone was amplified from Wei-0 cDNA and recombined into pEarleygate100 containing a 35 S promoter [50]. The binary vectors were mobilized into Agrobacterium tumefaciens strain GV3101 with tri-parental mating and used to transform Arabidopsis plants following the floral dip protocol [52]. Presence of the transgene in Arabidopsis was confirmed by PCR. Transient expression experiments in N. benthamiana were performed as previously described [35]. Protein expression was determined by Western blotting as described [31].
RNA was extracted from Arabidopsis seedlings infected with Hpa using the RNeasy plant mini kit (Qiagen) and reverse transcribed using Superscript III (Invitrogen) and oligo-dT primers. RT-PCR was performed with gene specific primers (Table S2) and 25 amplification cycles.
|
10.1371/journal.pcbi.1004016 | A Systematic Computational Analysis of Biosynthetic Gene Cluster Evolution: Lessons for Engineering Biosynthesis | Bacterial secondary metabolites are widely used as antibiotics, anticancer drugs, insecticides and food additives. Attempts to engineer their biosynthetic gene clusters (BGCs) to produce unnatural metabolites with improved properties are often frustrated by the unpredictability and complexity of the enzymes that synthesize these molecules, suggesting that genetic changes within BGCs are limited by specific constraints. Here, by performing a systematic computational analysis of BGC evolution, we derive evidence for three findings that shed light on the ways in which, despite these constraints, nature successfully invents new molecules: 1) BGCs for complex molecules often evolve through the successive merger of smaller sub-clusters, which function as independent evolutionary entities. 2) An important subset of polyketide synthases and nonribosomal peptide synthetases evolve by concerted evolution, which generates sets of sequence-homogenized domains that may hold promise for engineering efforts since they exhibit a high degree of functional interoperability, 3) Individual BGC families evolve in distinct ways, suggesting that design strategies should take into account family-specific functional constraints. These findings suggest novel strategies for using synthetic biology to rationally engineer biosynthetic pathways.
| Bacterial secondary metabolites mediate a broad range of microbe-microbe and microbe-host interactions, and are widely used in human medicine, agriculture and manufacturing. Despite recent advances in synthetic biology, efforts to engineer their biosynthetic genes for the production of unnatural variants are frustrated by a high failure rate. In an effort to better understand what types of genetic changes are most likely to lead to successful improvements, we systematically analyzed the ways in which biosynthetic genes naturally evolve to generate new compounds. We show that large gene clusters appear to evolve through the merger of sub-clusters, which function independently, and are promising units for cluster engineering. Moreover, a subset of gene clusters evolve by concerted evolution, which generates sets of interoperable domains that may enable predictable domain swapping. Finally, many biosynthetic gene clusters evolve in family-specific modes that differ greatly from each other. Overall, this quantitative perspective on the ways in which gene clusters naturally evolve suggests novel strategies for using synthetic biology to engineer the production of unnatural metabolites.
| Bacterial secondary metabolites are widely used as pharmaceutical, agricultural, and dietary agents. They consist of many classes of compounds including polyketides (PKs), nonribosomal peptides (NRPs), ribosomally synthesized and post-translationally modified peptides (RiPPs), terpenoids, saccharides, and a plethora of hybrids. The genetic basis for this rich molecular diversity can be found in biosynthetic gene clusters (BGCs), physically clustered groups of genes that encode the enzymatic pathways necessary to construct specific chemicals [1], [2].
The diversity of extant natural products and BGCs raises important questions about their evolutionary origin. These include the basic question of how Nature invents new molecules, and a series of applied questions relevant to biotechnology: for example, the evolutionary modularity of NRP and PK BGCs has long been seen as a feature that might allow large libraries of new compounds to be generated by mixing and matching their constituent domains and modules [3]. However, although there have been notable successes [4]–[6], the majority of combinatorially generated pathways appear to be nonfunctional [4]. More recently, advanced synthetic biology approaches to pathway engineering have been frustrated by the complexity and unpredictability of metabolic enzymes, particularly NRPSs and PKSs [7], [8]: unlike LEGO bricks, their constituent domains and modules do not ‘fit’ together universally, but only function effectively in specific pathway contexts.
Regardless of these apparent constraints to genetic change, Nature appears to have been quite successful at engineering biosynthetic pathways through the process of gene cluster evolution: even a conservative estimate suggests that the number of broad biosynthetic gene cluster families that have evolved exceeds 6,000 [8], most of which contain multiple BGCs that synthesize derivatives of a common scaffold. Hence, a detailed study of evolutionary patterns within various BGC families has the potential to offer a new inroad into effective BGC engineering, through mimicry of Nature's evolutionary design strategies.
So far, insights into the key principles underlying the evolution of BGC architectures and repertoires have been derived from limited case studies [9]–[13], which lack sufficient detail about the generality of the underlying mechanisms. Here, we systematically quantify the strategies that make evolution so successful at engineering BGC diversity. Through a detailed computational analysis of a recently generated dataset of 732 known and 10,724 predicted prokaryotic BGCs [8], we find that the rates of evolutionary events, such as insertions, deletions and duplications within BGCs, are much higher than those seen in comparable gene clusters involved in primary metabolism. Furthermore, distinct sub-clusters consisting of co-evolving genes appear to constitute relatively independent building blocks that play key roles in the evolution of larger BGCs encoding the biosynthesis of complex metabolites. Finally, BGC families encoding the production of polyketides and nonribosomal peptides evolve in family-specific modes, in many of which we observe an unexpectedly large role for concerted evolution [14], [15] driven by internal recombinations. Based on these observations, we offer several recommendations for establishing new modes of evolution-guided BGC engineering.
The large diversity of BGCs observed throughout the prokaryotic tree of life [8] suggests that BGCs evolve rapidly. Indeed, when we systematically quantified different evolutionary events by mutually comparing all gene clusters in our data set (Table S1), we found not only that they may have been transferred horizontally at high frequency (Fig. 1a and Figure S1), but also display exceptionally high rates of insertions, deletions, duplications and rearrangements (Fig. 1b). While the percentage of gene cluster pairs related by an indel is independent of gene cluster size, the distribution of indel sizes shows a long tail that includes 195 indels of 10 kb or more (Fig. 1c). As expected, these large indels are more commonly found in larger gene clusters, where they indicate either the merger of one gene cluster fragment with another or the loss of a gene cluster fragment from a larger cluster (see examples in Figure S2). Phylogenetic profiling [16] showed that many such BGC fragments – here termed sub-clusters – appear to evolve in a correlated fashion: 884 different motifs of adjacent Pfam domains (out of 7,641 found) were shown to co-evolve significantly more often than not (P<0.001), based on the χ2 test. These motifs comprise 591 different Pfam domains and have an average length of 5.3 domains (Table S2). As expected, they include many well-known and widely conserved motifs that appear to be linked to specific sub-functionalities of gene clusters, such as precursor biosynthesis, transport or synthesis of a specific chemical moiety, and motifs belonging to modular BGC architectures of NRPSs and PKSs (e.g., C-A-T and KS-AT-T [17]).
Earlier evidence has suggested complex mosaic patterns of sub-cluster sharing for some BGCs, such as those involved in the production of glycopeptides [18]. To further explore the role of sub-cluster sharing in the evolution of BGCs, we manually compiled a set of 35 BGCs that are rich in sub-clusters that have a known connection with a specific chemical moiety. We then used this data set to construct a network in which the nodes represent BGCs and the edges denote a sub-cluster that a pair of BGCs has in common (Fig. 2). Three observations were particularly notable (Fig. 2). First, >60% of the coding capacity of some BGCs (e.g., those encoding vancomycin and rubradirin [19]) is composed of individually conserved sub-clusters (note that this is not entirely reflected in the depiction of the rubradirin gene cluster in Fig. 2b, where only those sub-clusters are highlighted that are shared with other depicted BGCs). This supports a “bricks and mortar” model of gene cluster evolution in which gene clusters are composed of large, modular “bricks” (sub-clusters) that encode key building blocks and individual genes (the “mortar”) that encode functions such as tailoring, regulation and transport. During evolution, both bricks and mortar (scaffold and tailoring) may remain the same, only the tailoring may change or the scaffold itself may change. Second, the same sub-cluster commonly appears in otherwise unrelated BGCs, and multiple unrelated sub-clusters can be found in a single parent gene cluster, indicating that sub-clusters are independent evolutionary entities. Third, sub-clusters are not static; they are loosely organized around a core set of genes, but gene gain/loss leads to chemical changes in the corresponding part structure: for example, gene clusters encoding molecules such as everninomicin [20], simocyclinone [21] and polyketomycin [22] have different variants of deoxysugar sub-clusters, which lead to subtle variations in the final chemical structures.
Although the complex patterns of sub-cluster sharing, in which various sub-clusters are shared between otherwise completely different gene clusters (Fig. 2), indicate that BGCs may evolve by the successive merger of sub-clusters, this does not mean that every case where sub-clusters are shared points to an independent sub-cluster transfer event. For example, the KS domains of the diverse range of ansamycin type I PKS BGCs that harbor AHBA sub-clusters are almost completely monophyletic (Figure S3), indicating that the macrolactam- and AHBA-producing sub-clusters have been co-evolving for a long time (instead of multiple independent AHBA sub-cluster acquisitions having occurred in different macrolactam-producing polyketide BGCs). Hence, the multi-hybrid rubradirin gene cluster might have arisen from a rifamycin-like ancestor (most rubradirin KS domains are monophyletic with rifamycin KS domains, see Figure S3) that already harbored the combination of a modular type I PKS sub-cluster and an AHBA biosynthesis sub-cluster, and which then acquired new sub-clusters for the biosynthesis of the aminocoumarin, 3,4-dihydroxydipicolinate and nitrosugar moieties (which are not found in any other closely related ansamycins). Contrary to the shared evolutionary histories of AHBA and ansamycin type I PKS sub-clusters, a clear example of sub-cluster transfer between BGCs of different types can be seen for 6-methylsalicylic acid (MSAS)/orsellinic acid (OSAS) sub-clusters, as inferred from a maximum-likelihood phylogenetic tree of MSAS/OSAS-producing iterative PKSs (Figure S4). The topology of this tree strongly indicates that MSAS/OSAS sub-clusters have largely evolved independent of the scaffold types of their parent gene clusters (Figure S4), and that they have been transferred between multiple types of BGCs during their evolutionary past. In conclusion, in the context of the bricks-and-mortar analogy, some bricks move around between different structures more often than others. Finally, we should note that there are also BGC families which evolve over long periods of time without major changes to the gene cluster architecture or the scaffold of the core molecule made: for example, the large family of over >1,000 aryl polyene BGCs that we described recently [8] has not undergone any major sub-cluster transfers, aside from the inclusion of the dialkylresorcinol sub-cluster in the BGCs from some CFB group bacteria. The products of many of these BGCs are likely to be entirely identical, while remaining differences between the molecules mostly concern differential tailoring of the same scaffold.
Many chemical scaffold types of secondary metabolite classes are quite distinct, which raises the question of how BGC families encoding the synthesis of distinct scaffolds are related. To assess this question, we calculated the proportion and similarity of Pfam domains shared between all pairs of BGCs within our data set of 732 known gene clusters using multiple sequence alignments for each Pfam domain (Fig. 3) and looked specifically for close homologues of BGCs just outside their immediate family. Even though of course sequence similarity alone does not provide conclusive evidence on evolutionary histories, the analysis did suggest that unexpected evolutionary connections might exist between natural products of different scaffold types.
For example, the Streptomyces gene cluster encoding the lipopeptide antibiotic daptomycin [23] is surprisingly similar to Mycobacterium glycopeptidolipid (GPL) gene clusters [24] (Figure S5). When we performed a more in-depth analysis through a phylogenetic analysis of condensation domains, we indeed found that GPL domains consistently cluster together with domains from the NRPSs that synthesize daptomycin (Figure S6). Although both daptomycin and the GPLs are lipopeptides, the Mycobacterium GPLs are shorter (tetrapeptide vs. tridecapeptide), cell-wall-associated rather than diffusible, linear rather than cyclic, and originate from an actinomycete genus that is not closely related to Streptomyces.
Likewise, one of the strongest matches for the gene cluster encoding the immunosuppressant rapamycin [25], apart from the closely related FK520 [26] and meridamycin [27], [28] BGCs, was the gene cluster for pladienolide [29], a polyketide of unrelated structure with a distinct biological activity (inhibition of the splicing factor SF3b instead of TOR). Strikingly, based on phylogenetic trees of their constituent ketosynthase (KS) and acyltransferase (AT) domains, the meridamycin gene cluster is more closely related to the pladienolide BGC than to those encoding rapamycin and FK520, the molecules to which it is often compared (Fig. 3). These examples suggest that closely related sets of protein domains can be reconfigured by evolution to yield a new scaffold that is chemically and biologically distinct.
The phylogenetic trees of KS and AT domains from our data set of known BGCs revealed another unexpected finding: in spite of the structural similarity of rapamycin and FK520, 63% of the constituent domains of their polyketide synthases (PKSs) cluster into entirely separate clades (Fig. 3b, see also Figure S8 which shows that relevant bootstrap values are almost all above 90). Even more remarkably, 14 out of 16 domains responsible for the biosynthesis of the sub-structure shared between these two molecules (shown in red in Fig. 3c) do not cluster together with the corresponding domain from the assembly line for the other molecule. This pattern of homology is consistent with a phenomenon called ‘concerted evolution’, the homogenization of DNA sequences within a given repetitive family caused by high rates of internal recombination [14], [15]. Given the similar sizes and architectures of the gene clusters and the structural similarity of their products, this is a much more parsimonious explanation for the patterns observed than convergent evolution of multiple similar gene clusters through successive duplication of an ancestral single-module PKS. Notably, previous phylogenetic analyses of PKS domains have also observed BGC-specific clades of PKS domains [10], [30], but not to the extent observed here for such closely related gene clusters: the fact that such a strong pattern is even observed for the AT domains of two different gene clusters that encode the same molecule [27], [28], meridamycin, shows that the underlying process may operate on very short time scales, and that recombination can remove almost all traces of independent evolution of these PKS modules. In the case of the rapamycin family, recombinations are likely to occur neutrally and have no effect on the structure of the small molecule product (rapamycin, meridamycin and FK520), whereas in other cases, single crossovers within or between gene clusters may dramatically change the modular architecture of a synthase [30]. Near-neutral changes brought about by gene conversion may occur at higher rates for some domains or domain types than for others: in the meridamycin gene clusters, no signs of gene conversion could (yet) be observed for KS domains, even though gene conversion manifested itself clearly when comparing the meridamycin clusters with those encoding rapamycin, FK520 and pladienolide. On the contrary, AT domain gene conversion was widespread even between the two meridamycin gene clusters. We speculate that for these BGCs, gene conversion events get fixated in the population at lower rates for KS domains because not all KS sequences work equally well for different polyketide chain lengths that occur at different points of the assembly line, so that the changes brought about by a conversion event are less neutral than for AT domains. Mapping of rapamycin family PKS sequence mutations onto the 3D structure of an AT- and KS-containing protein further supports this hypothesis (Figure S7a), showing widespread sequence variability at almost every position in the AT domains, except for the residues near the substrate binding site (Figure S7b). Mutations in KS domains, on the other hand, are mostly restricted to the regions in vicinity (around the core) of the substrate-binding site and the dimerization interface (Figure S7c), suggesting their importance in influencing substrate selectivity.
Concerted evolution is not peculiar to the rapamycin family (Figure S8). For the gene clusters encoding the biosynthesis of the mutually closely related macrolides erythromycin [31], oleandomycin [32] and pikromycin [33], BGC-specific branching appeared to occur for both KS and AT domains, similar to the pattern for rapamycin, FK520, meridamycin and pladienolide. However, for the ansamycin antibiotics macbecin [34], geldanamycin [35] and herbimycin [36], and the antifungals pimaricin [37], nystatin [38] and amphotericin [39], BGC-specific branching occurs only for AT domains, and not for KS domains. Finally, corroborating earlier observations [40], domains from the trans-AT PKS gene clusters encoding pederin [41] and psymberin [42] do not show any BGC-specific branching at all. We observed that certain NRPS gene clusters also show signs of concerted evolution: a clear BGC-specific branching pattern pointing to concerted evolution can be seen for the A domains and most of the C domains of the gene clusters encoding the biosynthesis of the closely related calcium-dependent lipopeptides daptomycin [23], A54145 [43] and CDA [44]. However, the glycopeptide gene clusters encoding the biosynthesis of balhimycin [45], teicoplanin [46] and A40926 [47] showed no such pattern at all: almost all domains cluster in groups corresponding to domains in the same positions in the assembly line. Collectively, these observations suggest that concerted evolution is a key mechanism driving the evolution of NRPS and PKS gene sequences, but the extent to which it happens depends on family-specific functional constraints as well as on the presence of other evolutionary forces acting upon a gene cluster. Our qualitative model of PKS/NRPS evolution (Fig. 4), which summarizes the interplay of concerted evolution with other evolutionary mechanisms, is relevant to PKS/NRPS engineering efforts: the highly homologous sets of domains generated by concerted evolution are more likely to be mutually interoperable than domain sets chosen at random, and might therefore be attractive building blocks for synthetic biological engineering of biosynthetic pathways.
To understand more generally how PKS and NRPS BGCs evolve, we set out to measure the contributions of concerted evolution, duplication, and divergence to the evolution of all multimodular PKS and NRPS BGCs in both our known and predicted BGC data sets. We first collected and quantified 25 different features describing the nature of gene cluster sequences and the relationships among their constituent domains (see methods for details). A principal component analysis (PCA) and hierarchical clustering using these features can distinguish many of the well-known gene cluster families from our data set of known BGCs (Figure S9, Fig. 5a). Two features in particular, the ‘internal similarity index’ and the ‘vertical evolution index’, explain much of the variation in terms of the modes of evolution of different classes of gene clusters (Fig. 5b). At the level of individual domains, we find that there are four primary mechanisms by which NRPS and PKS BGCs evolve (Fig. 5c–f, Figure S10). Firstly, gene clusters encoding glycopeptides, calcium-dependent lipopeptides and macrolides/polyethers appear to be most repetitive, pointing to a history of module duplications and/or a prominent influence of concerted evolution. The syringopeptin NRPS [48] and mycolactone PKS [49] are extreme examples of this: both are likely to have evolved recently by subsequent module duplications and concerted evolution. Secondly, we sometimes observed gradients of the internal homology p-values from the N- to C-termini of large synthases, suggesting that some gene clusters evolve to encode the synthesis of larger molecules by iterative duplication of their most N-terminal module, would have the effect of extending an intermediate NRP or PK by the addition of a new starter unit. Thirdly, a group of BGCs including the ones that encode the polyketides psymberin [42] and erythrochelin [50] show a ‘vertical’ type of evolution, in which the domains appear to evolve independently, with perhaps occasional domain swapping with related gene clusters, as has been suggested previously [40]. Finally, there are many gene clusters showing a ‘mixed’ mode of evolution, in which one or more of the above mechanisms are combined. For example, NRP siderophore gene clusters show some signs of internal recombinations, but at the same time many domains show no high mutual similarity. Like the trans-AT PKS gene clusters, they seem to have a higher tendency to recruit domains from dissimilar gene clusters. This recruitment over larger evolutionary distances appears to be a general feature of NRPS gene clusters as opposed to PKS gene clusters, and might be related to the wider range of possible substrates for NRPSs, which often require BGC-specific sub-pathways for the synthesis of a dedicated monomer [51].
The observation of so many different evolutionary mechanisms of gene cluster evolution begs the question which circumstances lead to the birth and death of BGCs over evolutionary time. Are all BGCs that are detected bioinformatically also still intact and functional, or might many of them have degenerated and entered a nonfunctional state? The absence or presence of nonfunctional genetic units (e.g., pseudogenes or pseudo-gene-clusters) is largely governed by the evolutionary population dynamics of the species. Many bacteria live in large effective population sizes and have relatively short generation times, leading to very strong purifying selection and, consequently, rigorous genome streamlining [52]. Hence, BGCs that become nonfunctional will be quickly lost in such organisms if they do not provide any evolutionary advantage. Notably, some bacteria in fact occur in smaller population sizes and/or regularly go through population bottlenecks, leading to altogether different evolutionary dynamics [53]: in such cases, a range of pseudogenized gene clusters can sometimes still be observed that have not been purged from the genome yet [54]. On the whole, however, these appear to be rather the exception than the rule [55].
Concerning the birth of new gene cluster architectures, large effective population sizes and short generation times also suggest that BGC modifications should immediately confer an evolutionary advantage in order to be maintained; on the other hand, frequent changes in population size may affect the probability of mutations to be fixated in the population [56]. Alternatively, neutral mutations could hitchhike with strongly adaptive mutations within or close to the same gene cluster. Concerning the physical growth of gene clusters, it should be noted that new enzymes may already be recruited to a biosynthetic pathway before their genes are physically recruited to the gene cluster, and such an addition to a pathway could evolve through, e.g., positive selection acting on promiscuous enzyme activities or substrate specificities [57]. The precise reason for and evolutionary mechanism of clustering of biosynthetic genes in bacteria itself is still largely an unanswered question [58].
Our analysis of BGC evolution will enable new approaches to BGC engineering informed by the mechanisms by which BGCs evolve naturally. Our results suggest that efforts to engineer the biosynthesis of unnatural natural products could be more successful by observing the modes by which specific BGC classes evolve in nature.
For example, conglomerate molecules consisting of multiple different chemical moieties could be designed by engineering BGCs consisting of novel combinations of sub-clusters. Such an effort could be guided by information taken from evolutionary comparisons, which would offer suggestions about which sub-clusters are most likely to function together, based on how often evolution has successfully forged combinations between them.
Furthermore, our evolutionary analysis of NRPS and PKS gene clusters suggests that concerted evolution has created sets of domains within gene clusters that are highly homologous. These domain sets are more likely to be mutually interoperable than domain sets chosen at random, and might therefore be of great utility in future engineering efforts.
Also, evolutionary strategies towards generating larger and more complex compounds could be mimicked by N-terminally extending certain types of NRPS/PKS gene clusters by duplicating and then carefully modifying the first assembly-line module.
Overall, in combination with new synthetic biology techniques that may soon enable the rapid assembly of thousands of clusters from a common set of parts [59]–[61], our results suggest a new approach for re-engaging gene cluster engineering in a manner informed by the mechanisms by which gene clusters have naturally evolved.
To remove highly similar genomes from these analyses, we used the AMPHORA [62] (August 10th, 2010) dataset, which contains gene sequences from 562 organisms for 30 universally conserved genes. Genes from these organisms were compared using sequence identities based on MUSCLE [63] multiple sequence alignments. This resulted in 30 distances between each pair of organisms. The distributions of distances of all pairs were tested for normality using a Shapiro-Wilk test. An organism distance map was then built with distances defined as the mean distances of AMPHORA genes. The resulting distance map was clustered using default settings in MCL [64], and only one member of each cluster was kept for further analyses. This left us with total of 408 organisms.
To search for histidine and tryptophan biosynthetic operons, we modified ClusterFinder [8]. Pfam [65] IDs associated with the histidine biosynthesis pathway (PF00475, PF00815, PF01174, PF01502, PF01634, PF04864, PF08029, and PF08645) or with the tryptophan biosynthesis pathway (PF00218, PF00290, PF00465, PF00697, PF01220, PF01264, PF01487, PF04715, and PF08501) were acquired from JGI IMG [66]. Trp or His operons were defined as gene clusters containing at least one of these domains with a probability >0.5 and containing at least two of the domains in total. Among 408 organisms searched, 350 His and 288 Trp biosynthesis operons were identified in 271 and 248 different organisms, respectively. The average number of domains per predicted gene cluster were 2.9 and 3.1, respectively.
Best matching sequence homologs of a query protein domain from a biosynthetic or primary metabolic gene cluster were obtained using MUSCLE [63] multiple sequence alignments. The distance between the organism containing the query protein domain and the organism with the best matching sequence homolog was determined based on 16S rRNA sequence similarity. Best matching sequence homologs of all protein domains that are in Pfam are included in the organism similarity histograms (Fig. 1a).
For each BGC, a two-dimensional array of the size corresponding to the numbers of consecutive protein domains that are in Pfam database (rows) and 408 selected organisms (columns) (see “Comparison of HGT with primary metabolism”) was created. The cells in the array consisted of sequence identities between a given domain from a BGC and the most homologous domain (which is also predicted as part of a BGC) from a given organism. Next, we calculated a Pearson product-moment correlation coefficient (correlation coefficient) for each possible pair of protein domains (rows), resulting into a new matrix, a correlation matrix, of the size corresponding to the number of protein domains (rows from the initial array) in both dimensions. To take rearrangements into account, we reordered rows and columns of the correlation matrix based on hierarchical clustering of the correlation matrix in both dimensions. We then parsed linear motifs that are likely to evolve in a correlated fashion by selecting consecutive pairs of domains in this reordered correlation matrix (consecutive fields on the first offset diagonal) with correlation coefficient >0.5. The analysis was repeated by setting the correlation coefficient cutoff to >0.65 and >0.8. Each motif was divided into all possible sub-motifs of sizes between 2 domains and the total number of domains in a motif. To determine the significance of a (sub)motif occurrence, we next compared the number of (sub)motif occurrences to the number of all possible (sub)motif occurrences in all BGCs that did not pass the correlation coefficient cutoff. Pearson's χ2 test with Bonferroni correction was applied to test for statistical significance, with the null hypothesis stating that the two values are equal.
We performed an all-versus-all alignment of nucleotide sequences of known and predicted BGCs using the blastn algorithm. Gene cluster sequences were divided into blocks of 1 kb, and then mapped to the most homologous blocks from other gene clusters, as well as from the same gene cluster (to test for genomic duplications). 56% of the blocks (118,320 out of 212,176) did not map to any homologous regions in the same or other BGCs with >70% identity. Evolutionary events (insertions/deletions, duplications and rearrangements) were detected by a custom-made Python script (Data S1) comparing each alignment of two-gene clusters having at least three matching blocks with >70% identity. Rearrangements were defined as an identified difference in the order of 1-kb blocks in an otherwise conserved (piece of) gene cluster, such as when A1-A2-A3-A4-A5 matches to B1-B4-B3-B2-B5 in an alignment of two BGCs A and B. Indels were defined as 1-kb blocks present in one gene cluster but not in the other gene cluster, such as when A1-A2-A3-A4-A5-A6 matches to B1-B2-B5-B6 in an alignment. To make these inferences more reliable, a constraint was used that the flanking regions (of size > = 2 kb) of each indel breakpoint must be homologous between query and hit gene cluster, and the block order must be conserved between them. Finally, duplications were defined as 1-kb blocks that had the best hit towards another block in its own gene cluster, and having a higher copy number in one gene cluster than in the other, such as when A1-A2-A3-A2-A3-A4-A5 aligns to B1-B2-B3-B4-B5, while the mutual sequence identity between the A2 and A3 pairs is higher than between any of the A2/A3 blocks and B2 or B3.
For a given BGC pair, we first calculated sequence identities between all Pfam domain pairs of each Pfam ID, using MUSCLE [63] multiple sequence alignments. A BGC sequence similarity index was defined as the Jaccard index with the size of the intersection represented by the number of Pfam pairs whose sequence identities were higher than the best 10% alignments of all Pfam domains of the same Pfam ID. Taking into account the underlying distributions of sequence identities between all domain sequences prevented misinterpretation of simpler sequence similarity metrics (e.g., an absolute sequence identity threshold) when different evolutionary rates apply to different protein families. We define structural similarity of a given BGC product pair as the Tanimoto coefficient between the two SMILES strings, using linear-path fingerprints (FP2) from Open Babel [67].
Sub-clusters with known functions from experimentally characterized gene clusters were manually collected from the literature. Sub-cluster sharing between gene clusters from the training set was calculated using blastp [68]. The minimum requirement used to identify a shared sub-cluster between two BGCs was sharing either 75% of the genes with >45% average sequence identity, 50% of the genes with >50% average sequence identity, or 25% of the genes with 70% identity. To account for different modes of sequence evolution of different sub-cluster types, these values were adjusted with sub-cluster type-specific cutoffs to obtain a good match between genetic similarity and chemical similarity (Table S3). The final sub-cluster sharing network was drawn with Cytoscape [69].
To study patterns of evolution in multimodular NRPS and PKS gene clusters, a range of features was calculated describing key characteristics of these gene clusters. The first set of features was based on the topologies of intra-BGC domain similarity networks (with protein domains and sequence similarity representing nodes and edges, respectively) and consisted of the average clustering coefficient, average sequence similarity, graph transitivity, number of 2–4 node cliques, number of connected components in a graph with sequence similarity >50%, and average neighbor degree. We also included as features the number of different Pfam domain types in a BGC, the total number of domains in a BGC, the average number of domains per gene, and the averages and standard errors of best-matching pair sequence identities and internal BGC similarity indices. Two evolutionary indices were also added: the internal similarity index and the vertical evolution index. To obtain the internal similarity index of a gene cluster, we calculated for each of its NRPS/PKS domains the p-value of its closest blastp match inside the gene cluster, given the distribution of the percent identities of all within-gene-cluster blastp hits of all domains of that domain type in the complete set of gene clusters. The internal similarity index was then calculated from these numbers as the mean of all inverse p-values. The same inverse p-values were used for plotting the internal domain similarity across gene clusters. The vertical evolution index of a gene cluster was calculated as the average difference between the p-value of the top 10 percent identities of a domain's blastp hits to all domains from other gene clusters with the p-values of the Lin distances of the gene clusters to the host gene clusters of each of the top 10 hit domains. Consequently, gene clusters with domains with highly similar closest hits to domains in dissimilar gene clusters get a low value, while gene clusters with domains with dissimilar closest hits to domains in similar gene clusters get a high value.
PCA analysis was performed with the aforementioned features as an input. Compound types were assigned using the classifications taken from the primary literature.
|
10.1371/journal.pgen.1001366 | Genome Analysis Reveals Interplay between 5′UTR Introns and Nuclear mRNA Export for Secretory and Mitochondrial Genes | In higher eukaryotes, messenger RNAs (mRNAs) are exported from the nucleus to the cytoplasm via factors deposited near the 5′ end of the transcript during splicing. The signal sequence coding region (SSCR) can support an alternative mRNA export (ALREX) pathway that does not require splicing. However, most SSCR–containing genes also have introns, so the interplay between these export mechanisms remains unclear. Here we support a model in which the furthest upstream element in a given transcript, be it an intron or an ALREX–promoting SSCR, dictates the mRNA export pathway used. We also experimentally demonstrate that nuclear-encoded mitochondrial genes can use the ALREX pathway. Thus, ALREX can also be supported by nucleotide signals within mitochondrial-targeting sequence coding regions (MSCRs). Finally, we identified and experimentally verified novel motifs associated with the ALREX pathway that are shared by both SSCRs and MSCRs. Our results show strong correlation between 5′ untranslated region (5′UTR) intron presence/absence and sequence features at the beginning of the coding region. They also suggest that genes encoding secretory and mitochondrial proteins share a common regulatory mechanism at the level of mRNA export.
| The function and evolution of introns have been topics of great interest since introns were discovered in the 1970s. Introns that interrupt protein-coding regions have the most obvious potential to affect coding sequences and their evolution, and they have therefore been studied most intensively. However, about one third of human genes contain introns within 5′ untranslated regions (UTR). Here we observe that certain classes of genes, including those targeted to the endoplasmic reticulum and nuclear-encoded mitochondrial genes, are surprisingly depleted of 5′UTR introns. We offer and support a model that explains this observation and points to a surprising connection between 5′UTR introns and how mRNAs are exported from the nucleus.
| In humans, ∼35% of all genes have introns in their 5′ untranslated regions (UTRs) [1]–[3]. These introns differ from those in coding regions, for example, in typical length and nucleotide composition [1]–[3]. Previously, 5′UTR introns (5UIs) were suggested to be evolving under a neutral model of random insertion and deletion events with the sole constraint of avoiding upstream open reading frames [3]. Recently, we showed that presence and length of 5UIs correlates with the level of expression across cells and tissue types [1]. More importantly, we observed an uneven distribution of 5UIs amongst genes across specific functional categories [1]. Genes with regulatory roles, including non-receptor tyrosine kinases, regulators of cytoskeleton, transcription and metabolism, were enriched in having 5UIs [1]. Our results suggested that many 5UIs are evolving under complex selective forces as opposed to a simple model of neutral evolution [1]. However, it is unclear whether there is any widely used mode of regulation that is unique to 5UIs.
In eukaryotes, splicing is coupled to key mRNA metabolic processes. During the act of splicing, several different protein complexes are deposited onto mRNA. For example, the Transcription Export (TREX) complex promotes the nuclear export of fully processed transcripts [4]. In higher eukaryotes, the TREX complex is deposited primarily onto the 5′ end of nascent transcripts by the cooperative action of the cap-binding complex and the spliceosome [5]. Given that 5UIs are necessarily proximal to 5′ ends of transcripts, an intriguing possibility is that splicing of 5UIs could have a disproportionate impact on mRNA export by promoting TREX recruitment. Although the majority of transcripts follow the splicing-dependent export pathway, alternative pathways exist. Recently, Palazzo et al. demonstrated that mRNAs that encode secreted proteins can use an alternative route for mRNA export that is mediated by a nucleotide element within the signal sequence coding region (SSCR) [6]. In contrast to the splicing-dependent pathway, this alternative RNA export (ALREX) pathway does not require splicing or a 5′ cap [6]. Vertebrate SSCRs were found to be adenine-poor and silent mutations introducing adenines into the SSCR impair its ability to promote mRNA export [6]. However, beyond adenine-depletion this element has been poorly characterized. Furthermore, it has remained unclear which SSCR-containing transcripts use ALREX and to what extent, since the vast majority of SSCR-containing transcripts are also spliced and thus could potentially use the canonical export pathway. The fact that both ALREX signals and splicing signals are found near the 5′ end of genes, suggests the interesting possibility that competition between signals at the 5′end of transcripts determines how a given mRNA is exported.
Here, we extend our computational analysis of 5UIs to identify functional groups of genes that preferentially lack these introns. We find that 5UIs are depleted in genes containing SSCRs or mitochondrial-targeting sequence coding regions (MSCRs). We demonstrate that SSCRs and MSCRs derived from 5UI-lacking (5UI−) genes contain sequence features associated with ALREX and promote export in vivo. In stark contrast, SSCRs and MSCRs derived from 5UI+ genes do not exhibit ALREX-associated features. Furthermore, we show that 5UI+ genes do not support splicing-independent mRNA export. We then characterize ALREX elements more fully by identifying and validating new ALREX-associated motifs. Taken together, our results support a model wherein the 5′-most element in a newly synthesized transcript, be it an intron or an ALREX element, dictates which pathway is employed for export. Furthermore, our results provide the first known regulatory role that is unique to 5′ UTR introns and suggest that it is widely used.
Using a high quality set of 5UI definitions for human, we observed a depletion of 5UIs amongst genes with certain Gene Ontology [7] (GO) annotations (Table S1). Examples of 5UI-depleted GO terms include “MHC class II protein complex” (ratio of 5UI-containing genes to total genes annotated with particular GO term is 0/25), “aspartic endopeptidase activity” (0/23), “voltage-gated calcium channel activity” (2/35), “growth factor activity” (33/180), “electron carrier activity” (27/145), and “extracellular space” (108/497). In each case, these ratios are significantly lower than the ratio of ∼35% expected by chance (p<0.05 after adjusting for multiple hypothesis testing). More generally, we observed a depletion of 5UIs among nuclear genes encoding three protein classes.
The first class was composed of protein families encoded by mostly intronless genes. This group includes histone genes [8], olfactory receptors, G-protein coupled receptors [9], and keratins [10], [11]. Depletion of 5UIs in these gene classes does not suggest any 5UI-specific phenomena, as these genes are more generally intron-depleted.
The second class was composed of secreted or membrane-bound proteins that are trafficked through the endoplasmic reticulum (ER). We compiled a list of all genes with signal sequence coding regions (SSCRs), encoding N-terminal cleavable signal sequence peptides that target newly synthesized proteins to the ER [12] (see Materials and Methods). We observed that 5UIs were generally depleted among SSCR-containing genes (Figure 1; Fisher's Exact Test p = 8×10-8, odds ratio 0.84).
The last class included proteins localized to mitochondria. Nuclear-encoded mitochondrial genes are translated in the cytoplasm and are targeted via an N-terminal leader peptide sequence to mitochondria [13], [14]. We compiled a list of genes with mitochondrial-targeting sequence coding regions (MSCRs), and observed that 5UIs were depleted in MSCR-containing genes (Figure 1; Fisher's Exact Test p = 8×10-6; odds ratio 0.59). This depletion is even stronger than that observed for SSCR-containing genes. Thus, our results showed a general depletion of 5UIs among genes encoding either ER-targeted or mitochondrial proteins.
Next, we tested whether 5UI depletion in SSCR or MSCR-containing genes is a secondary effect of these genes having short 5′UTRs. Although 5UIs are more likely amongst genes with long 5′UTRs (Figure S1A, Wilcoxon Rank Sum Test p <2×10-16; a 99 nt greater median 5′UTR length in 5UI+ than 5UI− genes), we observed that genes encoding secreted and mitochondrial proteins have 5′UTRs that are only slightly shorter than other genes (Figure S1B, Wilcoxon Rank Sum Test p = 2×10-15, p = 9×10-9; a 25 nt and 51 nt difference in median 5′UTR length for SSCR- and MSCR-containing genes, respectively). Even after correcting for the differences in 5′UTR length, SSCR- and MSCR-containing genes were significantly depleted of 5UIs (see Text S1). Similarly, the depletion of 5UIs did not reflect an overall decrease in intronic content, as the total number of bases in non-5′UTR introns did not differ between genes containing or lacking SSCRs (Welch Two Sample t-test, p = 0.34; Figure S2).
A possible link between splicing and genes encoding secretory proteins is the nuclear export of mRNA. Several studies have indicated that export factors are loaded near the 5′ cap co-transcriptionally during the splicing of the more 5′-proximal intron [5], [15]. SSCRs, which similarly promote mRNA export via ALREX [6], are located at the 5′end of the open reading frame (ORF) and could also potentially be recognized by factors co-transcriptionally. Hence, we hypothesized that the 5′-most element in a given transcript, be it an intron or an SSCR, dictates the pathway by which that transcript is exported.
Signal peptide sequences contain a hydrophobic core with amino acids that are naturally encoded by codons with low adenine content. In addition, for pairs of biochemically similar amino acids that differ in the adenine content of their corresponding codons, SSCRs tend to prefer the amino acid with low adenine content codons [6]. We previously showed that adenine depletion in SSCRs is functionally linked to ALREX as silent adenine mutations partially inhibit ALREX [6]. Our hypothesis of a competition between export pathways, driven by whether the 5′-most element is a 5UI or an ALREX signal, predicts that the selection pressure to maintain sequence features important for ALREX-dependent mRNA export would be relaxed in transcripts with 5UIs. We therefore tested whether adenine depletion in SSCRs is attenuated in genes containing 5UIs. Remarkably, we found that SSCRs from genes lacking 5UIs contain 18.2% fewer adenines when compared to SSCRs from genes carrying 5UIs (Figure 2A; Wilcoxon Rank Sum Test p = 4×10-49). Next, we analyzed the amino acid preference of SSCR-containing genes for pairs of biochemically similar amino acids. Specifically, we observed that SSCRs of 5UI− genes have a significantly increased ratio of leucine (which has adenine-poor codons) to isoleucine (which has at least one adenine in all of its three codons) and of arginine (with relatively adenine-poor codons) relative to lysine as compared to SSCRs of 5UI+ genes (Figure 2B–2C; Fisher's Exact Test, p = 3×10-27 and 3×10-40, 95% confidence interval of odds ratio 1.4–1.7 and 1.9–2.4 respectively). SSCRs also exhibit a bias towards synonymous codons that lack adenine [6]. Importantly, this bias diminishes for 5UI+ genes (Figure 2D). This was true for codons for any given single amino acid, such as leucine or serine (Figure S3), or when all synonymous codons were aggregated (Figure 2D; Fisher's Exact Test p = 2×10-42; 95% CI of odds ratio 1.3-1.4). Taken together, our computational analysis indicates that the bias of SSCRs against adenines is relaxed in 5UI+ genes. Furthermore, this reduced bias appears to be due to a relaxation of nucleotide-level constraints, supporting the idea that the presence of 5UIs relieves selection maintaining ALREX signals.
To experimentally investigate this intriguing connection between sequence features in the coding region and the presence or absence of 5UIs, we tested whether SSCRs derived from genes with 5UIs are defective in promoting mRNA export. We inserted SSCR elements into a fragment of the fushi tarazu (ftz), just downstream of the start codon. Furthermore we generated versions of ftz that either contained (ftz-i) or lacked (ftz-Δi) its endogenous intron. Modified forms of these transcripts were previously used to study splicing- and SSCR-dependent mRNA nuclear export [6], [16]. Polyadenylated forms of the ftz mRNA were microinjected into the nuclei of NIH 3T3 mouse fibroblasts. After incubating the cells for one hour, mRNA export was visually monitored by fluorescence in situ hybridization (FISH, Figure 3A) and the amount of mRNA nuclear export was quantified (Figure 3B). Nuclear injection was confirmed by co-injecting fluorescently labeled 70 kD dextran, which is too large to passively diffuse through nuclear pores (see insets, Figure 3A). As demonstrated by several groups, we found that a version of the ftz mRNA that encodes a cytoplasmic protein, but contains neither an intron nor an SSCR (c-ftz-Δi), was not efficiently exported [6], [16] (Figure 3). Nuclear export could be rescued if an intron was incorporated (c-ftz-i). As reported previously, SSCRs from the MHC class 2 gene H2-k1, which lacks a 5UI, promoted efficient export of an intronless version of ftz (Figure 3, MHC-ftz-Δi; see Palazzo et al. [6] and Figure S4 for all ftz variant sequences). We next examined the parathyroid hormone (PTH) and the prion protein (PRP) SSCRs, both derived from genes with 5UIs. Consistent with trends we observed for 5UI+ genes in general, neither PTH nor PRP SSCRs are depleted in adenine content. Furthermore, neither promoted efficient export (Figure 3, PTH-ftz-Δi and PRP-ftz-Δi). Interestingly, elimination of adenines from the PRP SSCR (PRPΔA) only marginally stimulated export (Figure 3, PRPΔA-ftz-Δi) suggesting that this SSCR lacks other features crucial for stimulating export. In summary, only SSCRs from genes lacking 5UIs promoted efficient mRNA export, experimentally demonstrating a functional relevance for the computationally-discovered connection between coding sequence features and 5UI status.
Our investigation into the relationship between 5UIs and alternative export began with the observation that 5UIs were depleted amongst secretory genes. Because 5UIs are also depleted amongst nuclear-encoded mitochondrial genes (Figure 1), we wondered whether related phenomena might be at play. Like secreted proteins, mitochondrial proteins contain a cleavable leader peptide that dictates the ultimate localization of the polypeptide chain [13], [14]. We therefore wondered whether MSCRs exhibit the same nucleotide features that had been associated with ALREX in SSCRs. Indeed MSCRs, like SSCRs, were depleted in adenines overall. Also like SSCRs, this adenine depletion was restricted to MSCRs derived from 5UI− genes (Figure 4A; Wilcoxon Rank Sum Test p = 2×10-9). We found that MSCRs, like SSCRS, tend to encode leucine relative to isoleucine (Figure 4B), and arginine relative to lysine (Figure 4C). Just as with SSCRs, this phenomenon was more pronounced when the elements were derived from 5UI− genes (Fisher's Exact Test p = 0.16 and 10−9, 95% CI of odds ratio 0.9–1.9 and 1.9–3.7 respectively). Finally, only MSCRs from 5UI− genes displayed a bias for synonymous codons that lacked adenine (Figure 4D). This was true for codons coding for any given single amino acid examined, such as leucine or serine (Figure S3), or when results for all synonymous codons were aggregated (Figure 4D; Fisher's Exact Test p = 7×10-06; 95% CI for odds ratio 1.2-1.7).
We next experimentally tested whether MSCRs from 5UI− genes promoted mRNA export in tissue culture cells. Indeed, we found that MSCRs from both the F1 ATP Synthase A (F1) and ferroredoxin reductase (FR) stimulated efficient nuclear export of the ftz transcript (Figure 5A–5B, F1-ftz-Δi, FR-ftz-Δi – see Figure S4 for all modified ftz sequences). We note that the alternative export phenotype observed for these MSCRs is at least as robust as any previously observed for SSCR-containing genes. In contrast, we found that the MSCR from the mitochondrial translation initiation factor 2a (MTIF), a 5UI+ gene, does not promote efficient export (Figure 5A–5B, MTIF-ftz-Δi). Similar to previous observations with the MHC and Insulin SSCRs [6], the introduction of seven silent adenine mutations in the FR MSCR (FR7A) partially inhibited its ability to promote export (Figure 5, FR7A-ftz-Δi).
Microinjected mRNA may behave differently from mRNA that has been endogenously transcribed. Therefore, we microinjected plasmids encoding various ftz transcripts into the nuclei of NIH 3T3 cells. After allowing the plasmids to be transcribed (20 min), further mRNA synthesis was inhibited by treating cells with the RNA Polymerase II inhibitor α-amanitin. Export of the newly synthesized transcripts was assessed two hours after treatment. We found that transcripts produced from plasmids containing FR-ftz-Δi, but not c-ftz-Δi or PTH-ftz-Δi, were efficiently exported (Figure 5C), as was previously seen for MHC-ftz-Δi [6]. Thus, we have shown that MSCR-containing transcripts are capable splicing-independent mRNA export in a manner that depends on 5UI status. This result suggests that the scope of the ALREX pathway extends from ER-trafficked genes to include nuclear mitochondrial genes.
We next wished to assess whether export was dependent on the TAP/p15 nuclear transport receptor, which is required for both SSCR- and splicing-dependent export [6]. We co-injected the viral constitutive transport element (CTE) RNA (known to inhibit TAP/p15 [17]) with the plasmid and observed that export of in vivo-transcribed FR-ftz-Δi was inhibited. Taken together, these experiments indicate that MSCRs and SSCRs from 5UI− genes promote mRNA export using a similar if not identical pathway.
Although our experimental findings supported the importance of adenine-depletion for ALREX, they also indicated that other sequence features may be involved. For example, the PRP SSCR (from a 5UI+ gene) did not promote efficient export even after adenines were eliminated (Figure 3, PRPΔA-ftz-Δi). Furthermore, the incorporation of silent adenines only partially inhibited export by the FR MSCR (Figure 5, FR7A-ftz-Δi), or the MHC SSCR [6]. Therefore, we wished to search for additional ALREX-associated sequence features.
Identification of nucleotide motifs responsible for ALREX function is challenging, because enriched RNA-level motifs might arise due to recurrent patterns at the protein sequence level. Although numerous bioinformatics tools exist to search for nucleotide features (such as transcription factor binding sites) in non-coding regions, few are tailored to the problem of identifying RNA motifs within coding regions. We sought to exploit the idea that we have two collections of SSCRs that differ in the expected abundance of ALREX signals. Specifically, we compared SSCRs from genes with and without 5UIs to identify nucleotide signals exhibiting differential abundance between the sets. Although RNA-level features may be artifactually enriched relative to random RNA sequence due to protein sequence-level constraints, such an artifactual enrichment would not be expected in 5UI− relative to 5UI+ SSCR-containing genes.
We first extended codon usage analyses of the SSCR and MSCR regions to identify other representative signatures. In addition to previously noted adenine depletion, 5UI− SSCR and MSCR genes strongly preferred codons lacking thymine, with a ∼1.4 and a ∼1.7 fold enrichment relative to 5UI+ SSCR and MSCR genes (Figure 6A, Fisher's Exact Test p = 7×10-46 and 4×10-13; 95% CI for odds ratio 1.3–1.5 and 1.5–2.0, for SSCRs and MSCRs respectively).
Next, we searched for primary sequence elements using a discriminative motif finding approach. Specifically, we searched for nucleotide sequences that are significantly enriched among SSCR-containing 5UI− genes relative to 5UI+ genes using the DEME algorithm [18]. We found a likely candidate motif (Figure 6B), which can be roughly described by the consensus sequence CGSSGC (where S represents a mixture of C and G). This motif is highly depleted of adenines and thymines consistent with our analysis (Figure 2, Figure 3, and Figure 6A) and had high information content.
The motif did not show a strong preference for a particular frame of translation (Figure 6B) suggesting that this signal is relevant at the RNA as opposed to protein level. The motif not only appeared in a higher fraction of 5UI− SSCR sequences (47.5% versus 22.2% in 5UI+ SSCRs; see Materials and Methods), but also was much more likely to occur in multiple copies in the SSCRs of 5UI− genes (Figure 6C, 6D; 26.8% versus 7.14%). The CGSSGC motif also revealed a strong positional bias, occurring more frequently toward the 5′ end of coding regions from 5UI− genes (Figure 6E, Figure S5, Wilcoxon Rank Sum Test p = 0.002, median position was 39 and 45 among 5UI− and 5UI+ genes, respectively; see Materials and Methods).
We wished to further examine the question of whether the non-canonical mRNA export function of SSCRs is acting via the same mechanism as that of MSCRs. We therefore tested whether the CGSSGC motif (which was enriched among 5UI− SSCR genes) could also predict the absence of 5UIs among genes with an MSCR. We compared performance of the CGSSGC motif (discovered without use of any MSCR-containing genes) in discriminating 5UI− from 5UI+ MSCRs and found it to outperform at least 99% of 100,000 randomly generated motifs (Figure 6F; False-positive Rate range 10% to 70%; see Materials and Methods). This result indicates that MSCRs and SSCRs, despite differences in the protein sequences they encoded, each play host to a common RNA-level motif associated both with the lack of 5UIs and the ability to support non-canonical mRNA export.
To identify additional motifs, we used the AlignACE [19] algorithm on the set of SSCR sequences from 5UI− genes. This algorithm has the advantage that it can identify multiple nucleotide sequences and allows greater flexibility in motif length. We filtered the discovered sequences for their discriminative ability and found 19 motifs that were significantly enriched among 5UI− relative to 5UI+ genes (Table S2, see Materials and Methods). The discovered motifs displayed mutual similarity and included several close variants of the CGSSGC motif discovered by DEME (See Figure S6 for the PSSM logos of the most discriminative AlignACE motifs).
We next focused on the properties of the four most discriminative AlignACE motifs. All four motifs were more likely to occur in multiple copies among the 5UI− genes compared to 5UI+ genes (Figure S7). Even though these four motifs were discovered based on their ability to discriminate 5UI− from 5UI+ genes among those genes with SSCRs, these motifs were also predictive of 5UI absence for genes with MSCRs (Figure S8). All four motifs performed in the top quartile compared to 100,000 random motifs (Figure S8; see Materials and Methods). However, unlike the CGSSGC motif, three of these motifs displayed a significant bias for occurring in a particular frame of translation. These three motifs may thus be detecting protein sequence-level differences between 5UI− and 5UI+ genes (Figure S9). In fact, consensus sequences of many AlignACE motifs included CTGs that can encode leucines, which were highly enriched among SSCRs and MSCRs from 5UI− genes relative to their 5UI+ counterparts.
We next decided to test whether synthetic elements matching the discovered motifs could promote the export of ftz mRNA. We used versions of the ftz-Δi mRNA containing either three copies of an element matching the consensus CGSSGC motif (M1-ftz-Δi), a CUG repeat-containing element (M2-ftz-Δi), or a single copy of each (M3-ftz-Δi see Figure 7A for the sequences of all these constructs). We chose CUG repeats as they appeared in many of the consensus sequences of AlignACE motifs (Table S2). In addition, there are several RNA binding proteins, such as CUG-BP1 [20] and the Muscleblind family of proteins [21] that are known to recognize CUG repeats.
To assay for export activity we microinjected plasmids that contained versions of the ftz gene fused to segments containing elements matching ALREX-enriched motifs and their combinations (Figure 7A) into the nuclei of NIH 3T3 cells. After allowing the plasmids to be transcribed (20 min), further mRNA synthesis was inhibited by treating cells with α-amanitin. We found that all three motif-containing ftz constructs (M1-, M2-, M3-ftz-Δi) were exported more efficiently than c-ftz-Δi but substantially less efficiently than MHC-ftz-Δi mRNA (Figure 7B). Adenine depletion was required for export, as mRNA generated from plasmid containing a mutant form of M3-ftz-Δi bearing four silent adenine mutations (4A-M3-ftz-Δi, see Figure 7A) collectively disrupting each of the two component elements was not efficiently exported (Figure 7B–7C). To further validate these results, we transfected plasmids encoding the motif-containing ftz genes with elements corresponding to these motifs into COS-7 cells and measured the steady state distribution of mRNA. In agreement with our microinjection experiments, we found that the three motif-containing ftz constructs were exported to a level that was clearly higher than c-ftz-Δi but lower than MHC-ftz-Δi (Figure 7D). As observed for microinjected NIH3T3 cells (Figure 5), mRNA generated from a plasmid containing the 4A-M3-ftz-Δi construct was not efficiently exported from transfected COS-7 cells (Figure 7D).
The function and evolution of introns has been intensely studied since their discovery (reviewed in [22], [23]). Despite the presence of a large number of introns in untranslated regions, especially in the 5′ untranslated regions of transcripts, these studies have been largely focused on introns in coding regions [3]. We established that the distribution of 5UIs in the human genome is non-random, with specific functionally related groups of genes being enriched [1] or depleted (this study) for 5UIs. Here we show that, in both secreted and mitochondrial genes, the presence or absence of 5UIs correlates with sequence features at the beginning of the coding region. Minimally, our results further support the conclusion that complex selective forces govern the evolution of 5′UTR introns. Moreover, our results are best explained by the existence of a regulatory mechanism that is both special to 5UIs and has relevance to thousands of genes across the genome.
Our results show that nuclear transcripts encoding both secretory and mitochondrial proteins share RNA-level signals capable of directing mRNA export, even for an intronless message. It has frequently been observed that mRNAs of functionally related genes are co-regulated at the post-transcriptional level (‘the regulon hypothesis’ [24]). Our results suggest that, consistent with this phenomenon, the ALREX pathway can facilitate coordinated expression of functionally related genes at the level of mRNA export. Moreover, our analyses support a model whereby the first transcript element emerging from RNA Polymerase II during transcription—be it an intron or an ALREX-promoting element—determines which RNA export pathway is predominantly followed (Figure 8). Under this model, presence of a 5UI would supersede downstream SSCR or MSCR export signals and relax selection pressures that maintain ALREX-promoting sequence features.
Although we have made progress in defining some sequence features that mediate the ALREX function (see Figure 6 and Figure 7), it is clear that a more extensive description of ALREX features individually and in combination will be quite useful. We found specific nucleotide-level motifs in the 5′ end of coding regions which discriminate between genes with and without 5UIs. Substantial future efforts will be required to combine information about 5UI absence with the presence and placement of ALREX signals within a unified framework that can predict ALREX activity. This information could be used to compile a full list of transcripts using the ALREX pathway. It will be interesting to determine whether other genes, such as those that encode membrane-bound proteins but lack a signal sequence (and hence an SSCR), can use this alternative export pathway.
The most fundamental challenge for future studies will be to understand the biological role or roles of the ALREX pathway. Why is its selection maintained even in transcripts that contain coding region introns and are therefore enabled to use the canonical mRNA export pathway?
Although the functional downstream consequences of using either the splicing-dependent or ALREX-pathway remain unknown, silent mutations within the SSCR not only impair mRNA export but also disrupt proper ER-targeting of the transcripts [6]. This suggests that multiple post-transcriptional events, such as mRNA export, mRNA transport in the cytoplasm and mRNA translation, are coupled [25].
Here, we have discovered and validated two motifs that promote mRNA export, suggesting that ALREX may recruit more than one nuclear factor. Such factors could not only dictate RNA export but perhaps also dictate how the mRNA is distributed and translated once in the cytoplasm. Investigation of these questions awaits identification of ALREX factors, and of mRNA localization or other phenotypes associated with disrupted ALREX function. One of the motifs we discovered is a long CUG repeat that could potentially bind to CUG binding proteins. However, MHC-ftz-Δi mRNA is exported from HeLa cells that were depleted of both MBNL1 and MBNL2, two members of the muscleblind family of CUG-repeat binding proteins (unpublished findings), suggesting that these are not the responsible factors. Identification of the ALREX-element binding protein(s) will shed light onto how ALREX operates and provide insight into the biological role of this pathway.
The question of biological role is particularly intriguing in the case of nuclear-encoded mitochondrial genes. The textbook description of nuclear-encoded mitochondrial genes has translation of these genes occurring within the general pool of cytoplasmic proteins, with subsequent protein localization due solely to the mitochondrial targeting peptide sequence. However, there is evidence that nuclear-encoded mitochondrial transcripts can localize to the vicinity of mitochondria prior to translation [26], [27]. Although we do not detect any mitochondrial targeting of MSCR-bearing transcripts (Figure 5), it is possible that a fraction of these mRNAs are indeed localized. It will be interesting to learn what role ALREX could play in the localization and translation of nuclear-encoded mitochondrial genes.
Substantial future studies will be required to further explore mechanisms of the ALREX pathway. For example, it is unclear whether ALREX signals are inhibited by other complexes deposited on the transcript in a splicing dependent manner. One example is the Exon Junction Complex (EJC), which potentiates the translation of properly spliced mRNA [28], [29] and the nonsense-mediated degradation of improperly spliced transcripts [30], [31]. Some mRNAs, such as those of PrP and PTH genes, encode secreted proteins but lack any ALREX-promoting element. For such mRNAs, it is possible that the proper ER- targeting and efficient translation of these transcripts requires the recruitment of the EJC or TREX components to the 5′UTR. Identification of the nuclear proteins that associate with ALREX elements, and how these factors are coupled to other processes, will yield significant insight into the role of ALREX in mediating gene expression, and localization of both mRNAs and proteins.
NCBI's human Reference Gene Collection (RefSeq) [32] and the associated annotation table, retrieved from the UCSC genome browser genome assembly May 2004 (http://hgdownload.cse.ucsc.edu/downloads.html), were used to extract a high confidence set of 5UIs. The lengths of 5′UTR-associated genomic features were determined using RefSeq intron-exon definitions (downloaded June 2007). Out of a total ∼24.5 k RefSeq transcripts, ∼8.5 k contained at least one intron. Genomic coordinates of 5UIs examined were as previously described [1]. When multiple splice variants involving a given 5′UTR exhibited identical splicing patterns within that 5′UTR region, a single identifier was selected randomly as the representative for that 5′UTR.
For the remaining transcripts, total lengths of coding region introns were determined from the RefSeq Annotation (downloaded from UCSC genome browser, May 2004 genome assembly on May 15th 2009 http://hgdownload.cse.ucsc.edu/downloads.html).
DNA constructs encoding ftz isoforms were assembled by first digesting the pBR322 plasmid containing c-ftz-Δi [6] with Nco I and ligating oligonucleotides encoding various SSCRs and MSCRs (see Figure S4) so that the extra sequences were all inserted just downstream of the start codon. The constructs were then transcribed into mRNA, which was then polyadenylated, purified and then microinjected into NIH 3T3 fibroblast nuclei at 200 µg/ml with Alexa488 conjugated 70 kD dextran (1 mg/ml) as previously described [6], [33]. DNA microinjections were performed as previously described [6]. Briefly, ftz isoforms were subcloned into pCDNA3 using Hind III and Xho I and microinjected at 50 µg/ml with Alexa488-conjugated 70 kD dextran (1 mg/ml) into NIH 3T3 fibroblast nuclei. After allowing the RNA to be transcribed for 20 min, the cells were treated with α-amanitin (50 ng/ml) to prevent further transcription. CTE RNA was synthesized as previously described [6] and microinjected at a concentration of 200 µg/ml along with DNA and Alexa488-conjugated 70 kD dextran. All microinjected cells were incubated for the indicated time to allow for mRNA export, then fixed with 4% paraformaldehyde in phosphate buffered saline (PBS). DNA transfections into COS-7 cells were performed as described previously [6]. Transfected cells were incubated for 12–18 hrs, then fixed with 4% paraformaldehyde in PBS. The ftz mRNA was stained by fluorescence in situ hybridization followed by imaging and quantification of RNA nuclear export as previously described [6]. Cell imaging and mRNA quantification were also performed as previously described [33].
FuncAssociate [34], [35] beta version was used for Gene Ontology (GO) analysis, and Synergizer [36] was used for mapping RefSeq IDs into the ‘namespace’ of GO association files using Ensembl as the synonym authority. We restricted the space of genes in which GO correlations were sought to RefSeq because our 5UI genes were drawn only from this set. To quantify the effect size of GO correlations, the results in Table S1 were sorted according to their log10 odds ratio, with significance calculated by Fisher's Exact Test as previously described [35]. Multiple hypothesis correction was achieved via a resampling approach that preserves the dependency structure between the tested hypotheses [35]. Adjusted p-values were calculated using 10000 resampling simulations.
We retrieved the complete set of transcripts with signal peptide annotations from the Ensembl 50 database using Biomart [37] (downloaded on February 2009 http://www.ensembl.org/biomart/martview). Of the 38396 transcripts in this database, 4953 were annotated as having a signal peptide, and 4704 of these were in our set of RefSeq genes. The coding region sequences for all the genes in our set were downloaded from NCBI Refseq Collection release 33 (ftp://ftp.ncbi.nih.gov/refseq/H_sapiens/mRNA_Prot). The ratio of the amino acids, total adenine counts and the codon usage bias were calculated for the first 69 nt and the rest of the sequences. There were 135 coding region sequences that had a length that was not a multiple of three. These sequences in addition to those with total length less than 150 nt were removed from further analysis.
The list of mitochondrial genes was retrieved from the Organelle DB [38] website (downloaded on February 2009 from http://organelledb.lsi.umich.edu/). Identifiers were translated to RefSeq ID using Synergizer [36]. Nine genes were removed from this list as they were encoded by the mitochondrial genome. For some genes, Synergizer could not find a RefSeq ID corresponding to the “standard name”. These genes were manually inspected and the synonyms provided by Organelle DB website were used to find corresponding RefSeq IDs. When multiple splice variants were exact duplicates with respect to the first 69 nts of their coding region, a single identifier was selected as the representative. This procedure yielded 364 RefSeq transcripts out of ∼25 k transcripts having an MSCR. The manually edited list of mitochondrial genes is available in Dataset S1. The software package R 2.6.0 was used for all the statistical analyses, except where otherwise noted.
For motif discovery, the first 99 nt of SSCR-containing genes were used to ensure that all signal peptides were included in their entirety. Highly similar sequences were removed to avoid overweighting closely related sequences. Specifically, the first 99 nt from each sequence was aligned to all others using blastn [39]. A threshold (E-value <10−25) was used to group similar sequences, and one randomly selected representative from each such set was used after this filter.
We used the DEME [18] software to search for a motif that is highly enriched in the 5UI− set of sequences relative to the 5UI+ set. We also used the AlignACE software [19] to search for a set of highly enriched motifs in the 5UI− set. AlignACE searches for frequently occurring motifs in both the forward and complementary strands of DNA sequences. Choosing to focus on RNA motifs, we discarded 2 of the 20 motifs reported that were constructed from less than 10 representative forward-strand sites.
Positional Specific Scoring Matrices (PSSM) of the discovered motifs were extracted from the forward-strand sites of each motif. For a given sequence s and a motif with length m, all windows of size m within the first 99 bases were scored using the PSSM of the motif. To avoid calling multiple overlapping motifs, only the highest scoring window in a contiguous series of overlapping windows was selected. For each motif, an initial PSSM score threshold (t*) was selected such that t* yields the highest enrichment of motif-containing sequences among the SSCR-containing and 5UI− genes on the p-value generated from Fisher's Exact Test using the 2×2 contingency table (Table 1).
Given the total number of genes N, the number of 5UI− genes m, and the number of motif-containing sequences n, this test estimates the probability that k or more genes would be found to overlap between the 5UI− genes and the motif-containing sequences under the null hypothesis of independence:where the probability of observing exactly i overlaps given N, m and n follows from the hypergeometric distribution:
Among the 18 AlignACE motifs, we focused on the four that were most enriched among 5UI− genes compared to 5UI+ genes based on the resulting p-value. Further analyses on motif occurrences and positional distributions were performed on these four AlignACE motifs and the DEME motif.
While PSSM threshold selection using Fisher's Exact Test provided a quick way identify discriminative AlignACE motifs, the selection of thresholds did not take into account the likelihood that such discrimination may have occurred by chance.
To account for this possibility, we randomly generated four sets of PSSMs matching the discovered motifs' lengths (6, 10, 14, and 16 nt). We modeled each position of the PSSM as an independent sample from a Dirichlet distribution with parameters (αi) equal to the background nucleotide frequency such that Σαi = 1. The background nucleotide frequency was calculated among the first 99 nts of either SSCR-containing or MSCR-containing 5UI− genes. For each given motif length, we generated 40,000 random PSSMs for SSCR set and 100,000 random PSSMs for the MSCR set. We generated receiver operating characteristic (ROC) plots to compare the discriminative performance of these randomly generated PSSMs with that of the discovered motifs. First, we scanned each sequence to find the maximum score for each PSSM. We classified a sequence as motif-containing if its maximum PSSM score was greater than a given threshold t*. For all random and discovered motifs, we calculated the true positive rate (TPR) as the fraction of motif-containing 5UI− genes, and the false positive rate (FPR) as the fraction of motif-containing 5UI+ genes as a function t*. Therefore, each point on an ROC plot corresponds to (TPR, FPR) of a particular PSSM at some threshold t*. These ROC plots are informative about the analyzed motif's power to discriminate 5UI− from 5UI+ genes.
For each discovered motif, we used the ROC plots generated from SSCR-containing genes (Figure S8) to choose the PSSM score threshold value (t′) for subsequent analysis. The threshold t′ was chosen such that it maximizes the difference between the discovered motif's TPR and the median TPR of the random motifs was the most at the FPR value corresponding to t′. Since we discovered motifs using the SSCR-containing set only, the ROC plots for the MSCR set were not subject to any overfitting that might have occurred during motif discovery.
To assess whether there is any significant deviation in the positional distribution of motifs in the 5UI− set from that in the 5UI+ set, we performed the Wilcoxon Rank Sum test. We examined differences in distributions for the positions of all motif occurrences in each sequence. We also generated histograms for the reading frame at which motifs occur in the coding region to look for differences between the 5UI− and 5UI+ sets.
|
10.1371/journal.pcbi.1003921 | Linking Macroscopic with Microscopic Neuroanatomy Using Synthetic Neuronal Populations | Dendritic morphology has been shown to have a dramatic impact on neuronal function. However, population features such as the inherent variability in dendritic morphology between cells belonging to the same neuronal type are often overlooked when studying computation in neural networks. While detailed models for morphology and electrophysiology exist for many types of single neurons, the role of detailed single cell morphology in the population has not been studied quantitatively or computationally. Here we use the structural context of the neural tissue in which dendritic trees exist to drive their generation in silico. We synthesize the entire population of dentate gyrus granule cells, the most numerous cell type in the hippocampus, by growing their dendritic trees within their characteristic dendritic fields bounded by the realistic structural context of (1) the granule cell layer that contains all somata and (2) the molecular layer that contains the dendritic forest. This process enables branching statistics to be linked to larger scale neuroanatomical features. We find large differences in dendritic total length and individual path length measures as a function of location in the dentate gyrus and of somatic depth in the granule cell layer. We also predict the number of unique granule cell dendrites invading a given volume in the molecular layer. This work enables the complete population-level study of morphological properties and provides a framework to develop complex and realistic neural network models.
| Computational models of neurons and neural networks provide a valuable avenue to test our understanding of brain regions and to make predictions to guide future experimentation. Each neuron has a unique dendritic tree, features of which can vary depending on the location of the neuron within the particular brain region. In this study, we generated a complete population of dendritic trees for the most numerous type of neuron in the hippocampus, the dentate gyrus granule cell, using a realistic three-dimensional structural context to drive the generation process. Morphological properties can now be studied at the level of complete neuronal populations, and this work provides a foundation to build upon in the construction of large-scale, data-driven neuroanatomical and network models.
| Growing evidence for the importance of dendritic structure on neuronal function has inspired the construction of morphologically realistic computational models of single neurons. Dendritic morphology has been shown to have a significant impact on neuronal firing properties, both between neurons of different classes [1] and within the same class [2], [3], as well as on signal integration and propagation [4]–[6]. The intra-class morphological variability could have a significant impact on the integration of individual neurons into the circuit and their resulting role in network computation. Correspondingly, this has led to the development of detailed three-dimensional morphological reconstructions of single cells [7] and functional models incorporating this level of detail [8], [9]. Not only does the incorporation of realistic morphology enable more accurate reproduction of measured electrophysiology, it also allows for a more detailed representation of network connectivity. These together enable a better understanding of the underlying computation in the network. Advances in computational power as well as in parallel computing, such as the development of parallel versions of neurophysiological simulation environments [10]–[14], have made the simulation of large networks with detailed neuron models accessible. Currently, however, the majority of functional electrophysiological network models utilize uniform single models or very small subsets of models to describe neurons of a given class, overlooking the inherent biological diversity. In addition, the connectivity is usually oversimplified in almost all functional neural networks, whether by the use of probabilistic methods rather than explicit connectivity or by making connections using only a subset of the neurons in the network population, while in most applications it should in fact reflect the full morphological architecture of dendrites and axons. The generation of full-scale, population-level morphological models is, therefore, an important and timely goal. Since experimental reconstructions are to date available only in small sample sizes, techniques to generate population-level morphological models will require the amplification of these data sets to fully realistic and diverse populations [15].
Aside from quantifications based on reconstructions of single cells, existing neuroanatomical data encompasses a large number of measures at multiple levels, such as density estimates for synaptic zones using electron microscopy or cell counts and population analysis using molecular techniques as well as entirely macroscopic features of neural tissue [16]. The optimal arrangement of elements of neural circuits in the brain has been extensively studied [17]–[21], and a recent trend has been to put neuroanatomical single cell reconstructions in the macroscopic context in which they originally existed [22]–[24]. In particular, recent work in Drosophila has focused on generating a standardized structural model [25] and taken steps toward generating a complete network connectivity map by placing all reconstructed neurons into a standard brain [26], which is possible given the smaller population of neurons in invertebrate model organisms. Conventional light microscopy does not have the resolution to reconstruct circuits in densely labeled neuropil [27], and as a result, modern techniques such as large-scale serial block-face scanning electron microscopy have started to provide reconstruction methods for which both the microscopic details of all cells and the macroscopic circuit-level features are present in the same biological tissue samples [28]–[30]. These data, however, are rather large and complex, and it will be important to develop novel approaches to facilitate the study of such neuroanatomical connectomes [31]. The development of large-scale morphological models with macroscopic constraints will enable the analysis of these large data sets and the study of connectomes before full anatomical reconstructions are available.
Current methodologies for the generation of morphological models primarily employ reconstructions and their branching characteristics independent of their originating context. Several studies have relied solely on the reconstructions themselves, involving pure duplication [32], making small variations in the lengths and angles of tree branches [33], or resizing to fit within a spatial context [24]. Other methods have focused on the branching properties of the reconstructions and have used a wide variety of algorithms, including the simulation of growth cones with NETMORPH [34], modeling self-referential forces [35], or mapping one-dimensional structures to 3D trees [36]. Several of these tools, such as L-Neuron [37], EvOL-Neuron [38], and NeuGen [39], [40], create variable dendritic trees by stochastically sampling branching parameters from extracted statistical distributions. While the stochastic sampling methodology is able to generate realistic synthetic trees, it was too inefficient in our previous work to generate a complete and distributed population [15], even without the constraint of fitting within a three-dimensional context. The current study reverses the direction of previous methods by starting with the macroscopic neuroanatomy and enables complete population-level construction and analysis.
Here we report a method that allows us to match generated single cell morphologies to measured data as a function of macroscopic features. We do this by devising a computational model that generates morphologies of all single neurons in a population while considering the broader neuroanatomical context in which they grow. First we model the volume of the rat dentate gyrus based on a recent detailed reconstruction of the entire structure [24]. We then generate single cell morphologies of all granule cells (GCs) as described previously [41] constrained within this volume. The resulting population data matches the known variability in GC morphology as well as some known key dependences of GC features on location within the dentate gyrus. We then use our model to develop measures and predictions for dendritic features at the population level. This work provides a valuable framework for the study of complete populations of neuronal morphologies and represents a major step in the development of large-scale neural network models.
In order to grow dentate gyrus granule cell (GC) dendritic tree structures within their structural context, we first generated a parameterized volume representing the dentate gyrus (DG) shape. Smoothed surfaces for the boundaries of the DG granule cell layer (GCL) and molecular layer (ML) were obtained from a recent high-resolution, 3D serial reconstruction of the rat hippocampus [24]. Parametric 2D manifolds were then fitted to these boundary surfaces (Figure 1A; see also Methods for detailed equations) in order to provide a coordinate system in which depth in the GCL and ML as well as the septo-temporal and infra- versus suprapyramidal axes are mapped. This in turn enabled the subdivision of the ML volume into inner (IML), middle (MML), and outer molecular layers (OML) using intermediate surfaces, since several aspects of GC morphology have previously been associated to these reference structures. The resulting model DG closely matched the structural features of the experimentally reconstructed volume (Figure 1B). The model DG had the same overall GCL volume, 3.78 mm3, and ML volume, 9.02 mm3, as the experimental reconstruction. Also, the ML width throughout the structure, 247±33 µm, closely matched a previous experimental measurement, 249±33 µm [42]. Slices from the model DG possessed the characteristic curved structure of the biological dentate gyrus, which is known to be more “V”-shaped in the septal region and “U”-shaped in the temporal region (Figure 1C). The volume created by the parametric surfaces thus served as a realistic structural context within which to drive GC dendrite generation. Furthermore, the parametric character of the surfaces subsequently enabled the mathematical tractability of the transformation between a planar two-dimensional sheet and the curved two-dimensional manifolds in 3D space.
The dendritic field spanned by the GC dendritic tree can be approximated by an elliptical cone [41]–[43]. GC trees can be synthesized by connecting a somatic coordinate to target points distributed in such cone-like volumes while minimizing total dendrite length as well as path lengths within the dendrite [41]. To generate the complete GC forest, somata were first distributed within the GCL. The rat DG GCL is estimated to contain approximately 1.2 million tightly-packed GCs [44], [45], and the GC soma is an ellipsoid with an average width of 10.8 µm and height of 18.6 µm [42]. For the purposes of this study, spheres with 12.54 µm diameter corresponding to the volume of an average GC ellipsoid soma were arranged on a large hexagonal grid. Those spheres with any portion located outside of the GCL volume were discarded, and the remaining spheres were selected as somata for the GC population (Figure 2A). In this way, 1.19 million somata separated by a 3.5 µm distance were well-packed within the GCL volume. An elliptical cone could now be placed at each of these soma locations to select the target points necessary to grow GC dendritic trees within the ML boundaries.
The optimal wiring algorithm connects points by performing a dual minimization of total dendritic length and path lengths, under a constraint (balancing factor bf) that weighs the importance of one over the other. Low values for bf lead to strongly minimizing the total wiring which can result in long conduction paths to the soma, while larger bf values lead to trees with short conduction times. Target points for all cells were first distributed in the GCL and ML according to proportions estimated from experimentally reconstructed dendritic trees (Figure 2B, see also Figure S1). Points within each elliptical cone (Figure 2C, shaded area) were then isolated, and a subset of these points (Figure 2C, larger dots) was selected to result in realistic numbers of branch and termination points per layer when connected. These target points were then connected using the optimal wiring algorithm, which results in specific portions of target points becoming branch, continuation, or termination points in the tree depending on the balancing factor [46] (Figure 2D). Spatial jitter of two different spatial frequencies was added to reproduce the tortuosity of real dendritic trees (Figure 2E; see Methods). Finally, a realistic quadratic tapering in diameter was mapped onto the dendritic topologies (Figure 2F), based on both the tapering present in real granule cells and previous work showing that a quadratic taper optimizes synaptic democracy [47], or the equalization of current transfer between all dendritic locations and the root. This process was then repeated for each GC in the DG, varying the parameters to reproduce the variability in the resulting population (details in Methods).
Synthetic GC dendritic trees were statistically and visually indistinguishable from real GCs. Since the generative wiring algorithm connects target points to form tree structures, it is an important validation of the procedure that both the laminar distribution of branch points and of dendritic length in the synthetic GC population matched the experimental data [42] (Table 1). Example dendritic topologies are shown in Figure 3A. Experimental reconstructions and synthetic dendritic trees had similar branching properties, exemplified by the classical Sholl analysis [48] (Figure 3B), for which the number of intersections between the dendrite and a sphere of increasing diameter centered on the dendrite root are counted. Also, the distributions for contraction values (the ratio of Euclidean distances and path distances for all branches in the tree) were similar between reconstructed and synthetic GCs (Figure 3C), validating the balancing factor between costs of total dendrite length and path distances in the synthetic trees as well as the added spatial jitter. Because diameter measurements are not available for our reference GC morphologies [42], the diameter tapering was constrained to a more recent set of experimental reconstructions [49] independent of the context-dependent study. The match of the diameter tapering between reconstructed and synthetic GCs is visualized in Figure 3D. The branching structure and diameter tapering of the synthetic trees were thus indistinguishable from experimental reconstructions.
The complete forest of 1.19 million synthetic GC dendritic trees was constructed by varying the parameters in the generation process (i.e., cone radii, number of stems, total number of nodes, laminar distribution of nodes, balancing factor, amplitude of spatial jitter, and diameter taper) for each individual GC. The resulting properties of the complete population matched values from reconstructed granule cells (Table 2), with small differences arising from a different relative composition of GCs from context-dependent subgroups. Significant differences have been described in GC dendritic morphology depending on the location of the soma within the GCL, i.e. for GCs with somata in the suprapyramidal versus infrapyramidal blade as well as in the deep versus superficial parts of the GCL [42]. Accordingly, choosing parameters in the generation process based on the location of each generated granule cell somata allowed for each statistically significant context-dependent difference reported previously to be recreated in the synthetic tree population (Table 3, p<0.001 for all comparisons, Student's t-test). For example, the balancing factor for the suprapyramidal deep granule cells was set to 0.9 to match the higher maximum branch order, which was lower than the 1.35 value used for suprapyramidal superficial granule cells and the 1.22 value used for infrapyramidal granule cells. The lower balancing factor in the suprapyramidal deep subgroup signifies that minimizing dendritic length is more important for these cells. A complete and realistic population of 1.19 million context-dependent GC dendritic trees was created that matched the observed biological variability and recreated context-dependent differences, in addition to fitting within a realistic three-dimensional DG structure (Figure 4A–C). The distribution of these trees within the neuroanatomical space enables the study of the input organization and spatial occupancy of the complete GC forest.
In the following, we show how simple analyses that become possible with such a model can be informative about the network constituency in the hippocampus and about the location-specific distances of dendritic structure within the DG volume. An important question regarding the connectivity in the circuit is to know how many unique GCs an axonal arborization would reach within a given volume of the molecular layer. This can now simply be visualized as exemplified for a sample transverse slice from the center of the dentate gyrus (Figure 4D), divided into 25 µm cubic volumes. The overlap of dendrites from unique GCs in each sample volume (312±114 GCs, range 45 to 650) is a small portion of the 1.19 million GC population, signifying that the macroscopic neuroanatomy of the DG promotes a sparse connectivity. This large range also results in a diverse amount of complexity required for an axon to arrive within 5 µm of all GCs in the sample cubic volumes (64±22 branch points, range 17 to 142, see Methods for axon construction details). The distribution was location dependent, as there was a greater overlap in unique GC dendrites in the OML versus the IML, which coincides with the increased dendritic length in the OML (Table 1) and an increased cable density, i.e. dendritic length per volume (Figure 4E). The volume occupied by the GC forest, on the other hand, decreased toward the OML (Figure 4F), signifying that the increased cable density in the OML does not counteract the diameter tapering implemented into synthetic GC dendrites. The overlap of unique GCs, cable density, and volume occupied by GC dendrites were larger in the supra- and infrapyramidal blades as opposed to the crest, which is in accordance with the increased ML volume at the crest reported in the experimental reconstruction [24]. There were significant positive correlations between the overlap of unique GC dendrites, cable density, and volume occupied (Figure S2, p<0.001). However, the correlation between the number of unique GC dendrites and cable density (r = 0.95) was much stronger than the correlation between the number of unique GC dendrites and volume occupied or the cable density and volume occupied (r = 0.27 and 0.50, respectively). This likely resulted from the implemented variability in diameter tapering.
While the occupancy features impact strongly on network connectivity, measures of dendritic morphology that vary along with spatial coordinates in the DG impact strongly on the electrotonic constituency and resulting dendritic computation and synaptic integration in individual GCs. Using our complete population model, we can compare simple measures such as total dendritic length (Figure 4G) and maximum tip distances (Figure 4H) in a location-dependent manner. Even ignoring the difference between GCs from the suprapyramidal and infrapyramidal blades since these were directly incorporated into the model, the total dendritic length varied by a factor of 2× between the most distal septal or temporal tips of the DG as compared to the center of the model. The transverse axis, or size of the “C”-shape, is higher toward the center compared to the septal and temporal tips in the experimental reconstruction [24], so the GCs in the center have an increased length in order to reach the outer edge of the OML. GCs with somata deep in the GCL had less variability in total length compared to the GCs with somata in more superficial parts of the GCL (Figure 4G). As expected, maximum tip distances in deep GCs were longer than in superficial GCs (Figure 4H).
We therefore have provided here simple measures linking the macroscopic scale of the DG volume with the microscopic details of single neuron morphologies and extracted useful information for network connectivity and neural computation. In future studies, novel population-level measures can be designed and tested utilizing this framework as a foundation.
In the present study, we used a realistic structural context based on a reconstructed rat dentate gyrus [24] to drive the generation of dendritic trees with a recently developed algorithm based on optimal wiring constraints [47], [50]. By varying the relatively few parameters in the generation process, we were able to reproduce the observed biological variability in the morphology of dentate gyrus granule cells and match key location-specific differences. While some properties were obtained from parameter optimization, several features were emergent and not the result of direct parameter constraints, including the total dendritic length, branch lengths, path lengths, and asymmetry in Table 2 as well as the Sholl intersections in Figure 3. In addition, all population-level measures, such as the cable density, are emergent properties.
The set of synthetic dendritic trees represents the largest collection of realistic morphologies to date, a complete forest of 1.19 million granule cell dendritic trees, with each tree requiring less than two seconds to be constructed. The method that we devised enables population-level analysis, and we can link the larger neuroanatomical features with the resulting branching characteristics. Due to the small number of existing reconstructions that are registered to a macroscopic context and the limited information about the properties of granule cells in the crest, granule cells were split into subgroups differentiating infrapyramidal versus suprapyramidal and deep versus superficial granule cells based on previous reconstructions [42], and a single balancing factor parameter was specified for each group. As more context-aware reconstructions become available, this abrupt transition can be modified to create a more continuous variation of the parameters. The speed conferred by utilizing parallel computing in the generation process and the relatively few parameters involved provide flexibility to incorporate future experimental observations to improve the model.
While the current model provides a valuable framework for the exploration of macroscopic and microscopic neuroanatomical links, there are inherent simplifications that deviate from the biological condition that should be improved upon in future studies. The current generation process allows for multiple dendrites to occupy the same point in space, so a form of avoidance could be implemented into the spatial tortuosity, instead of solely low-pass filtered noise, in order to create a more realistic spatial occupancy. In addition, the packing of spherical somata can be improved to implement the variable and tightly-packed elliptical somata observed in experimental studies [42]. The current study also does not include the newborn granule cells, which constitute approximately 10% of the total granule cell population [51], transiently exhibit basal dendrites [52]–[54], and possess a significantly smaller total dendritic length [53]. This subpopulation has recently come under intense focus for their unique participation in hippocampal network functions [55]–[57], and the neuroanatomical properties of this subpopulation could be contrasted with the more numerous mature granule cells constructed in this study.
The linking of macroscopic and microscopic neuroanatomy presented in this study provides a framework that can be expanded upon with additional cell types and axons, but it also provides an avenue to link neuroanatomical features with electrophysiological function. The breadth of anatomical data being collected, including recent experimental reconstructions of excitatory mossy cells [49] and inhibitory interneurons [58], [59] in the dentate gyrus, will make it possible to construct even more biologically realistic DG models. In addition, the context-driven generation methodology can be applied to axons to create realistic connectivity for comparison to the growing connectomics literature. As noted in the introduction, dendritic morphology can have a dramatic impact on electrophysiological function, and the framework provided in this study allows for this relationship to be studied on the level of the complete population. All generated morphologies can be exported to simulation environments [60] for the insertion of ion channel conductances or other biophysical mechanisms. For the example case of granule cells, the measured properties of dendritic integration [61], [62] and action potential initiation [63] should serve as valuable constraints. This structure and function relationship can eventually be linked to both the macroscopic neuroanatomical and network context.
The model dentate gyrus structure and granule cell synthetic trees were created and analyzed in MATLAB using the TREES toolbox [41], [60] on University of California Irvine's High Performance Computing cluster. The model structure and generation process will be made available at ModelDB (http://senselab.med.yale.edu/ModelDB/). The standard deviations for the literature values [42] were determined by multiplying the reported standard error by the sample size. All values are presented as mean ± standard deviation.
The following parametric equations defined the layer boundaries:where v defined the “C”-shape and ranged from −0.23π to 1.425π, u defined the septotemporal extent and ranged from 0.01π to 0.98π for the GCL and −0.016π to 1.01π for the ML, and L defined the layer and was −1.95 for inner GCL, 0 for outer GCL, 1 for IML, 2 for MML, and 3 for OML. The experimental reconstruction GCL volume was calculated based on the average 0.6 GCL to hilus volumetric ratio and their combined volume of 6.30 mm3 [24]. The experimental value for the molecular layer width was determined by combining the means and standard deviations for the infrapyramidal and suprapyramidal group measurements (240±17 and 254±3, respectively) reported in a previous study [42]. The model ML width was determined by distributing 2 million points on the outer GCL and OML boundary, and then calculating the closest distance to the OML boundary from 10,000 randomly sampled outer GCL points.
In order to recreate the context-dependent differences in the synthetic tree population, the size of the elliptical cone, number of stems, total number of nodes, and balancing factor governing the wiring were modified based on the location of each soma. Superficial neurons were defined as having somata in the half of the GCL closest to the ML, whereas deep neurons had somata in the half farthest from the ML. The infrapyramidal/suprapyramidal split was located halfway around the characteristic “C”-shape of the transverse slice of the dentate gyrus, which was defined by the midpoint of the v parameter in the GCL boundary equation. The number of stems was set by sampling from a truncated Poisson distribution and ranged from 1 to 4, as observed in experimental reconstructions [42], [49]. The elliptical cone was oriented by pointing the center axis toward the closest of two million points distributed on the OML boundary and orienting the longitudinal and transverse elliptical cone radii within the structure. The transverse spread of generated GCs was analyzed by orienting cells based on their mean transverse axis and measuring the distance between the outermost dendritic tips. The widest spread in the majority of granule cells is reported to be close to the transverse axis [42], so the elliptical cone transverse radius was set greater than the longitudinal radius in the generation process. All trees were resampled to a 5 µm fixed segment interval, and low-pass filtered homogenous spatial noise was applied to all points similar to previous methods [41], using length constants of 10 µm and 50 µm. Diameter mapping was implemented using a variable quadratic tapering from previous studies [41], [47] and adding an additional scaling function exp(x) – 1, where x is the distance from the soma, to better approximate the initial diameter taper close to the soma. Experimental reconstructions [42] used in the target point laminar distribution estimation (see Supporting Information) and synthetic tree validation were obtained from the www.NeuroMorpho.Org database [7].
The ray-tracing images in Figure 4B–C were created with the Persistence of Vision Ray tracer (POV-ray) software (http://www.povray.org/download/). The location of dendrites within each 25 µm cube was determined by testing the points in each dendritic tree, which specify the center of each segment. Because the granule cell dendritic trees were resampled at 5 µm before the spatial jitter addition, the length and volume measurements within each cubic volume are approximations. To get an estimate of the complexity required for an axon to contact all granule cell dendritic trees invading each cubic volume, random points were selected for each cubic volume and connected using the optimal wiring algorithm with a balancing factor of zero (to minimize total dendritic length). The number of target points was increased until the simulated axon reached within 5 µm of all granule cell dendritic trees present in the each volume. The results from 10 different random collections of target points were averaged together to determine the complexity required (number of branch points) for each cubic volume. In order to map the dendritic length and maximum tip distance onto the GCL, triangulations of the inner and outer GCL surfaces were created with 5000 faces, and the closest face for all trees was determined. The values for the trees associated with each respective face were then averaged together.
|
10.1371/journal.pbio.2001323 | What makes a reach movement effortful? Physical effort discounting supports common minimization principles in decision making and motor control | When deciding between alternative options, a rational agent chooses on the basis of the desirability of each outcome, including associated costs. As different options typically result in different actions, the effort associated with each action is an essential cost parameter. How do humans discount physical effort when deciding between movements? We used an action-selection task to characterize how subjective effort depends on the parameters of arm transport movements and controlled for potential confounding factors such as delay discounting and performance. First, by repeatedly asking subjects to choose between 2 arm movements of different amplitudes or durations, performed against different levels of force, we identified parameter combinations that subjects experienced as identical in effort (isoeffort curves). Movements with a long duration were judged more effortful than short-duration movements against the same force, while movement amplitudes did not influence effort. Biomechanics of the movements also affected effort, as movements towards the body midline were preferred to movements away from it. Second, by introducing movement repetitions, we further determined that the cost function for choosing between effortful movements had a quadratic relationship with force, while choices were made on the basis of the logarithm of these costs. Our results show that effort-based action selection during reaching cannot easily be explained by metabolic costs. Instead, force-loaded reaches, a widely occurring natural behavior, imposed an effort cost for decision making similar to cost functions in motor control. Our results thereby support the idea that motor control and economic choice are governed by partly overlapping optimization principles.
| Economic choice in humans and animals can be understood as a weighing of benefits (e.g., reward) against costs (e.g., effort, delay, risk), leading to a preference for the behavioral option with highest expected utility. The costs of the action associated with a choice can thereby affect its utility: for equivalent benefits, an action that requires less physical effort will be preferred to a more effortful one. Here, we characterized how human subjects assess physical effort when choosing between arm movements. We show that the effort cost of a movement increases with its duration and with the square of the force it is performed against but not with the distance covered. Therefore, the subjective cost that determines decisions does not reflect the objective energetic cost of the actions—i.e., the corresponding metabolic expenditure. Instead, the subjective cost has commonalities with the cost that our central nervous system is believed to minimize for controlling the motor execution of actions. Our findings thus argue in favor of action selection and action control sharing common underlying optimization principles.
| Should I rather bring the groceries from the car trunk to the kitchen in 1 trip or in 2 trips? Even in a seemingly simple decision like this, multiple decision parameters are at odds. When doing a single trip, this bothersome task will certainly be finished more quickly but will require an intense physical effort. This choice might also put one at risk to drop everything on the way. On the other hand, when making 2 trips, each will be less effortful and safer but the task will take longer to complete. When examined through the prism of economics, this example shows 2 alternatives with an equal reward but different amounts and types of costs: risk, time, and effort. Utility theory [1] posits that these decision parameters are combined in a single value, the utility, which characterizes the desirability of each choice as whole.
The ways in which costs affect the utility of an option have been well described for risk (prospect theory [2]) and delay (hyperbolic temporal discounting [3]). Defining such a relationship is not straightforward for effort. Physical effort [4], in contrast to mental effort [5], can at least be related to an external, physically measurable variable, in the same way that reward delay is used in the example of temporal discounting. Therefore, we focus on physical effort, defined here as the subjective cost or negative utility associated with a given motor action, independent from the costs resulting from its success rate (risk) or delayed reward (temporal reward discounting).
Studies focusing on the brain circuits involved in physical effort-based decision making in humans have used handle-squeezing tasks to produce different effort levels, but they just assumed that subjective effort increases monotonically with isometric squeezing force, without further characterizing the dependency [6,7]. Using a similar task, Hartmann and colleagues [8] showed a quadratic discounting of monetary rewards by squeezing force, suggesting that effort for isometric force production grows proportionally to the square of the force amount. In contrast, by pitting isometric force production with different parameters directly against each other, Körding and colleagues defined effort as a function of both the duration and magnitude of force production, without the need to use an external monetary scale [9]. By using 2 parameters in a force-production task, this latter study highlighted the multifaceted nature of physical effort. The use of isometric force-production tasks to probe physical effort discounting is, however, still limiting compared to the full range of effort-related parameters one could experience when deciding between actual movements. Here we characterize the influence of duration, distance, direction, and force on subjective effort costs in actual reaching movements.
From the perspective of motor control, planning and executing a movement, even towards an unambiguous goal, requires commitment to a specific motor act among an infinite amount of potential ways (“choices”) to acquire the goal. Decision making in this context can be seen as part of a continuum that includes motor planning and motor control, and minimizing various cost functions is a core concept of motor control: the stereotypical nature of movement trajectories and velocity profiles has been attributed to minimization of hand jerk [10], endpoint variance [11], and even control effort itself [12,13]. The potential tight link between decision making and motor control is supported by studies showing that action selection can take into account parameters that are related to movement execution, such as biomechanics [14,15] or motor accuracy [16]. Conversely, the vigor with which a movement is executed was shown to be explainable through delay discounting [17]. This raises the question of whether the subjective cost of an action as computed in a decision-making context (i.e., what we call effort here) is comparable to potential cost functions used for optimization in motor control or, as an alternative, to the metabolic cost of the movement.
Here we address this question by investigating how humans assess subjective physical effort in action-selection tasks involving binary choices between different reaching movements. In a first experiment, we varied movement duration, amplitude, and direction as well as resistive force in order to derive isoeffort curves in this duration–amplitude–direction–force space. This allowed us to independently test the sensitivity of subjects to impulse (force × duration) and work (force × amplitude) exerted during movements. In a second experiment, we pitted repeated identical movements against single movements with different resistive forces in order to obtain more precise estimates of the relationship between force and subjective effort in reaching movements.
In both experiments, subjects performed 2-alternative forced choice (2-AFC) tasks in which they compared 2 different actions and were asked to choose the least effortful one (Fig 1). To make informed choices in each trial, subjects first performed both proposed actions (sampling) and then reported the least effortful action by executing it again (choice). The need to repeat the chosen action rendered the choice relevant for the subjects, since genuine selection minimized the overall task effort. Both actions consisted of reach movements performed against different levels of resistive force. In each trial, one of the proposed actions served as a reference action, while the other served as a test action. Note that this distinction was not indicated and not relevant to the subjects but was part of our adaptive experimental design. Within each task condition, the trial-to-trial resistive force level in the test movements was selected with a staircase algorithm [18], while the force level of the reference movements was kept constant. As a consequence, the staircases converged to the force level at which the test action was perceived as being equally effortful as the reference action (equivalent force).
In experiment 1, reference and test actions consisted of single movements, differing primarily in amplitude or duration. More precisely, in each trial of the amplitude session, subjects had to choose between 2 movements that differed in amplitude, direction, and force (after sampling both). Conversely, in each trial of the duration session, the choice was between movements that differed in duration, direction, and force. In both sessions, the staircase algorithm adjusted the forces of 1 of the movements, depending on the choice of the subjects, until both movements were subjectively equivalent in effort for the subject. This allowed us to construct isoeffort curves in the force–amplitude–duration movement parameter space (Fig 1A–1D and Methods).
In experiment 2, the reference action consisted of 2 identical repeated movements and the test action of a single movement. This allowed us to determine the scaling of subjective effort with force (Fig 1E and 1F).
In experiment 1, we asked subjects to conduct naturalistic reach movements against different force levels, either with varied durations independent of amplitude (duration session) or with varied amplitudes independent of duration (amplitude session). As we used constant force profiles, these constraints correspond to dissociations either in impulse (force integrated over time) or work (force integrated over distance), respectively. Fig 2A and 2B depict the average work and impulse produced by the manipulator as a function of force for the different duration and amplitude conditions in both sessions for a representative subject (both integrated from the time of movement onset minus 100 ms to movement offset plus 400 ms). Impulse values were well separated in the duration session but not in the amplitude session, while work values were well separated in the amplitude session but not in the duration session. This confirms that visual instructions about reach-target location and requested movement duration together with the manipulator-controlled resistive force successfully constrained the actual movements of the subjects to the desired parameter ranges in each session (sample trajectories and generated force profiles in S1 Fig and S1 Text). Importantly, the forces imposed by the manipulator had an additive effect on the torques that the subject’s arm actually needed to produce to generate the movements. Since the imposed forces were independent from arm kinematics, a simple biomechanical model of the arm (S3 Text, S3 Fig, S4 Fig) showed that over the duration of a movement, the torques the subjects produced to compensate the imposed forces outweighed the torques produced to compensate for the inertia of the arm. As a consequence, the total work and impulse actually produced by the subjects in the different conditions showed dissociations comparable to those observed in Fig 2A and 2B.
Subjects’ choices did not systematically depend on performance differences in the various task conditions of experiment 1, but they depended reliably on force levels (S2 Fig, S2 Text). As a consequence, we could use equivalent forces to titrate the effort subjects associated with the explored movement parameters.
In the amplitude session of experiment 1, the equivalent forces did not vary significantly with movement amplitude (linear mixed-effect model [LME], p = 0.4, effect size 0.3 N) but varied significantly with reference force level (LME, p < 0.001, effect size 3.4 N) (Fig 2C). The result indicates that the 4-N difference between the reference forces was large enough for subjects to judge it as different in effort but that subjects were unaffected in choice by movement amplitude over the tested range when movement duration was kept constant. In other words, a movement of 120 mm against a force of 6 N was rated as effortful by the subjects as a movement of 200 mm against a force of 6 N. Since movement duration was constant in the amplitude session, the observed insensitivity to movement amplitude can also be interpreted as insensitivity to movement speed.
In contrast, subjects were sensitive to movement duration when movement amplitude was kept constant in the duration session (Fig 2D). Equivalent force levels were lower for long-duration movements than short-duration movements (LME, p < 0.001, effect size 3.8 N). Here, a movement in the 1,300–2,000 ms range against a force of 5 N was judged as effortful as a movement in the 0–800 ms range against a force of 9 N. This indicates that long-lasting reaches were perceived more effortful than brief reaches against the same force level. Additionally, the equivalent forces scaled with the reference force levels (p < 0.001, effect size 3.1 N) without interaction between the factors force and duration (p = 0.4).
Physical movement parameters like work or impulse, as defined above, describe the movement properties at the manipulator handle (endpoint movement), irrespective of the required multijoint arm movement. Instead, effort evaluation and the resulting choice were based on subjective experience, to which the biomechanics of the movement could have contributed. In fact, biomechanics of the movement influenced effort judgment in our experiment. In both the amplitude and the duration sessions of experiment 1, equivalent forces depended on the reference movement direction (Fig 2C and 2D). Test movements performed inward (i.e., towards the left for right-handed subjects and vice versa) required higher force levels to be judged as equally effortful as outward movements (LME, amplitude session: p < 0.001, effect size 1.5 N; duration session: p < 0.001, effect size 0.9 N). This indicates that at the same force level, outward movements were considered more effortful than inward movements. This difference in equivalent force is likely linked to the use of larger muscles and the higher available strength for inward movements.
After showing that duration and biomechanics but not amplitude had an influence on the effort judgment, we asked how effort would depend on the force itself. In Fig 2D, a thin dotted line represents an example isoeffort curve in the force–duration space: it connects the point representing parameters of a reference movement to the points representing parameters of equivalent-force test movements. Similar to this example curve, when averaging over movement directions in the duration session, isoeffort curves are convex for both reference force levels. This indicates that the putative effort cost function supporting the subjects’ choices was a nonlinear combination of force and movement duration. This is because if effort was a linear combination of force F and duration d, such as E(F,d) = αF + βd, the isoeffort curve defining the equivalent force of the test movement would have to be a straight line defined as FT=αFR+βdRα−βαdT. If, on the other hand, effort was a purely multiplicative function of force and duration—i.e., assuming that the effort cost function is impulse E(F,d) = Fd—then this would lead to convex isoeffort curves shaped as the inverse function FT=FRdRdT. The isoeffort curve in Fig 2D is not straight, and the curvature does not fit the impulse model but is shallower instead. In experiment 2, we determined the precise shape of the force–effort relationship for constant movement durations.
Binary choices between 2 options only allow ranking the options in terms of preferred or nonpreferred. To describe effort as a function of force, additional information is needed to turn such a ranking into a scale with a continuous metric. This could be achieved by trying to compensate the effort cost of an action with an independent scalable benefit (e.g., monetary reward) to achieve equal preference for movements of different force. But the utility function of the benefit must then be known, a task which might be as difficult to achieve as the task of determining the effort cost function itself. As an alternative, in each trial of experiment 2, subjects chose between 2 options after having sampled both: either they opted for performing 2 similar movements in rapid succession against a reference force level FR (the endpoint of the first movement is the starting point of the second movement), or they chose to perform a single movement against a test force level FT (adjusted between trials by a staircase). Here we assume that an action consisting of 2 identical movements is twice as effortful as an action consisting of 1 of these movements, but both give the same benefit (finishing the trial). Under this assumption, the equivalent forces FTeq (staircase convergence point for forces FT in single test movements) as a function of the reference movement force level FR (fixed forces in double movements) should follow the rule E(FTeq) = 2E(FR), that is FTeq = E−1(2E(FR)), with E(F) being the function linking force and subjective effort (= cost function) and E−1 being its inverse. As in experiment 1, the observed decision behavior in experiment 2 was best explained by force-based choices rather than performance-based choices (S2 Text). Results from experiment 2 thus allowed to test and fit models for both the cost function and its link to decisions, which we carried out using a Bayesian modeling approach (see Methods).
A direct observation of the equivalent force as a function of the reference force level is suited to highlight the properties required of E(F). We computed equivalent forces in 2 ways: first as averages of the test-force levels at the staircase inversions (i.e., the asymptotic force to which the staircase procedure converged as in experiment 1) and second via points of subjective equality of the psychometric functions that resulted from the Bayesian model. Results of both approaches are illustrated for 2 example subjects in Fig 3A and 3B. Equivalent to the considerations regarding the isoeffort curves in experiment 1, the simplest putative effort cost function is a linear function, E(F) = αF. For this, equivalent forces would have to obey the equation FTeq = 2FR (steepest dashed red line), as we required E(FTeq) = 2E(FR). However, in our data we observed that for large reference forces, the equivalent forces were smaller than predicted from the linear model. For example, for a FR of 9 N, subjects JP and MK showed equivalent forces of 12 and 14 N, respectively, instead of 18 N. This observation indicates that there was a convex nonlinear relationship between force and effort, in line with the results obtained from experiment 1. Additionally, we observed that the equivalent forces for a FR of 0 N were larger than 0 N, confirming the intuition that a movement against no external force still has nonzero effort (i.e., that E(0) > 0). Therefore, a power function with an offset, (F) = Fα + β, appears as a reasonable minimal model for the force–effort relationship (Eq 2; see Methods), which we will use in the following sections.
Our Bayesian modeling approach used the trial-by-trial choices of subjects as dependent data. Thereby, within the same unified framework, we simultaneously modeled (1) how forces affected effort values (see paragraph above, Eq 2) and (2) how the subjects’ choices depended on these effort values (Eq 1). For the dependence of choice on effort, we answered 2 questions. Does utility show a subtractive discounting by effort (i.e., the difference in utilities on which decisions are based is equivalent to a difference between efforts in our task [Eqs 4 and 5]) or a hyperbolic discounting by effort (i.e., the difference in utilities corresponds to a difference between effort inverses [Eqs 6 and 7]). Within each of these 2 alternatives, we tested whether effort was represented in a linear (Eqs 5 and 6) or logarithmic scale (Eqs 4 and 7). This 2 x 2 design resulted in 4 alternative models obeying E(FTeq) = 2E(FR), all with the same number of free parameters (see Methods). Notably, all tested models had the same equation for the equivalent force curve (Eq 3); the models only varied in the shapes of the choice curves—in particular, how the reference force FR affected their slope. We assessed model quality on the basis of 3 criteria. First, the percentage of correctly predicted choices (binary predictions based on a comparison between the actual test force and fitted equivalent-force levels) reflected the validity of the equivalent force curve as a decision boundary (in red on Fig 3A and 3B). Second, we evaluated the fit of the model to the full choice probability curves (in blue on Fig 3A and 3B) by examining the corresponding residual distribution (Fig 3F). Last, each fit was also tested by using the Watanabe–Akaike Information Criterion (WAIC) [19], an approximation of cross-validation that allowed us to compute the relative likelihood between models. The model we ultimately selected (and show in Fig 3A and 3B) expressed choice probability as a function of the difference between the logarithms of reference and test efforts (Eqs 1, 2 and 4) and outperformed all other models, which we discuss below.
The simplest model obeying E(FTeq) = 2E(FR) was based on a difference between test and reference efforts (Eq 5, no logarithms). The difference model predicted subjects’ choices as well (68.3% correct predictions) as the selected logarithmic difference model (68.7%), meaning that the equivalent force curves were similar. But the difference model did not fit the choice probability curves nearly as well (WAIC of 4.50e3) as the selected model (4.43e3), making the difference model less likely by a factor of 6.3e–16 compared to the selected logarithmic difference model. Indeed, the distribution of residuals for the simplest model (in red, Fig 3F), is wider than the distribution of residuals for the selected model (in grey, Fig 3F). Both alternative models based on hyperbolic discounting of effort, whether computing choice probability from the difference between the inverse of efforts (Eq 6) or the inverse of effort logarithms (Eq 7), showed lower prediction performances (66.9% and 63.8%), higher WAICs (4.76e3 and 5.09e3), and wider residual distributions than the selected model and therefore had to be rejected.
The results of this model-selection approach thus favor a subtractive discounting of utility by effort (no hyperbolic discounting) and a logarithmic internal representation of effort. By examining the posterior distributions of the selected model’s parameters (Fig 3C–3E, Eqs 1 and 2), we can interpret the equivalent force curves obtained in our subject population. Our main parameter of interest was the population average of the force exponent α in the power function described in Eq 2, as it describes the nonlinear dependency of effort on force. The exponent α showed a narrow posterior distribution centered around 2. This means that on average, the subjective effort rose with the square of the resistive force against a movement. Since the distribution is narrow, the confidence in this estimate is high (95% credible interval for the population average: 1.56–2.48). For completeness, the posterior for the effort offset β (95% CI: 4.9–22), reflecting the effort that subjects associated with performing a movement against no resistive force, and the posterior for γ (95% CI: 1.1–4.2), reflecting the sensitivity of subjects to effort differences, are represented in Fig 3D and 3E, respectively.
The biomechanics seemed to have played less of a role in experiment 2. For the same reference force levels, we did not obtain different equivalent test forces between the 2 movement directions inward and outward. However, in contrast to experiment 1, subjects never directly compared movements with different directions, as both options were in the same direction. This difference likely made experiment 2 much less sensitive in that respect.
In this study, we used an action-selection task to characterize how physical effort discounts the utility associated with arm movements and controlled for potential confounding factors such as delay discounting and performance. By repeatedly asking subjects to choose between 2 arm movements of different amplitude or duration that were performed against different levels of force, we were able to construct isoeffort curves in the amplitude–duration–force parameter space. These isoeffort curves indicated that for a choice between 2 arm movements against resistive forces, the movement amplitude did not influence effort cost but its force and duration did: movements with a longer duration were judged more effortful than shorter movements against the same force. Biomechanics of the movements also influenced their utility, as movements towards the midline of the body were less effortful than movements away from it. In a second experiment, by using the same approach but introducing movement repetitions as factor, we determined that the cost function in effort-based decisions had a quadratic relationship with force, and that choices were made on the basis of the logarithm of these cost functions.
Most studies on physical effort in human decision making operationalize effort by asking subjects to squeeze a handle that measures hand grip force [7,8,20,21], a device that is easy to use and fMRI-friendly. The effort is then an isometric contraction of varying magnitude, expressed as a percentage of the maximum voluntary contraction (%MVC) that each subject can produce. Subjects are typically required to choose between 2 squeezes with different grip forces, each associated with different rewards or additional decision factors (delay or risk). Hartmann and colleagues associated monetary rewards with grip forces and reported that among linear, hyperbolic, and quadratic effort cost functions, the quadratic cost function explained the subjects’ behavior best [8]. Klein-Flügge and colleagues used a similar task to compare effort discounting and delay discounting and reported that effort cost seemed best described as a sigmoidal function, i.e., showed a convex dependency for lower forces, as Hartmann and colleagues did, but Klein-Flügge and colleagues found a concave relationship for forces closer to MVC because of saturation of effort cost [20]. Burke and colleagues, instead, compared the integration of physical effort and risk in a similar task and reported a sharp increase of effort cost when approaching MVC [21]. In contrast, Prevost and colleagues used rewarding erotic images instead of money, but found effort cost to fit a hyperbolic function [7]. This means that for isometric force production, the effort cost function is still uncertain or at least depends on the choice task (reward, risk, or delay discounting), while for actual movements hardly any previous data exist.
To answer how effort depends on force in our movement task, we first have to address the question of how choice is best linked to effort, since choice is the behavioral readout, while effort is a hidden decision variable. Apart from the study by Prevost and colleagues, the aforementioned studies rejected the idea that choice behavior takes into account physical effort in a similar fashion to delay (i.e., by hyperbolic discounting). This is not surprising since, intuitively, high physical effort cannot make the subjective utility of a choice decrease asymptotically to zero (as hyperbolic discounting does because of the increasing denominator). Indeed, the subjective utility of a high-effort, low-reward action could well be negative, in which case doing nothing would be preferable. As a consequence, the cost of effort is a value that should be offset from the benefits of an action in the utility space—i.e., the utility of each action should be computed by subtracting the associated effort. This intuition is confirmed by the results of our experiment 2: models in which the decision variable was a difference between efforts predicted the subject’s movement choices better than models in which the decision variable was a difference between the inverse of efforts. Effort studies with isometric force production modeled the probability of choosing each option by transforming the difference of efforts between the 2 alternatives with the softmax function [7,20]. Here we used an equivalent probit transformation but showed that using the difference of the effort logarithms yielded significantly better results than using the difference of efforts. Indeed, the difference of logarithmic effort as choice variable best captured the decrease in sensitivity we observed for higher efforts (Fig 3A and 3B).
How does force affect effort in transport movements? Previous isometric contraction studies do not allow generating a good prediction for this question. The aforementioned studies modeled the subjective cost of effort as convex functions of force expressed as %MVC and titrated efforts against rewards. However, both these features of the experimental design could overemphasize the convexity of the effort cost function. First, producing a force stronger than MVC is by definition impossible. Therefore, the effort cost likely has to undergo a sharp increase when approaching this discontinuity in designs that use the full 0%–100% range of MVC as a force scale. Indeed, Hartmann and colleagues and Burke and colleagues noted that subjects always chose the effortful option when it provided more reward, except when the effort was close to MVC [8,21]. Second, the tendency to almost always make reward-based choices while ignoring moderate efforts also suggests that monetary rewards are too strongly motivating for typical subjects and may not be appropriate to study the cost of the moderate efforts. However, moderate efforts are essential in experimental settings to allow large numbers of trials and to stay away from the MVC discontinuity. In conclusion, a paradigm that uses moderate absolute forces instead of %MVC forces and in which effortful actions are not associated with monetary or social rewards but are compared directly seems more suited to precisely determine effort cost functions.
Such a paradigm, which we partly adopted here, was introduced for isometric forces by Körding and colleagues in a study in which subjects had to resist against imposed force profiles of variable magnitudes and durations [9]. This previous study led to a different effort cost function than the one we found. When subjects had to choose between dual and single contractions with the same force profiles, but in which force and duration were varied together, a loss function of the form (FT)α gave the best fit for = 1.1. Assuming that this fit can be generalized to constant durations, the resulting F1.1 relationship would indicate a quasilinear influence of isometric force on subjective effort. In contrast, in our experiment 2, in which we extended this approach to actual effortful transport movements and isolated force dependency by keeping duration constant, we found a more convex F2 relationship. Hence, our result is closer to the results obtained in studies using %MVC despite the use of a different task, force scale, and fitting procedure (the force exponent was a free parameter in our model, in contrast with [8]). Nevertheless, our use of moderate forces prevents us from generalizing our findings to movements realized against higher levels of forces (closer to MVC), for which large accuracy and duration changes might bias choice preferences independent of force-dependent effort.
In summary, the cost of effort as a function of isometric muscle contraction force has previously been shown to take various forms. Yet, we mainly attribute differences to the quadratic discounting that we observed here to 2 facts that were not fully considered previously. First, compensating effort with rewards is difficult because of uncertainties about the reward utility itself and the need to use large forces. Second, even when avoiding reward–effort competition and comparing effortful actions directly, interactions between force and duration need to be avoided or compensated for, since duration contributes to both effort and reward discounting and thereby may distort measured force–effort relationships, as will be discussed in the following paragraph.
To properly understand physical effort in movements, it is important to disentangle the different contributions of force and duration. A previous study obtained V-shaped isoeffort curves in the duration–force space when pitting isometric contractions with force profiles of different durations and magnitudes against each other [9]. For durations below 250 ms, an increase of duration required a sharp decrease of force to maintain effort constant; for durations above 500 ms, the opposite was observed. In other words, contractions of long durations (1–2 s) were considered as effortful as shorter contractions, even with slightly lower forces. This observation contradicts the intuition that longer contractions should be more effortful than shorter contractions. Moreover, performing short contractions allowed finishing the experiment more quickly since total trial duration was not controlled for; therefore, delay discounting should have additionally devalued longer movements. Körding and colleagues interpreted their result as a consequence of increased control difficulty for fast force changes: it was easier for subjects to resist against stronger force profiles when the onset and offset of the forces were slower, which was the case for long-duration force profiles. In this sense, their results marked a compound effect. In contrast, in our task we tied the onset and offset of forces to self-timed movements. This rendered force control less difficult, and, as a consequence, we observed monotonically decreasing isoeffort curves in the duration-force space (Fig 2D)—i.e., effort increased monotonically with both duration and force, as intuitively expected.
Other recent work argued that increasing movement durations requires smaller and smaller decreases of force to maintain effort constant (i.e., that perceived effort reaches an asymptote instead of growing linearly with duration in isometric force productions). To explain such a counterintuitive effect, the authors assume that effort costs are subject to the same temporal discounting as can typically be observed for reward in economic choice behavior [22]. This hypothetical explanation is, however, not applicable to our results, as we kept the total duration of trials constant. The isoeffort curves observed in our experiment 1 can be explained by the quadratic relationship between force and effort.
In conclusion, for effortful transport movements like reaches, effort increases monotonically with movement duration, suggesting that effort is integrated over time.
The observations from experiment 1 and 2 provide insight into the internal cost function used by subjects to decide between arm movements. This effort cost function for decisions could potentially be paralleled with motor control cost functions or with the actual metabolic cost of movements. In other words, we can examine whether the choice made by subjects between proposed movements with constrained parameters (duration, speed, force, etc.) reflect natural preferences in the execution of unconstrained movements or minimization of energy expenditure.
After providing informed written consent, 17 subjects participated in experiment 1, and 16 other subjects participated in experiment 2 (ages 19–30 years, 6 left-handed subjects, normal or corrected-to-normal vision, no overlap in subjects between the experiments). Experiments were in accordance with institutional guidelines for experiments with humans, adhered to the principles of the Declaration of Helsinki, and were approved by the ethics committee of the Georg-Elias-Mueller-Institute for Psychology at the University of Goettingen. In both experiments, subjects answered a postexperiment questionnaire.
Subjects performed the tasks by holding and moving the spherical handle of a parallel-type haptic manipulator (Delta.3, Force Dimension, Nyon, Switzerland) with their dominant arm (Fig 1A). The manipulator was connected to a computer running our own custom-written software (C++, OpenGL) in charge of visual stimulus presentation, task event control, force computation, and associated data recording. The manipulator and the computer communicated bidirectionally at 2 kHz, with the manipulator sending the 3D position of the handle and the computer requesting forces to be applied at the handle for each iteration of this 0.5-ms haptic cycle.
The movements of the manipulator handle were reproduced in real time for the subject via a spherical yellow cursor displayed in a stereoscopic augmented-reality (3D-AR) environment. Display and haptic device latency were fully compensated by a forward prediction to achieve synchrony between visual cursor and handle movement (Kalman filter with position, speed, and acceleration as state variables). The 3D-AR environment consisted of 2 computer monitors (BenQ XL2720T, screen size 590 x 338 mm, 60-Hz refresh rate, distance 45 cm, Matrox DualHead2Go DisplayPort splitter) that were placed to either side of the subject with the screens facing each other. The subject viewed the screens through a pair of semitransparent mirrors that were angled at 45° relative to the screens. This allowed for the creation of stereoscopic 3D visual stimuli that were perceived as being projected into the haptic device’s workspace. In addition to the visual cursor, which always coincided with the handle's current physical position, other visual stimuli indicated the starting points and targets of the reaching movements as well as text information.
The 3D-AR haptic interface was calibrated for each subject. For this, we made the actual manipulator handle visually coincide with multiple visual targets sequentially presented in the virtual space. Since the control software selected the visual target locations, the manually adjusted handle position could be used to compute the manipulator-to-display transformation matrix for the current geometry of the setup. This calibration was then further adjusted for each subject by setting the location and projection matrix of the virtual openGL cameras according to the subject’s interpupillary distance.
To allow the subjects to comfortably operate the haptic manipulator, both monitors and mirrors were tilted to lower the location of the 3D representation (Fig 1A; angle relative to horizontal: 30°). For the same reason, we defined a virtual plane in front of and parallel to the monitor image plane in which all movement targets appeared (distance to mirrors was 430 mm). Subjects were also encouraged to take breaks and relax their arm as frequently as desired. In order to limit the force output of the manipulator to task-relevant forces only, and to allow for natural movement trajectories, subjects could freely move the cursor around the entire spherical workspace of the haptic device. Correct depth perception of the 3D stimuli was thus required for the subjects to be able to acquire the movement targets.
Before each experimental session, subjects were trained on simple versions of the tasks in order to familiarize themselves with the setup and the task requirements, notably 3D vision, resistive forces, and time constraints.
The haptic manipulator produced forces that resisted the subjects' movements. Our aim was to produce a force with constant magnitude that was only present during the movements and that opposed the instantaneous movement direction, similar to a kinetic friction force. The direct definition of this friction force would thus depend on the velocity of the handle. Yet, the force command sent to the manipulator was not computed directly from online estimates of handle velocity (which is difficult at low speeds) to prevent sudden force onset and direction inaccuracies at low handle speeds. Instead, we implemented the friction force using a virtual point-mass (virtual mass = 100 g) that was connected to the handle via a virtual spring (coefficient = 1 N.m-1). In other words, subjects dragged a virtual mass with the help of a spring, and the kinetic friction force was computed according to the speed of the virtual mass and applied to it. The magnitude of this kinetic friction force (in N) was varied in order to produce the different resistive force levels. For each iteration of the haptic cycle, the dynamic state of this virtual mass was updated according to the forces applied to it (= sum of the spring force and the friction force), and the force resulting from the virtual spring was sent to the haptic manipulator as a command. The position of the virtual point-mass was reset to the handle location, and thereby the spring force set to 0, before the start of each movement. Additionally, the commands sent to the haptic manipulator were modulated by an envelope function (a constant function with linear tapers of varied durations at onset and offsets), which allowed controlling force output outside of the defined movement periods.
The course of events of a trial in experiment 1 is presented in Fig 1B (example trial from the amplitude session). Each sampling subtrial started with the subject placing the cursor (6-mm diameter, yellow sphere) within a fixation sphere (20-mm diameter, grey, brightening upon acquisition) and holding this position for a duration randomized between 500 ms and 800 ms. The movement target sphere (diameter 30 mm) was displayed from the start of the subtrial; its color indicated to the subject the modalities of the movement to be executed: across both sessions, a green target indicated the need for a rapid movement (a short-duration movement in the duration session or a large-amplitude movement in the amplitude session), whereas blue and red indicated medium and low speed, respectively. During the acquisition and hold stages, onscreen text announced which sampling movement was currently being performed (“sampling 1” or “sampling 2”). When the fixation sphere disappeared (“go” cue), the subject had to execute the required movement by placing the cursor within the target sphere within the requested time constraints. Movement duration was computed from movement onset (determined online by a combination of speed and distance thresholds) to target acquisition (determined only based on cursor position relative to the target). Resistive forces were turned on when the fixation sphere was acquired (on-taper: linear increase to the desired force value within 200 ms) and were turned off when the target was acquired (off-taper: linear decrease to 0 within 600 or 500 ms in case of successful or failed acquisition). Note that the actual force production by the manipulator was dependent on the subject’s movement and only started at movement onset (see "Force generation" section). If the subject acquired the target faster than the minimum duration set for the movement, or if the subject did not reach the target before the maximum duration set for the movement, the subtrial was interrupted and onscreen text indicated to the subject the type of error committed (“too fast” or “too slow”). Failed subtrials, which also included trials in which the subject broke fixation or left the target too early, were restarted until executed correctly. Once the target was acquired (with movement duration dm), which was signaled by the target sphere becoming brighter, the subject had to hold the cursor within the sphere for a total duration (in ms) dh = 100 + dmmax − dm, with dmmax being the maximum potential movement duration in the session (2,000 ms in the duration session, 1,250 ms in the amplitude session). This ensured that every subtrial had the same duration across conditions within a session, thus preventing temporal discounting, here equivalent to the desire to terminate the experiment early by preferably selecting short movements.
After sampling both the test movement and the reference movement by performing each sampling subtrial successfully, the subject had to indicate in the choice subtrial which movement felt less effortful. This subtrial, announced by a “choice” onscreen text, started with the subject acquiring a pre-fixation sphere. Then, the fixation spheres and targets for the 2 alternatives were displayed, and the subject indicated their choice by acquiring the fixation sphere of the chosen movement (Fig 1B, right column). With acquisition of the chosen movement’s fixation sphere, the fixation and target spheres of the nonchosen movement disappeared, and the rest of the subtrial was identical to the sampling subtrial that corresponded to the chosen movement.
Subjects performed experiment 1 over 2 sessions on different days, with each session lasting on average 70 minutes. In the duration session, reference and test movements differed in allowed movement duration but had the same amplitude. This allowed us to construct isoeffort curves in the force–duration space. Conversely, in the amplitude session, the reference and test movements differed in amplitude, but not in duration, which allowed us to construct isoeffort curves in the force–amplitude space. With the use of constant magnitude force profiles, the duration and amplitude session allowed us to double-dissociate total impulse (Jx=∫tstarttstopFxdt) and work (Wx=∫tstarttstopFxdxdtdt), respectively. The reference movements had a medium duration in the duration session (800–1,300 ms) and medium amplitude in the amplitude session (160 mm). In both sessions, reference movements were performed against either 6 N or 10 N of resistive force and could be directed either to the right or to the left. The test movements required either low or high values for the session variable of interest (low or high duration or amplitude) and were carried out in the direction opposite to the reference movements (Fig 1D). These combinations lead to 8 conditions per session, which were presented to the subject in a randomly interleaved manner (2 reference movement force levels × 2 reference movement directions × 2 test movement levels of duration or amplitude).
For each of these 8 conditions, independent pairs of staircases determined the force level against which the test movements were performed. These staircase pairs followed a one-up one-down rule with a step size of 2 N, with one staircase starting at 0 N and the other at 16 N (the highest force the manipulator could sustainably produce) to compensate hysteresis. In other words, when the subject chose the reference movement in a given trial, the force level of the test movement in the next trial of the same condition and staircase would be decremented by 2 N; and vice-versa, when the subject chose the test movement, the force level of the next test movement of the same condition and staircase was incremented by 2 N. Data collection for each staircase was considered complete after 7 inversions in the subject’s choices. In this subjective-choice task, the choices of the subject could sometimes lead the staircase procedure to propose force values beyond the capabilities of the manipulator (above 16 N) or the interest of the task (below 0 N). As a consequence, the force values were clamped between 0 and 16 N, and each time the force stayed at these boundaries for 2 trials in a row, an inversion was counted in order to allow the staircase to terminate eventually. We used the average force at staircase inversions to determine isoeffort forces in experiment 1. Because of the clamping of the staircase forces, the isoeffort forces were also bounded between 0 and 16 N, which could have caused an underestimation of the observed effects in the rare cases in which the subject stayed at the clamped force limits.
Target and fixation locations were selected to avoid confounding biases. Across all 8 conditions, the targets were not placed further than 100 mm from the workspace vertical midline to prevent the effort of reaching towards large eccentricities, which would be considered a confounding factor. To achieve this, the different movement amplitudes in the amplitude session were created by offsetting the locations of the fixation spheres while keeping the targets at constant eccentricities (Fig 1C, vertical dotted lines). For this reason, the 2 alternative movements (towards left and right) also had to be placed at different heights on the workspace (40-mm vertical distance). This made them visually more distinguishable for the subjects, especially in the choice subtrial, in which both targets and both corresponding fixation spheres are displayed. Importantly, the prefixation sphere in the choice subtrial was placed halfway between the 2 alternative fixation spheres to prevent subjects from choosing the movement starting closest to the current cursor location. Both fixation spheres were visually identical and were identified by their vertical position, which was the same as the target of the corresponding movement.
While experiment 1 was designed to explore isoeffort curves in the force–duration–amplitude spaces, the similar experiment 2 was designed to provide more details about the shape of the force–effort relationship. Instead of executing a single movement in the reference action, subjects performed an identical movement twice in experiment 2, while the amplitudes and durations of all movements were kept the same across conditions. Assuming that executing a movement twice doubles the associated effort, the experiment allowed us to determine how much force in a test movement was needed to double the effort of a single movement from the reference action. Subjects performed experiment 2 in a single session (average duration 140 minutes).
To repeat the reference movement, 2 targets were presented in the corresponding subtrial, and subjects performed 2 reaches in succession. The location of the first target was used as a starting point for the second movement such that no additional movements were required (additional movements would cause more than doubling of effort). Targets were placed such that the 2 movements matched in reach direction and amplitude (Fig 1E). In a reference action, after the movement to the first target, the subject had to maintain the cursor in its location for 500 ms, after which the first target disappeared, indicating to the subject to perform the movement to the second target. The resistive force was tapered in and out for each of these movements (50-ms onset taper on “go” cue and 400-ms offset taper on target acquisition). In all other aspects, the course of events for experiment 2 is identical to experiment 1.
In experiment 2, individual movements had a 120-mm center-to-center amplitude and were time constrained between 800 and 1,300 ms (matching the short-amplitude, mid-duration reaches of experiment 1, Fig 1F), for both individual reference movements and the test movements. The total duration of each action, starting from the time the subject left the fixation point to the end of the subtrial, was maintained constant over all subtrials by adding an additional waiting time when holding the target, resulting in a total subtrial duration of 4,000 ms. Contrary to experiment 1, both reference movements and the test movement were in the same direction in each trial. Four levels of reference movement force were probed (0, 3, 6, and 9 N), while the forces for the test movements were determined using the same staircase procedure as in experiment 1. Therefore, there was 1 staircase pair for each of the 8 conditions (4 reference movement force levels times 2 movement directions).
Data processing and statistical analysis were carried out using Matlab and the gramm [34] toolbox for plotting.
In experiment 1, averaged staircase inversion points were analyzed using LMEs (fitlme function in Matlab). For each session, we constructed mixed-effect models fitting the average force at staircase inversion points depending on the varied parameter of the session (duration or amplitude: low, high), reference movement direction (relative to subject handedness: inward, outward), and reference movement force (low, high). All these independent variables were treated as categorical variables. The mixed-effect model included separate random intercepts and random slopes for movement duration and for amplitude across subjects. Main effect sizes were extracted from models without interaction terms. Interactions were tested in separate models and their significance was evaluated by model comparison.
Choice data from experiment 2 were modeled using Stan [35], a probabilistic programming language, through its Matlab interface. We used Stan to perform Bayesian inference, using its default implementation of a Markov chain Monte Carlo sampler (NUTS). We fitted a probit hierarchical model, in which the choice of the reference movement in each trial is modeled as a Bernoulli distribution in which the associated probability P(R|FT,FR) is a function of the difference in utility between the test and the reference movement, and the utility for each movement depends on the corresponding movement force. Variations of the model will differ in the way utility is expressed as a function of force-dependent effort.
Therefore, for subject i:
P(R|FT,FR)i=Φ(U(2Ei(FR))–U(Ei(FT))γi)
(1)
where Φ is the cumulative density function (CDF) of the standard normal distribution (probit link), Ei(F) is the effort of a movement executed against the force F for subject i, and U is the utility as a function of force-dependent effort. The factor 2 reflects our assumption that repeating a movement twice should double the effort compared to a single movement and thus imposes the constraint E(FTeq) = 2E(FR). Effort itself in all variations of the model was modeled as power-law function of force:
Ei(F)=Fαi+βi
(2)
The constraint E(FTeq) = 2E(FR), applied on Eq 2, yields the following equation for the equivalent force curve:
FTeq=2FRαi+βiαi
(3)
The force exponent αi, the effort offset βi, and the effort sensitivity γi for each subject were drawn from normal distributions αi ∼ N(μα,σα), βi ∼ N(μβ,σβ), γi ∼ N(μγ,σγ). The resulting parameters of these normal distributions characterize the population-level distributions for α, β, and γ. Bayesian inference requires providing prior distributions for these parameters, which were chosen wide to not constrain the model: μα ∼ N(1,10), μβ ∼ N(0,100). The parameters μγ,σα,σβ,σγ were positive scale parameters and their priors each followed the same half-Cauchy distribution [36] with parameters (location = 0; scale = 20). Posterior distributions were sampled using 4 Markov chains with 1,000 samples each (after a warmup of 1,000 samples).
To test our model against alternative hypotheses, we varied the function of the choice probability (Eq 1). We then compared the individual model fits using the WAIC [19], an approximation of cross-validation.
In our first model, utility is the negative logarithm of effort and choice probability thus depends on the difference between effort logarithms:
U(E)=−logE
(4)
As second model, we used a simpler model for the choice probability that is based on the difference of effort values and not the log-ratio (difference of their logarithms):
U(E)=−E
(5)
Third, we tested the hypothesis of hyperbolic effort discounting with a model in which effort is on the denominator in each utility term (inverse effort):
U(E)=1E
(6)
Finally, in a fourth model we modified the hyperbolic model to use logarithmic effort:
U(E)=1logE
(7)
|
10.1371/journal.pgen.1006698 | VAMP3/Syb and YKT6 are required for the fusion of constitutive secretory carriers with the plasma membrane | The cellular machinery required for the fusion of constitutive secretory vesicles with the plasma membrane in metazoans remains poorly defined. To address this problem we have developed a powerful, quantitative assay for measuring secretion and used it in combination with combinatorial gene depletion studies in Drosophila cells. This has allowed us to identify at least three SNARE complexes mediating Golgi to PM transport (STX1, SNAP24/29 and Syb; STX1, SNAP24/29 and YKT6; STX4, SNAP24 and Syb). RNAi mediated depletion of YKT6 and VAMP3 in mammalian cells also blocks constitutive secretion suggesting that YKT6 has an evolutionarily conserved role in this process. The unexpected role of YKT6 in plasma membrane fusion may in part explain why RNAi and gene disruption studies have failed to produce the expected phenotypes in higher eukaryotes.
| The constitutive secretory pathway delivers newly synthesised proteins and lipids to the cell surface and is essential for cell growth and viability. This pathway is required for the secretion of molecules such as antibodies, cytokines and extracellular matrix components so has both significant physiological and commercial importance. The majority of secreted proteins begin their journey at the endoplasmic reticulum, pass through the Golgi, and are transported to the cell surface in small vesicles/tubules which fuse with the plasma membrane. Surprisingly, the molecular understanding of this fusion step is still unclear and in higher eukaryotes it is not known which SNARE proteins drive this process. To address this problem we have developed a powerful, quantitative assay for measuring secretion and used it in combination with gene depletion studies in Drosophila cells. Using this assay we identified three SNARE complexes driving the fusion of secretory vesicles with the plasma membrane and uncovered an unexpected role for the R-SNARE YKT6 in this process. Using this knowledge we have re-examined the role of SNAREs in the fusion of secretory carriers with the plasma membrane in mammalian cells and have found that YKT6 has an evolutionarily conserved role in this process.
| Constitutive secretion delivers newly synthesised proteins and lipids to the cell surface and is essential for cell growth and viability. This pathway is required for the exocytosis of molecules such as antibodies, cytokines and extracellular matrix components so has both significant physiological and commercial importance. The majority of constitutive secreted proteins are synthesised at the endoplasmic reticulum, pass through the Golgi, and are transported to the cell surface in small vesicles and tubules which fuse with the plasma membrane [1, 2]. Constitutive secretory vesicles are not stored within the cell and do not require a signal to trigger their fusion with the plasma membrane which is in contrast to dense core secretory granules or synaptic vesicles [3, 4]. In some cell types, such as MDCK cells and macrophages, there is evidence that constitutive secretory cargo passes through a endosomal intermediate on its way to the cell surface [5, 6]. However, in non-polarised cells endosomal intermediates do not appear to play a major role in this pathway [7].
Vesicle fusion is driven by a family of molecules known as SNAREs. SNARE are generally small (14-42kDa), C-terminally anchored proteins that have a highly conserved region termed the SNARE motif that has the ability to interact with other SNAREs [8, 9]. For membrane fusion to occur, SNAREs on opposing membranes must come together and their SNARE motifs zipper up to form a SNARE complex [10, 11]. Detailed characterisation of the neuronal SNARE complex (syntaxin 1A/VAMP2/SNAP25) required for synaptic vesicle fusion has provided a mechanistic framework for understanding the function of SNAREs [4, 12, 13]. There are 38 SNAREs encoded in the human genome and they can be classified as either R or Q-SNAREs depending on the presence of a conserved arginine or glutamine in their SNARE motif [14–16]. Q-SNAREs can be further subdivided into Qa, Qb and Qc SNAREs based on their homology to syntaxin and SNAP25. A typical fusogenic SNARE complex will contain four SNARE motifs (Qa, Qb, Qc and R)[17]. Qbc-SNAREs such as SNAP23, 25, 29 and 47 contribute two SNARE motifs to the SNARE complex. R-SNAREs can also be further classified as either longin or brevin type SNAREs. Longin type R-SNAREs contain a longin type fold and are found in all eukaryotes and while brevin type SNAREs are less widely conserved across species [18].
Over the past twenty years significant progress has been made defining the SNARE complexes required for the majority of intracellular transport steps within eukaryotic cells (reviewed in [19–23]). In addition, there are an increasing number of examples where the SNARE complexes required for the secretion of specific cargo such as Wnt, TNF and IL-6 have been identified [24–26]. However, these proteins are not delivered directly to the cell surface from the TGN but pass through an endosomal compartment. Many labs, including our own, have attempted to identify the machinery which drive the fusion of constitutive secretory vesicles with the plasma membrane and on the whole very little progress has been made [27–34]. This in part may be due to the fact that there are multiple routes to the cell surface from the Golgi and redundancy in the fusion machinery. If we just consider the R-SNAREs, the human genome encodes seven post-Golgi SNAREs (Table 1) and a typical mammalian cell line can express at least five R-SNAREs so disruption of just one R-SNARE is unlikely to block secretion if they are functionally redundant [15, 27]. To overcome this problem we have decided to analyse SNARE function in Drosophila cells as they have a simpler genome with less redundancy. The Drosophila genome encodes 26 SNAREs with 16 of them predicted to be localised to post-Golgi membranes based on their homology to mammalian SNAREs [14, 15]. The complexity is reduced even further as Drosophila cell lines just express two post-Golgi R-SNAREs, Syb and VAMP7 (based on publically available microarray data generated by the modENCODE project)[35].
In this study, we have developed a novel, quantitative assay for measuring constitutive secretion based on a reporter cell line that can be effectively used to monitor secretion by flow cytometry, immunoblotting and fluorescence microscopy. Depletion of known components of the secretory pathway in Drosophila cells (STX5, SLH and ROP) causes robust blocks in ER to Golgi and Golgi to plasma membrane transport, therefore validating this approach. As predicted, there is redundancy in the post-Golgi SNAREs and multiple SNAREs must be depleted to obtain robust blocks in secretion. We have detected strong negative genetic interactions between Drosophila STX1 and STX4, SNAP24 and SNAP29, STX1 and Syb, and SNAP24 and Syb. We have also detected a novel and unexpected genetic interaction between Syb and YKT6. Depletion of YKT6 and VAMP3 in mammalian cells also causes a robust block in secretion indicating that this negative genetic interaction is conserved across species and provides evidence that these two R-SNAREs function in the late secretory pathway.
We previously used a ligand-inducible reporter system to measure constitutive secretion in mammalian cells [27, 36]. This system utilizes a GFP-tagged reporter construct (cargo) that is retained in the ER until the addition of a small molecule (AP21998 or D/D solubiliser), which causes the cargo to exit the ER in a synchronous pulse (Fig 1A). The transport of the cargo can be monitored using flow cytometry, microscopy and immunoblotting. The cargo contains a furin cleavage site so changes in its molecular weight can be used to determine if it has reached the trans-Golgi network (TGN), where the furin endoprotease normally resides. We have moved this reporter system into Drosophila S2 cells and generated a clonal cell line (C3). C3 cells have similar secretion kinetics to mammalian cells and secrete approximately 80% of their cargo in 80 minutes (Fig 1B) [27].
To validate the C3 cells we used RNAi to deplete the Drosophila orthologues of syntaxin 5 (STX5) and Sly1 (SLH), genes previously shown to be essential for ER-Golgi transport in human cells [27, 37, 38]. Amplicons to both of these genes were designed using FLYBASE, synthesised and transfected in to C3 cells. The mRNA level for both STX5 and SLH were reduced by over 80% as determined by qRT-PCR (S1 Fig). Depletion of STX5 and SLH cause a significant block in biosynthetic transport as almost no cargo is secreted from the cells as determined by flow cytometry (Fig 1C and 1D) and immunoblotting (Fig 1E). Similar results were obtained using alternate amplicons indicating that the observed block in secretion is not due to off-target effects (S1 Table). In the STX5 depleted cells the trapped cargo is found in the Golgi (co-localisation with Golgi marker GM130) and reticular and tubular structures most likely the ER (Fig 1F and S1 Fig).
To determine whether the assay could be used to detect blocks in post-Golgi trafficking we depleted ROP, the Drosophila Sec1 homolog [39, 40] and STX7/Avalanche an endosomal Q-SNARE. Immunoblotting for ROP and STX7 confirmed that both proteins were efficiently depleted (Fig 1E). Depletion of ROP caused a significant defect in secretion, while depletion of the endocytic SNARE STX7 did not (Fig 1C–1E). An alternate ROP amplicon give a similar phenotype indicating that the defect in secretion is not due to off-target effects (S1 Table). In the ROP depleted cells, a significant proportion of the retained cargo has been furin-processed suggesting that it has reached a post-Golgi compartment (Fig 1E, appearance of lower molecular weight band in GFP blot and accumulation of processed GH in the cells). In support of the biochemical data we observe cargo accumulating in small vesicular structures in the ROP depleted cells (Fig 1F). These membranes are distinct from the Golgi (GM130 negative) and start appearing approximately 20 minutes after the induction of secretion suggesting that they may be post-Golgi transport carriers which have been unable to fuse with the plasma membrane.
To determine which post-Golgi SNAREs mediate fusion of biosynthetic vesicles with the plasma membrane we depleted syntaxin 1 (STX1), syntaxin 4 (STX4) and synaptobrevin (Syb). These SNAREs are the closest homologs of the yeast genes SSO1/2 and SNC1/2 previously shown to mediate the fusion of biosynthetic vesicles with the plasma membrane [41, 42]. We depleted these SNAREs individually, or in combination and the knock down efficiency was determined by immunoblotting and RT-PCR (Fig 2B and S2 Fig). Depletion of STX1 or Syb leads to a partial block in secretion while depletion of STX4 had no effect (Fig 2A and 2C)(S2 Fig). Depletion of STX1 or Syb leads to a similar phenotype to that observed with ROP knock down, where furin-processed cargo is retained inside the cell (Fig 2B). The block in secretion became more pronounced when STX1 and STX4, or STX1 and Syb were depleted in combination indicating a negative genetic interaction between these genes. In the STX1-STX4 depleted cells the retained cargo is found in small vesicular structures scattered throughout the cytoplasm (Fig 2D). No genetic interaction was detected between STX4 and Syb. The STX1-Syb genetic interaction can be replicated using an alternative amplicons targeting Syb. Alternative amplicons targeting STX1 did not efficiently knockdown the protein (S1 Table).
As depletion of Syb did not produce a complete block in secretion it was possible that another R-SNARE might be able to substitute for the loss of Syb. To address this we used publically available microarray data to determine which Drosophila R-SNAREs are expressed in S2 cells (modENCODE project). The R-SNAREs Syb, VAMP7, YKT6 and Sec22b are expressed in S2 cells, but not the neuronal R-SNARE n-Syb. This is consistent with previous studies indicating that n-Syb is exclusively expressed in neuronal tissue [43].
We depleted the R-SNAREs individually or in combination and determined the knock down efficiency by immunoblotting (Fig 3B). Depletion of Syb, YKT6 and Sec22b caused a partial block in secretion as determined by flow cytometry and immunoblotting of the cargo (Fig 3A–3C) (S3 Fig). Depletion of Syb or YKT6 causes retention of furin-processed cargo indicating a late block in secretion (Fig 3B, GFP and GH blots). This block became more severe when YKT6 and Syb were depleted in combination. The level of block was comparable to that observed when STX5 is depleted as almost no processed GH was detected in the media (Fig 3B, GH media blot). In support of the role of Syb and YKT6 in the fusion of secretory carriers with the plasma membrane we observe an accumulation of secretory carriers in cells depleted for both of these genes (Fig 3D). The observed genetic interaction between Syb and YKT6 are not due to off-target effects as they can be reproduced using alternate amplicons targeting both genes (S1 Table). Importantly, no genetic interaction was detected between the R-SNARE Sec22b and Syb indicating that YKT6-Syb interaction is specific and not due to general toxicity (Fig 3A–3C). In support of YKT6 having a role in the fusion of secretory carriers with the plasma membrane we were able to immuoprecipitate YKT6 in a complex with STX1 from S2 cells (Fig 3E) (Table 2).
In S. cerevisiae, it has previously been reported that YKT6 and Sec22 function redundantly in ER to Golgi transport [44]. To determine if this is also the case in Drosophila cells we depleted YKT6 and Sec22b individually and in combination (Fig 4B). As in S. cerevisiae, we see a robust block in constitutive secretion when YKT6 and Sec22b are depleted in combination (Fig 4A and 4C)(S4 Fig). The level of inhibition is very similar to that seen when STX5 is depleted. In the Sec22b/YKT6 depleted cells the cargo is trapped in the ER and has failed to reach the Golgi (Fig 4B and 4D). This is in contrast to what is observed when YKT6/Syb are depleted where there is an accumulation of furin processed cargo (Fig 4B). We also depleted YKT6 in combination with STX1 and STX4. No genetic interaction was detected between YKT6 and STX1 or YKT6 and STX4 (Fig 4A and 4C).
Our data suggests that the Qa-SNAREs STX1/4 and the R-SNAREs Syb and YKT6 mediate the fusion of secretory carriers with the plasma membrane. A canonical SNARE complex also requires Qb and Qc SNARE domains, often provided by a Qbc-SNARE. The Drosophila genome encodes three SNAP genes: SNAP24, 25 and 29 (ubisnap) [15]. Only SNAP24 and SNAP29 are expressed in S2 cells based on publically available microarray data (modENCODE project). We depleted SNAP24 and SNAP29 individually or in combination and validated the knock down for SNAP29 using immunoblotting (Fig 5B). Depletion of SNAP24 or SNAP29 did not block secretion of the reporter construct. However, depletion of SNAP24 and SNAP29 in combination caused a significant block in secretion (Fig 5A–5C) (S5 Fig), similar to that seen when ROP is depleted. Similar results were obtained using alternate amplicons indicating that the observed block in secretion is not due to off-target effects (S1 Table). In the SNAP24-SNAP29 depleted cells a significant amount of furin-processed cargo is retained within the cells suggesting a late block in secretion (Fig 5B, GH and GFP blots). In support of the biochemical data we observe an accumulation of secretory carriers is the cells depleted for both SNAP24 and SNAP29 (Fig 5D). Consistent with SNAP24 having a role in the fusion of secretory vesicles with the plasma membrane we were able to immunoprecipitate SNAP24 in a complex with STX1 (Fig 3E)(Table 2).
We also investigated the effect of depleting SNAP24 and SNAP29 in combination with Syb. Depletion of SNAP24 and Syb in combination caused a robust block in secretion (Fig 5A–5C). The retained cargo was furin-processed indicating a late block in secretion (Fig 5B, GH and GFP blots). As in the SNAP24-SNAP29 knock down the cargo accumulated in small transport vesicles which did not co-localise with the Golgi (Fig 5D). No genetic interaction was detected between Syb and SNAP29. Depletion of SNAP29 in combination with SNAP24 and Syb did not cause a stronger block in secretion. In support of these observations we obtained similar results using alternate amplicons (S1 Table).
We have uncovered an unexpected role for YKT6 in the fusion of biosynthetic vesicles with the plasma membrane in Drosophila cells. We next sought to determine if human YKT6 has a similar role. We have previously shown that combinatorial depletion of the human post-Golgi R-SNAREs VAMP3, 4, 7, and 8 does not block secretion in HeLa cells [27]. We depleted VAMPs 3, 4, 7, 8, and YKT6 individually or combination and determined the effect on secretion using our mammalian reporter line (HeLa-M C1). As previously reported depletion of VAMPs 3, 4, 7 and 8 individually causes little retention of the secretory cargo (Fig 6A and 6B). However, depletion of YKT6 causes partial retention of the cargo consistent with our previous results [27]. As in Drosophila cells combinatorial depletion of YKT6 and VAMP3 causes an almost complete block in secretion. This genetic interaction is specific because no interaction was detected with either VAMP4 or VAMP7. To investigate the specificity of the genetic interaction further we depleted the R-SNARE Sec22b in combination with YKT6 or VAMP3 (Fig 6A and 6B). As observed in Drosophila cells (Fig 3A–3C) no genetic interaction was detected between Sec22b and VAMP3 suggesting that the observed phenotype when Syb and YKT6 are depleted is not simply caused by a general defect in trafficking or toxicity. As in S. cerevisiae and the Drosophila cells we detect a strong negative genetic interaction between Sec22b and YKT6 [44].
The aim of this study was to identify the SNAREs required for the fusion of constitutive secretory carriers with the plasma membrane in higher eukaryotes. To address this we have developed a simple and robust assay for measuring secretion in Drosophila cells. Using well characterised targets (STX5, SLY1 and ROP) we have validated the system and have shown that the assay is capable of differentiating blocks in ER to Golgi and Golgi to plasma membrane transport based on proteolytic processing and accumulation of the secretory cargo in post-Golgi transport vesicles. Our experimental data suggests that there are at least three fusion complexes operating at the Drosophila PM (Fig 7A). The first complex consists of STX1, SNAP24/29 and Syb. The second complex consists of STX4, SNAP24/29 and Syb. The third complex consists of STX1, SNAP24 and YKT6. The reason we have excluded the possibility of a STX4, SNAP24/29, YKT6 complex is because depletion of both STX1 and Syb led to a complete block in secretion. Indicating that STX4 and YKT6 are unable to form a SNARE complex that can substitute for the loss of STX1 and Syb. Genetic interaction data also suggests that SNAP29 is unable to substitute for the loss of SNAP24 under conditions when both SNAP24 and Syb are depleted. This data suggests that the third SNARE complex specifically consists of STX1, SNAP24 and YKT6. At present it is unclear whether these SNARE complexes define parallel pathways to the plasma membrane or simply reflect the ability of these SNAREs to substitute with each other.
The most striking observation in this study is that we have uncovered an unexpected role for YKT6 in the fusion of secretory carriers with the plasma membrane. Depletion of YKT6 and Syb/VAMP3 in combination causes a complete block in secretion and leads to an accumulation of post-Golgi transport vesicles within Drosophila cells. YKT6 is a lipid anchored R-SNARE that has been shown to function on many pathways including ER to Golgi transport, intra-Golgi transport, endosome-vacuole fusion, endosome to Golgi transport and exosome fusion with the plasma membrane [24, 45–51]. YKT6 actively cycles on and off membranes in a palmitoylation dependant manner so potentially it is well suited to function on a wide variety of intracellular pathways [52]. Due to the promiscuous nature of YKT6 some caution must be taken when interpreting our functional data. It is possible that loss of YKT6 may be indirectly affecting post-Golgi transport and fusion at the plasma membrane. However, the simplest interpretation of our data is YKT6 is directly involved in this process as we are able to biochemically detect an interaction between YKT6 and STX1.
Using the knowledge obtained from the Drosophila system, we re-examined the role of R-SNAREs in constitutive secretion in mammalian cells. As previously reported, depletion of VAMP3 and other post-Golgi R-SNAREs did not perturb secretion in HeLa cells [27]. However, depletion of VAMP3 and YKT6 in combination caused a complete block in secretion. This data suggests that YKT6 and VAMP3 may be functioning in the fusion of secretory carriers with the plasma membrane in mammalian cells. We have made significant efforts to localise endogenous YKT6 and VAMP3 on post-Golgi secretory carriers. However, our attempts have been hampered by the fact the endogenus YKT6 is expressed at very low levels and over expressed YKT6 does not target correctly to membranes and remains cytoplasmic.
As expected, there is redundancy in the Q-SNAREs required for the fusion of secretory carriers with the plasma membrane. However, it is clear that certain SNAREs have a more prominent role in this process. The main Q-SNAREs at the Drosophila plasma membrane are STX1 and STX4 (share homology with SSO1 and 2). Depletion of STX1 causes a partial block in secretion while depletion of STX4 does not. It is unclear why STX1 is the favoured Qa-SNARE. It could simply be that STX1 is more abundant than STX4 or has a higher affinity for the R-SNARE on the vesicle [53]. It may also reflect the route by which the synthetic cargo traffics to the cell surface. We have also observed redundancy between the Qbc-SNAREs SNAP24 and SNAP29 (orthologues of Sec9). We are only able to detect a complete block in secretion when both are depleted. It has previously been shown that SNAP29 interacts with STX1. However, the complexes it forms are not SDS-resistant suggesting that they may not be fusogenic [54].
A potential problem with gene disruption and RNAi mediated depletion studies is compensation by other genes in the same family. For example, VAMP2 and 3 are upregulated in certain tissues of the VAMP8 knockout mouse and VAMP3 is upregulated in VAMP2 deficient chromafin cells isolated from VAMP2 null mice [55, 56]. Based on our immunoblotting data we did not observe any compensation between R-SNAREs when they are depleted using RNAi in Drosophila cells (Fig 3B). We also did not see any evidence of this in our previous work performed in HeLa cells [27]. We initially thought that STX1 and STX4 were being upregulated in STX5 and Syb depleted cells based on immunoblotting (Figs 2B and 4D). However, when the samples were directly prepared in Laemmli sample buffer, rather than a TX100 based extraction buffer, no difference in the levels of these SNAREs was observed (S2 Fig). It is possible that the change in extractability may be caused by an alteration in the localisation of the Q-SNAREs from TX100 insoluble micro-domains at the plasma membrane [57]. However, we have not tested this hypothesis. To directly assess changes in gene expression during the RNAi experiments we measured the mRNA levels several SNAREs using RT-PCR (S2 Fig). Depletion of STX1 leads to an upregulation of STX4 and Syb. However, we did not observe a significant change in the protein level of these SNAREs by immunoblotting. Thus it is unclear how significant these changes are. In the future, it will be interesting to determine how the expression levels of SNAREs, which function on the same pathway, are co-ordinated and regulated.
To validate our genetic interaction data we have interrogated a published S. cerevisiae proliferation-based genetic interaction map to determine if the yeast homologues share similar genetic interactions to those observed in Drosophila cells (under the assumption that constitutive secretion is essential for growth) [58]. We have detected negative genetic interactions between Drosophila STX1 and STX4, STX1 and Syb, Syb and SNAP24, SNAP24 and SNAP29, YKT6 and Sec22b and Syb and YKT6 (Fig 7C). Similar genetic interactions were also observed in S. cerevisiae indicating that the data generated from Drosophila cells is physiologically relevant and the genetic interactions are evolutionary conserved. Importantly the homologues of YKT6 and Syb/VAMP3 were also found to genetically interact in yeast (YKT6 and SNC2).
In summary, we have identified the SNARE complexes required for the fusion of constitutive secretory vesicles with the plasma membrane in Drosophila cells. We have uncovered a novel role for YKT6 in the fusion of secretory vesicles with the plasma membrane which is conserved from yeast to man. This observation may in part explain why RNAi and gene disruption studies in higher eukaryotes have failed to yield the expected phenotypes. In the future, it should be possible to use our secretion assay in combination with SNARE depletion as a tool to further dissect the post-Golgi pathways involved in secretion and generate post-Golgi secretory carriers for proteomic profiling.
Rabbit polyclonal antibodies were raised against GFP and the cytoplasmic domains of Drosophila STX4, SNAP29, Syb, VAMP7 and Sec22b. The antibodies were affinity purified as in [59]. The rabbit polyclonal antibody against Drosophila STX7 was a generous gift from H. Krämer. The mouse monoclonal antibodies against Drosophila STX1 (8C3, depositors Benzer, S. and Colley, N.), ROP (4F8) and Actin (JLA20, depositor Lin, J. J.-C) were purchased from the Developmental Studies Hybridoma Bank [39]. The Rabbit polyclonal to Drosophila GM130 was purchased from Abcam. The mouse monoclonal to human growth hormone (2H81G10) was a generous gift from Genentech Inc.,. The rabbit polyclonal antibody that cross-reacts with Drosophila YKT6 was a generous gift from Jessey Hay [60]. Secondary antibodies for immunoblotting were purchased from Jackson ImmunoResearch Laboratories. Secondary antibodies for immunofluorescence microscopy were purchased from Invitrogen Molecular Probes.
Drosophila D.mel-2 (Invitrogen) and C3 cells were maintained in Express Five® SFM media (Invitrogen,) supplemented with 100 IU/mL penicillin, 100 μg/mL streptomycin, and 2 mM glutamine (Sigma-Aldrich) at 25°C in an cooled incubator. Expression of the reporter construct in C3 cells was maintained by the addition of 5μg/mL Blasticidin (PAA Laboratories). HeLa-M and C1 cells were grown in high glucose DMEM supplemented with 10% fetal calf serum, 100 IU/mL penicillin, 100 μg/mL streptomycin, and 2 mM glutamine (Sigma-Aldrich) at 37°C in a 5% CO2 humidified incubator. Expression of the reporter construct in C1 cells was maintained by the addition of 1.66μg/mL puromycin (PAA Laboratories). siRNA transfections were performed as in Gordon et al., 2010. The sequence of the siRNA used in the experiments can be found in (S2 Table).
The reporter construct used to generate the C3 cell line was generated by subcloning the expression cassette from pC4S1-eGFP-FM4-FCS-hGH (Ariad Pharmaceuticals) into pAC-V5-His-A expression vector (Invitrogen). 2μg of pAC-S1-eGFP-FM4-FCS-hGH was co-transfected with 50ng of pCoBLAST into 500,000 S2 cells using the TransFast transfection reagent (Promega). A population of cells stably expressing the reporter construct was generated by the addition of 25 μg/mL blasticidin (PAA Laboratories). The cells were selected for two weeks and then autocloned into a 96 well plate using a MoFlo Flow cytometer (Beckman Coulter) based on GFP fluorescence. We were initially unsuccessful in this process until we supplemented the media with 5% FCS and put two cells in each well of the plate. 96 well plates were sealed with Parafilm M (Pechiney Plastic Packaging) to minimize evaporation during cell culture. Positive wells were identified using fluorescence microscopy. Clonal cell lines were screened for their ability to efficiently secrete the reporter construct and Clone 3 cells chosen as they have the most uniform expression of the reporter construct.
HA tagged Drosophila STX1 was generated using PCR and cloned into the copper inducible expression vector pMT/V5-HIS (Invitrogen). A stable population of cells was generated by co-transfecting the plasmid with pCoBLAST and selected as above.
Primers for generating dsRNA amplicons were designed using the Harvard Drosophila RNAi Screening Center database (http://www.flyrnai.org) or the GenomeRNAi database (http://rnai2.dkfz.de/GenomeRNAi). Amplicons were chosen which were predicted to have the fewest off-target hits. Primers sequences were copied verbatim from the websites and T7 sequences added to the 5’ end of both primers for each amplicon (S2 Table). Primers were synthesized by Sigma Genosys. A cDNA library was generated from S2 cells and used as a template for amplicon synthesis. The cDNA library was made by purifying RNA from S2 cells using a QIAshredder and RNeasy Protect Mini purification kit; followed by cDNA synthesis using the QuantiTect Reverse Transcription kit (Qiagen). The DNA template for the amplicons was generated using two rounds of PCR from the cDNA library. A sample of this DNA was sequenced to confirm that the correct target had been amplified. Double stranded RNA was synthesized using the DNA template and T7 Ribomax Express RNAi System (Promega) according to manufacturers’ instructions. The reaction was cleaned up using a DNAse and RNAse digestion step followed by column purification using the RNeasy Midi kit (Qiagen). A small amount of the reaction was run on agarose gel to confirm that the amplicon was the correct size. The RNA concentration was determined using a Nanodrop spectrophotometer (Thermo Scientific). Knock downs were performed by transfecting 20μg of dsRNA into 500,000 S2 cells using TransFast (Promega). The cells were then analysed 96 hours post transfection.
S2 cells were lysed and the RNA purified using a QIAshredder and RNeasy Protect Mini purification kit following the manufacturer’s instructions (Qiagen). The mRNA levels of specific genes were quantified using the Taqman RNA-to-CT 1-Step Kit (Applied Biosystems). Pre-designed sets of primers and FAM-labeled fluorescent probes designed against target genes were ordered from Applied Biosystems, and these were used according to manufacturers’ instructions (S3 Table). qRT-PCR reactions were run on an Applied Biosystems 7900HT Fast Real-Time PCR System. To quantify knockdown efficiency, relative quantification was performed using the ΔΔCT method [61]. For a list of qRT-PCR primes used in this study please see (S3 Table).
Secretion of the reporter construct was induced in Clone 1 (HeLa M) or Clone 3 (S2) cells by the addition of AP21998 (Ariad Pharmaceuticals) or D/D solubilizer (Clontech). Following secretion, the cells were placed on ice for 10 minutes to halt vesicle trafficking. C1 cells were detached using cold EDTA-Trypsin solution (PAA Laboratories) for 2 hours on ice. C3 cells are semi-adherent so were detached with pipetting. The fluorescence of the cells was measured using a BD FacsCalibur equipped with an HTS 96-well sampling robot (BD Biosciences). Live cells were gated using forward and side scatter and dead cell exclusion (2 μg/mL 7-AAD for clone 1 cells or 1 μg/mL PI for clone 3 cells) (Molecular Probes Invitrogen). A minimum of 2000 cells were analysed for each sample. FlowJo (Treestar) was used to calculate the geometric mean fluorescence for each sample. GraphPad Prizm (GraphPad Software) was used for generating statistics and graphs. Each sample is set up in duplicate. One sample receives AP21998 or D/D solubilizer and the other does not. The percentage of cargo remaining after secretion is then calculated by taking a ratio between the two samples.
To measure secretion by immunoblotting, equal numbers of C3 cells were resuspended in fresh media containing AP21998 and incubated for 80 minutes at 25°C. Secretion was halted by cooling the cells to 4°C and the media and cells collected by centrifugation. The media and cells were solubilized in Laemmli sample buffer and separated using SDS-PAGE. Proteins were transferred overnight onto PVDF membranes using wet transfer conditions. The membranes were blocked using 5% milk, 1% Tween-20 in PBS and probed with antibodies against actin (loading control) and growth hormone. Secondary antibodies conjugated to HRP were used to detect the primary antibodies and Supersignal West Pico Substrate (Pierce) used to develop the blots. Super RX Medical X-Ray Film (Fujifilm) was used to capture the signal and densitometry performed using ImageJ software. To evaluate knock down efficiency, C3 cells were counted and equal numbers of cells collected by centrifugation. The cells were resuspended in TX100-based extraction buffer (100 mM NaCl, 5 mM MgCl2, 50 mM Tris pH 7.4, 1% TX100), incubated for 15 minutes on ice, centrifuged at 15,000 g for 15 min at 4°C. The supernatants were normalised for protein concentration using the Bradford protein assay, (BIO-RAD), boiled in reducing SDS sample buffer and separated by polyacrylamide gel electrophoresis. Antibodies against actin (loading control) and SNAREs were used to probe the membranes.
To isolate HA-tagged syntaxin 1/SNARE complexes, cells were resuspended in lysis buffer (100 mM NaCl, 5 mM MgCl2, 50 mM Tris pH 7.4, 0.5% Igepal CA-630) with a complete protease inhibitor tablet (Roche) and incubated for 30 minutes. Insoluble material was removed by centrifugation at 5,000 rpm for 5 minutes and then followed by centrifugation at 50,000 rpm for 30 minutes. The lysate was then passed through a 0.2 μm syringe filter. Cleared lysate was incubated for two hours with anti-HA resin (Roche). Following multiple wash steps, samples were eluted twice with one column volume of 1 mg/mL HA-peptide (Roche) and acetone precipitated. The samples were then solubilized in Laemmli sample buffer and separated using SDS-PAGE. The gel was stained using SYPRO Ruby (Molecular Probes Invitrogen) and the bands visulaised using a Typhoon Trio Variable Mode Imager (GE Healthcare). The bands were excised using a scalpel blade and in-gel trypsin digestion performed. Analysis was performed using an AB Sciex 4800 MALDI TOF/TOF. The instrument is configured to acquire an MS spectrum between m/z 700 and 4000. From these MS spectra 7 peptides above a predetermined s/n threshold are selected for fragmentation. The MS spectra of intact peptides are used to determine protein identity by peptide mass fingerprinting (PMF) using the MASCOT search engine (NCBInr database 20/10/2010, 12061831 sequences). For further confirmation, the MSMS spectra are used to perform fragment ion searches to determine peptide sequence but if they fail to yield any identifications, it may be because peptides above the s/n threshold gave poor fragmentation patterns.
C3 cells were grown on 13 mm No. 1 round coverslips (VWR) coated with Concanavalin-A (Sigma-Aldrich) and allowed to adhere over night. Cells were incubated in the presence of AP21998 or D/D solubilizer for 80 minutes. Cells were then fixed and stained as described in [62]. Coverslips were mounted using ProLong Gold (Molecular Probes Invitrogen) and sealed with clear nail polish. Images were captured using either a 63x or 100x oil objective on a Zeiss Axioplan fluorescence microscope (Zeiss) equipped with a Hamamatsu Orca-R2 C10600 camera (Hamamatsu Photonics), and SEDAT quad pass filter set (Chroma). The brightness and contrast of microscopy images were adjusted using ImageJ (NIH).
|
10.1371/journal.pntd.0001980 | Distinct Transcriptional Signatures of Bone Marrow-Derived C57BL/6 and DBA/2 Dendritic Leucocytes Hosting Live Leishmania amazonensis Amastigotes | The inoculation of a low number (104) of L. amazonensis metacyclic promastigotes into the dermis of C57BL/6 and DBA/2 mouse ear pinna results in distinct outcome as assessed by the parasite load values and ear pinna macroscopic features monitored from days 4 to 22-phase 1 and from days 22 to 80/100-phase 2. While in C57BL/6 mice, the amastigote population size was increasing progressively, in DBA/2 mice, it was rapidly controlled. This latter rapid control did not prevent intracellular amastigotes to persist in the ear pinna and in the ear-draining lymph node/ear-DLN. The objectives of the present analysis was to compare the dendritic leukocytes-dependant immune processes that could account for the distinct outcome during the phase 1, namely, when phagocytic dendritic leucocytes of C57BL/6 and DBA/2 mice have been subverted as live amastigotes-hosting cells.
Being aware of the very low frequency of the tissues' dendritic leucocytes/DLs, bone marrow-derived C57BL/6 and DBA/2 DLs were first generated and exposed or not to live DsRed2 expressing L. amazonensis amastigotes. Once sorted from the four bone marrow cultures, the DLs were compared by Affymetrix-based transcriptomic analyses and flow cytometry. C57BL/6 and DBA/2 DLs cells hosting live L. amazonensis amastigotes do display distinct transcriptional signatures and markers that could contribute to the distinct features observed in C57BL/6 versus DBA/2 ear pinna and in the ear pinna-DLNs during the first phase post L. amazonensis inoculation.
The distinct features captured in vitro from homogenous populations of C57BL/6 and DBA/2 DLs hosting live amastigotes do offer solid resources for further comparing, in vivo, in biologically sound conditions, functions that range from leukocyte mobilization within the ear pinna, the distinct emigration from the ear pinna to the DLN of live amastigotes-hosting DLs, and their unique signalling functions to either naive or primed T lymphocytes.
| The rapid and long term establishment of parasites such as L. amazonensis, otherwise known to strictly rely on subversion of macrophage and dendritic leucocyte (DL) lineages, is expected to reflect stepwise processes taking place in both the skin dermis where the infective form of the parasite and the skin-draining lymph node (DLN) were inoculated. Relying on mice of two distinct inbred strains—C57BL/6 and DBA/2—that rapidly and durably display distinct phenotypes at the two sites of establishment of L. amazonensis, we were curious to address the following question: could live L. amazonensis-hosting DL display unique signatures that account for the distinct phenotypes? Based on flow cytometry, genechip and real-time quantitative PCR analyses, our results did evidence that, once subverted as cells hosting live L. amazonensis, DLs from C57BL/6 or DBA/2 do display distinct profiles that could account for the i) distinct parasite load profiles, ii) as well as the distinct macroscopic features of ear pinna observed once the L. amazonensis metacyclic promastigotes completed their four day developmental program along the amastigote morphotype.
| Leishmania (L.) amazonensis perpetuates in South and Central America, its main location being the wet forests of the Amazon basin. The perpetuation of this Leishmania species relies successively on two hosts which cohabit more or less transiently within this ecosystem: blood-feeding sand flies and mammals, including wild rodents and humans. A broad spectrum of clinical manifestations, ranging from single cutaneous lesions to multiple, disfiguring nodules [1], [2], [3] assess the durable establisment of L. amazonensis as intracellular amastigotes in the dermis. As model rodents, the laboratory mice of different inbred strains can be subverted as hosts by L. amazonensis, the establishment of parasites in the dermis being more or less rapid. In C3H, BALB/c and C57BL/6 mice high parasite loads, coupled to non healing skin-damages are displayed at site of L. amazonensis inoculation and in multiple skin sites reached by parasites emigrating from the primary inoculation site [4], [5], [6], [7], [8]. By contrast, in DBA/2 mice, at the inoculation site, the L. amazonensis population size is rapidly controlled, a process coupled to a controlled inflammatory process with limited parasite dissemination in distant tissue(s), if any [9].
Knowing that once in the dermis of the mouse, amastigotes are hosted by mononuclear phagocytes including macrophages and dendritic leukocytes (DLs) [10], [11], [12], [13], [14], [15], we have addressed the following question: could the DLs harbouring live amastigotes contribute to the distinct phenotypes observed in C57BL/6 and DBA/2 mice? Since the frequency of DLs hosting live Leishmania amastigotes within the skin and skin-draining lymph nodes (DLNs) remains very low [16], [17] we decided to first conduct an in vitro study relying on bone marrow-derived DLs (BMD-DLs) from C57BL/6 and DBA/2 mice exposed or not to live L. amazonensis amastigotes.
Based on flow cytometry (FCM), genechip (Affymetrix Mouse GeneChip) and real-time quantitative PCR (RT-qPCR) analyses performed on sorted DLs hosting live DsRed2-expressing L. amazonensis transgenic amastigotes [17] many distinct features have been highlighted. DBA/2 DLs displayed transcriptional signatures and markers that could be related to the early phenotype observed in vivo, in contrast to live amastigotes-hosting C57BL/6 DLs. The data are consistent with rapid and sustained immune regulatory functions accounting for the remodeling of the DBA/2 ear as L.amazonensis protective niche. All together this study provides, for the first time, a solid base for exploring i) the inflammatory processes that maintain the amastigote population under control in DBA/2 mice and ii) the inflammatory processes coupled to extended parasite dissemination and to poor parasite population control in C57BL/6 mice.
Six week old female DBA/2, C57BL/6 and Swiss nu/nu mice were purchased from Charles River (Saint Germain-sur-l'Arbresle, France).
All animals were housed in our A3 animal facilities in compliance with the guidelines of the A3 animal facilities at the Pasteur Institute which is a member of Committee 1 of the “Comité d'Ethique pour l'Expérimentation Animale” (CEEA) - Ile de France - Animal housing conditions and the protocols 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 B75-15-28 in accordance with the Ethics Charter of animal experimentation that includes appropriate procedures to minimize pain and animal suffering. TL is authorized to perform experiment on vertebrate animals (licence 75-717) issued by the Paris Department of Veterinary Services, DDSV) and is responsible for all the experiments conducted personally or under his supervision as governed by the laws and regulations relating to the protection of animals.
DsRed2-transgenic L. amazonensis strain LV79 (WHO reference number MPRO/BR/72/M1841) amastigotes were isolated from Swiss nude mice inoculated 2 months before within a BSL-2 cabinet space as described previously [17]. These amastigotes did not present any antibodies at their surface [18]. Promastigotes derived from amastigotes were cultured at 26°C in complete M199 medium. The metacyclic promastigote population (mammal-infective stage) was isolated from stationary phase cultures (6 day-old) on a Ficoll gradient.
Ten thousand metacyclic promastigotes in 10 µl of PBS were injected into the ear dermis of C57BL/6 and DBA/2 mice. Increased ear thickness was measured using a direct reading Vernier caliper (Thomas Scientific, Swedesboro, NJ) and expressed as ear thickness.
DLs were differentiated from bone marrow cells of DBA/2 or C57BL/6 mice according to a method described previously [18], [19]. Briefly, bone marrow cells were seeded at 4×106 cells per 100 mm diameter bacteriological grade Petri dish (Falcon, Becton Dickinson Labware, Franklin Lakes, NJ) in 10 ml of Iscove's modified Dulbecco's medium (IMDM; BioWhittaker Europe, Verviers, Belgium) supplemented with 10% heat-inactivated foetal calf serum (FCS; Dutscher, Brumath, France), 1.5% supernatant from the GM-CSF producing J558 cell line, 50 U/ml penicillin, 50 µg/ml streptomycin, 50 µM 2-mercaptoethanol and 2 mM glutamine. Cultures were incubated at 37°C in a humidified atmosphere with 5% CO2. On day 6, suspended cells were recovered and further cultured in complete IMDM supplemented with 10% of the primary culture supernatant before seeding on day 10 in hydrophobic 6-well plates (Greiner, St Marcel, France) at a concentration of 9×105 cells/well in 3 ml complete IMDM.
On day 4 post the distribution of DLs in the 6 well plate culture, DLs were exposed or not to freshly isolated DsRed2-LV79 amastigotes or to live BCG at micro-organism-DL ratios of 5∶1 and 10∶1, respectively. DL cultures were placed at 34°C and sampled at 24 hours post micro-organism addition. Recovered DLs were incubated first in PBS-FCS supplemented with 10% heat-inactivated donkey serum for 15 minutes, second in PBS containing 10% FCS and 0.01% sodium azide in presence of antibodies directed against surface antigens. Extracellular staining procedures were performed with specific monoclonal antibodies (mAbs) directed against MHC class II molecules (M5/114 clone) conjugated to PE-CY5 (0.2 µg/ml) and either of the following biotinylated mAbs directed against CD86 (GL1 clone), CD80 (K-10A1 clone), CD54 (3E2 clone), CD11c (HL3) and IgG control (B81-3 clone) at 0.5 µg/ml (eBioscience,San Diego, USA). Biotinylated mAbs were revealed using 1.5 µg/ml Streptravidin conjugated to Phycoeythrin (Molecular Probes, Cergy Pontoise, France). PE-conjugated mAb directed against CXCR-4 (2B11 clone) was purchased from eBioscience. Analysis was performed on the FACSCalibur. DLs were selected on FSC-SSC parameters (to excluded debris), and on the basis of MHC class II expression to discard the fraction of “contaminating” cells expressing no surface MHC class II molecules.
Intracellular staining of amastigotes was performed after fixation in PBS containing 1% paraformaldehyde (PFA) for 20 minutes at 4°C with the 2A3-26 mAb which was shown to strictly bind to the L. amazonensis amastigote [18]. DLs were washed in Perm/Wash solution from the BD Cytofix/Cytoperm™ Plus Kit (BD Bioscience) and incubated with 5 µg/ml of Alexafluor 488- conjugated 2A3-26 mAb in Perm/Wash buffer for 30 minutes at 4°C in the dark. Then DLs were washed in Perm/Wash buffer and fixed with in PBS −1% paraformaldehyde (PFA).
DLs were exposed or not to freshly isolated DsRed2-LV79 amastigotes at a parasite -DL ratio of 5∶1. DL cultures were placed at 34°C and sampled at 5, 24 and for 48 hours post parasite addition. Detached DLs were centrifuged on poly-L-lysine-coated glass coverslips and incubated at 34°C for 30 minutes. Cells were then fixed with 4% PFA for 20 minutes, permeabilised with saponin and incubated with 10 µg/ml of the amastigote-specific mAb 2A3-26-AlexaFluor 488 and 1 µg/ml of biotinylated-mAb (M5/114) directed against MHC class II molecules. The revelation was performed using 1.5 µg/ml streptravidin conjugated to Texas Red (Molecular Probes, Cergy Pontoise, France). Finally, they were mounted on glass slides with Hoechst 33342-containing Mowiol. Incorporation of Hoechst into DNA allowed the staining of both host cell and amastigote nuclei. Epifluorescence microscopy images were acquired on an upright microscope Zeiss Axioplan 2 monitored by the Zeiss Axiovision 4.4 software.
DsRed2-LV79 amastigotes were added or not to cultures of C57BL/6 and DBA/2-DLs. Twenty four hours later, three samples collected from three distinct cultures of either unexposed DLs or DLs exposed to DsRed2-LV79 amastigotes were carefully sorted as previously described by Lecoeur et al. [20]. Briefly cells were first incubated in PBS-FCS containing 0.2 µg/ml of the anti-MHC class II mAb (M5/114) conjugated to PE-Cy5-conjugated mAb (eBioscience). After two washes, cells were resuspended at 5×106 cells/ml in PBS containing 3% FCS and 1% J558 supernatant. The cell sorting was performed using a FACSAria (BD Biosciences, San Jose, CA) equipped with completely sealed sample injection and sort collection chambers that operate under negative pressure. PE-Cy5 and DsRed2 fluorescences were collected through 695/40 and 576/26 bandpass filters respectively. FSC and SSC were displayed on a linear scale, and used to discard cell debris with the BD FACSDiva software (BD Biosciences) [17]. L. amazonensis amastigote-hosting DLs were sorted by selecting cells expressing both surface MHC Class II molecules and DsRed2 fluorescence and immediately collected for RNA extraction by using the RNeasy Plus Mini-Kit (Qiagen) as previously described [21]. Whatever the readout assays-Affymetrix or RT-qPCR - the RNA populations used were prepared from the same samples. The quality control (QC) and concentration of RNA were determined using the NanoDrop ND-1000 micro-spectrophotometer (Kisker, http://www.kisker-biotech.com) and the Agilent-2100 Bioanalyzer (Agilent, http://www.chem.agilent.com).
Two hundred ng of total RNA per sample were processed, labelled and hybridized to Affymetrix Mouse Gene ST 1.0 arrays, following Affymetrix Protocol (http://www.affymetrix.com/support/downloads/manuals/expression_analysis_technical_manual.pdf). Three Biological replicates per condition were run. Following hybridization, the arrays were stained and scanned at 532 nm using an Affymetrix GeneChip Scanner 3000 which generates individual CEL files for each array. Gene-level expression values were derived from the CEL file probe-level hybridization intensities using the model-based Robust Multichip Average algorithm (RMA) [22]. RMA performs normalization, background correction and data summarization. An analysis is performed using the LPE test [23](to identify significant differences in gene expression between parasite-free and parasite-harbouring DLs, and a p-value threshold of p<0.05 is used as the criterion for significant differential expression. The estimated false discovery rate (FDR) was calculated using the Benjamini and Hochberg approach [24] in order to correct for multiple comparisons. A total of 1,340 probe-sets showing significant differential expression were input into Ingenuity Pathway Analysis software v5.5.1 (http://www.ingenuity.com), to perform a biological interaction network analysis. The symbols of the modulated genes are specified in the text (fold change [FC] values between brackets), while their full names are given in additional file 1. MIAME-compliant data are available through GEO database http://www.ncbi.nlm.nih.gov/geo/ accession GSE
Total RNAs from DLs cultures were reverse-transcribed to first strand cDNA using random hexamers (Roche Diagnostics) and Moloney Murine Leukemia Virus Reverse Transcriptase (Invitrogen, Life Technologies). A SYBR Green-based real-time PCR assay (QuantiTect SYBR Green Kit, Qiagen) for relative quantification of mouse target genes was performed on a 384-well plate LightCycler 480 system (Roche Diagnostics). Crossing Point values (Cp) were determined by the second derivative maximum method of the LightCycler 480 Basic Software. Raw Cp values were used as input for qBase, a flexible and open source program for qPCR data management and analysis [25]. Relative expression for 8 transcripts (ccl2, cl17, ccl19, ccr1, ccr2, cxcr4, cd274, tnfsf4) were calculated for sorted LV79-hosting DLs using sorted DLs from Leishmania unexposed cultures as calibrators. For normalization calculations, candidate control genes were tested (pgk1, h6pd, ldha, nono, g6pd, hprt, tbp, l19, gapdh, rpIIe and ywhaz) with the geNorm [26] and Normfinder programs [27]. Tbp and nono were selected as the most stable reference genes for the C57Bl/6 DLs. RpIIe and tbp were selected for the DBA/2 DLs.
At day 4 and 7 post the inoculation of 104 metacyclic promastigotes, three mice were sacrificed, the abundance of some transcripts being determined by real time RT-qPCR. Control, naïve mice were analyzed in parallel. Whole ear pinnas and ears-DLN were removed and fragmented using the Precellys 24 System [21]. Total RNAs were extracted and processed for RT-qPCR as described above. Ldha and nono were selected as the most stable reference genes for the C57Bl/6 and DBA/2 ears. tbp and nono were selected for the as the most stable reference genes for C57Bl/6 DLNs while ywhaz and nono were selected for the DBA/2-DLNs.
The experimental procedure for quantifying Leishmania in tissues was done as previously described by de La Llave et al [21]. Briefly, serial 10-fold dilutions of parasites (from 108 to 101) were added to either ears or ear-DLN recovered from C57BL/6 or DBA/2 naive mice. Total RNAs were extracted and processed for RT-qPCR as described above. The primers for Leishmania gene target (ssrRNA) to quantify the number of parasites were F- CCATGTCGGATTTGGT and R- CGAAACGGTAGCCTAGAG [28]. A linear regression for each standard curve was determined: number of parasites against the relative expression of ssrRNA values.
Two-sided Student's paired t-tests were used to compare FCM experiments (4<n<6). A Mann-Whitney test was used to compare ear thickness measurements and number of parasites.
C57BL/6 and DBA/2 mice were given into the ear pinna dermis a low number (104) of L. amazonensis (LV79 strain) metacyclic promastigotes. The monitoring of ear macroscopic features up to 100 days post inoculation (PI) has evidenced mouse inbred strain-specific features (Figure 1). C57BL/6 mice did not display any significant inflammatory signs during the early phase (ranging from day 0 to day 22 PI, phase 1), whereas they later display sustained inflammatory signs (after 22 days, phase 2; figures 1A, 1B). During the early phase, only a few parasites can be quantified in the ear pinna, the ear pinna-DLN displaying lower number of parasites (<100 parasites/DLN; figure 1C). In contrast, in DBA/2 mouse ear pinna, a mild inflammatory process was observed immediately post the inoculation whereas a rapid increase of the amastigote population size was noted in both the ears and ears-DLN. The second phase was delineated by the persistence of inflammatory process (Figure 1) coupled to the control of parasite load in the ear pinna and ear-DLN (data not shown).
We reasoned that early distinct DLs-dependent immune processes- promoting either rapid or slow remodeling of the dermis as amastigote-protective niches- could account for the distinct features displayed, over time, by the L. amazonensis amastigotes-hosting ear pinna of the C57BL/6 and DBA/2 mice. Being aware that, whatever the tissues, the DL frequency is very low, we considered biologically sound to start the comparative analysis with GM-CSF-dependent C57BL/6 or DBA/2 cultured DLs, once they were hosting, or not, live L. amazonensis amastigotes. Briefly, C57BL/6 and DBA/2 bone marrow cell suspensions were exposed or not to live DsRed2 L. amazonensis amastigotes and carefully sorted from otherwise heterogeneous cultures. The immunolabelling of surface MHC class II allowed us to exclude the low fraction of amastigote-hosting cells that did not express surface MHC class II. The subsequent step of such an approach was to first monitor, at the transcriptional level with the Affymetrix-based technology any potential distinct reprogramming of live L. amazonensis amastigotes-hosting DLs.
We used a carefully designed in vitro model [20] based on cultures of mouse BMD-DLs in which more than 97% of cells expressed CD11c, CD11a and CD11b (data not shown). When the presence/absence of surface MHC class II molecules was monitored on whole cell cultures by fluorescence microscopy and FCM, three phenotypically distinct cell subsets were evidenced (Figures S1A–C). The population of cells that did not express surface and intracellular MHC class II molecules were considered as “Contaminating” Cells (CC). The two other cell populations partition between i) a majority of cells displaying a moderate surface MHC class II amount (MHC IIlow; bona fide immature DLs) and ii) a minority of cells expressing very high levels of MHC II molecules (MHC IIhigh; bona fide mature DLs). DsRed2 L. amazonensis/LV79 amastigotes were put in contact with BMD-DLs (MOI of 5/1) and analysed 5, 24 or 48 hours later (Figure 2). Intracellular amastigotes (2A3-26+) detected by immunofluorescence microscopy analysis were evidenced in all BMD-DL subsets with much higher number of amastigotes in CC (data not shown). Low percentages of DLs hosting 2A3-26+ parasites were also documented by FCM analyses at 24 hours post amastigote addition (23.0%+/−12.6 and 26.0%+/−8.1 of 2A3-26+ cells in C57BL/6 and DBA/2 BMD-DLs, respectively, for n = 9 experiments). Interestingly, while the percentage of DLs housing amastigotes did not change from 5 hours to 24 hours (Figure 2A), the number of intracellular amastigotes did slowly expand whatever the mouse genotype (Figure 2B) over the otherwise limited temporal window we did focus on. L. amazonensis amastigote-hosting DLs were sorted by selecting cells expressing both surface MHC Class II molecules and DsRed2 fluorescence (see below).
|
10.1371/journal.pntd.0006076 | Virulence, pathology, and pathogenesis of Pteropine orthoreovirus (PRV) in BALB/c mice: Development of an animal infection model for PRV | Cases of acute respiratory tract infection caused by Pteropine orthoreovirus (PRV) of the genus Orthoreovirus (family: Reoviridae) have been reported in Southeast Asia, where it was isolated from humans and bats. It is possible that PRV-associated respiratory infections might be prevalent in Southeast Asia. The clinical course of PRV is not fully elucidated.
The virulence, pathology, and pathogenesis of two PRV strains, a human-borne PRV strain (isolated from a patient, who returned to Japan from Bali, Indonesia in 2007) and a bat-borne PRV (isolated from a bat [Eonycteris spelaea] in the Philippines in 2013) were investigated in BALB/c mice using virological, pathological, and immunological study methods.
The intranasal inoculation of BALB/c mice with human-borne PRV caused respiratory infection. In addition, all mice with immunity induced by pre-inoculation with a non-lethal dose of PRV were completely protected against lethal PRV infection. Mice treated with antiserum with neutralizing antibody activity after inoculation with a lethal dose of PRV showed a reduced fatality rate. In this mouse model, bat-borne PRV caused respiratory infection similar to human-borne PRV. PRV caused lethal respiratory disease in an animal model of PRV infection, in which BALB/c mice were used.
The BALB/c mouse model might help to accelerate research on the virulence of PRV and be useful for evaluating the efficacy of therapeutic agents and vaccines for the treatment and prevention of PRV infection. PRV was shown for the first time to be a causative virus of respiratory disease on the basis of Koch’s postulations by the additional demonstration that PRV caused respiratory disease in mice through their intranasal inoculation with PRV.
| It is assumed that Pteropine orthoreovirus (PRV) is a causative agent of acute respiratory tract infection (RTI) in humans. PRV was isolated from patients and fruit bats in Southeast Asia. Furthermore, the genome of PRV was detected in patients with respiratory symptoms, suggesting that PRV causes RTI in humans. There is a potential for PRV to cause RTIs in tropical or sub-tropical regions in Southeast Asia more widely than thought. The virulence, pathology, and pathogenesis of PRV in BALB/c mice were elucidated in detail. To develop specific countermeasures and prophylactics against PRV infection, an animal model of lethal PRV infection is needed, and we developed such a model using BALB/c mice. This model allows investigation of the pathogenicity of PRV and evaluation of the efficacy of drugs and vaccines and might help to accelerate research on the virology of PRV.
| Pteropine orthoreovirus (PRV), a member of genus Orthoreovirus in the family Reoviridae, was originally isolated from the heart blood of a grey-headed flying fox (Pteropus poliocephalus) in Australia in 1968 [1].
PRV was isolated from a patient with respiratory tract infection (RTI) as a causative agent in Malaysia in 2006 [2]. Seven patients with symptoms of influenza-like illness, such as fever, cough, and sore throat, caused by PRV were reported between 2006 and 2017 [2–8]. Three patients with PRV infection were reported in Malaysia in 2006 and 2010 [2, 3, 5, 8]. Four cases of PRV infection imported from Indonesia to Japan and Hong Kong were identified in 2007, 2009, and 2010 [4, 6–8]. The presence of these RTI cases in Southeast Asia suggests that PRV might be the causative viral pathogen of RTI. Some patients with PRV also showed the symptoms of abdominal pain, watery diarrhea, and vomiting [3–5, 8]. Antibodies to PRV were detected in 13% of the residents of Tioman Island, Malaysia [2], and 4.4% of patients with nonspecific symptoms in central Vietnam [9]. Furthermore, PRV genomes were detected in 17% of patients with RTIs in Negeri Sembilan state, Malaysia [10]. These reports raise the concern that the prevalence of human PRV infection in Southeast Asia might be higher than previously thought. However, the disease spectrum and the pathogenesis of PRV infection in humans also remain unclear.
Fourteen strains of PRV have been isolated from fruit bats (Pteropus poliocephalus, P. hypomelanus, P. vampyrus, Rousettus leschenaultia, Eonycteris spelaea, and R. amplexicaudatus) in Australia, Malaysia, Indonesia, PR China, and the Philippines from 1968 to date [1, 8, 11–15]. The Indonesia/2010 strain was isolated from the salivary swab of P. vampyrus imported from Indonesia to Italy in 2010 [14]. PRV-neutralizing antibodies were also detected in 83% of fruit bat species (R. amplexicaudatus, E. spelaea, and Macroglossus minimus) in the Philippines, suggesting that PRV is generally prevalent in some species of wild bats in Southeast Asia [15]. It is still not known whether PRV causes illnesses in fruit bats [8], whereas bat-borne PRV is a potentially pathogenic to humans. Therefore, it is important to characterize both human-borne and bat-borne PRV.
A PRV strain isolated from a patient with RTI was found to be lethal in C3H mice, but the virulence and pathology of this strain in mice were not investigated in detail [16]. In the present study, the virulence, pathology, and pathogenesis of PRV in BALB/c mice were elucidated to validate respiratory disease caused by PRV and to develop an animal model of PRV infection.
Two PRV strains that were isolated in previous studies [7, 15] were used in this study. The PRV strain Miyazaki-Bali/2007 (PRV-MB) was isolated from a patient with PRV infection, who returned to Japan from Bali, Indonesia in 2007 [7, 17]. The PRV strain Samal-24 (PRV-Samal-24) was isolated from E. spelaea in the Philippines in 2013 [15]. The nucleotide sequences of the 10 segments of each of these two PRV strains are deposited in GenBank (Table 1).
PRVs were propagated in human embryonic kidney-derived 293FT cells (Thermo Fisher Scientific, Inc.) for the preparation of the working virus solution. Cells infected with each strain of PRV were cultured at 37°C in Dulbecco’s modified eagle’s medium (DMEM; Sigma-Aldrich Co., LLC) supplemented with 5% heat-inactivated fetal bovine serum (FBS) and 1% antibiotics (penicillin and streptomycin; Pen-Strep, Thermo Fisher Scientific, Inc.) (DMEM-5FBS). After 2 days of culture, the medium was centrifuged at 800 × g for 5 min to remove cellular debris. The supernatant was overlaid onto 20% sucrose in a 50 ml tube (Becton Dickinson, Ltd.) and centrifuged at 100,000 × g for 2 h to concentrate the virus. The concentrated viruses were dissolved with DMEM with 2% FBS and 1% Pen-Strep (DMEM-2FBS), and the aliquots were stored at -80°C until use.
The infectious dose of each virus was determined in a plaque assay in Vero cell (ATCC, CCL-81) monolayers as described previously [7]. The cells were inoculated with a serially diluted virus solution of PRV-MB or PRV-Samal-24 and incubated for 1 h at 37°C for adsorption. The cell monolayers were washed with phosphate buffered saline solution (PBS), and the cells were cultured with DMEM-5FBS supplemented with 0.8% agarose for 2 days at 37°C. Plaque was visualized by staining the cells with neutral red solution. Plaques were counted, and the virus titers were calculated in plaque-forming units per milliliter (PFU/ml).
Nine-week-old female BALB/c mice (Japan SLC, Inc.) were used. The mice used were healthy and weighed approximately 20 g.
The mice, which were anesthetized with a combination of ketamine (100 mg/kg) and xylazine (4 mg/kg) in 0.9% sodium chloride solution, were inoculated with each strain of PRV. Five mice per group were intranasally inoculated with 1.0 × 103 to 1.0 × 106 PFU of each PRV strain in 20 μl DMEM-2FBS. The clinical signs and body weight of the mice were monitored for 14 days, and the 50% lethal dose (LD50) of PRV (for mice) was calculated according to the method of Reed and Muench [18]. Mice that were intranasally inoculated with 20 μl DMEM-2FBS (vehicle) were used as the control. The changes in body weight and the survival rates were plotted using the GraphPad Prism software program (GraphPad Software, Inc.) and were analyzed statistically by a one-way ANOVA.
Five mice were intranasally inoculated with 1.0 × 105 PFU of the PRV-MB or PRV-Samal-24 strain as described above. The mice were sacrificed on the 5th or 6th day post-infection (DPI), and then blood and the organs (the head including the brain and nasal cavity, trachea, lung, liver, kidney, spleen, and intestine) were collected. The viral RNA load in each organ and blood was determined by a quantitative real-time RT-PCR (qRT-PCR) as described below.
Blood samples were collected from the mice (5 per group) infected with each strain of PRV by cardiac puncture after euthanasia. Each of the blood samples was mixed with Isogen LS (Wako Pure Chemical Industries, Ltd.), and total RNA was extracted from each blood sample according to the manufacturer’s instructions. The organs and tissues; the brain, nasal cavity, trachea, lung, heart, liver, spleen, kidney, and intestine were collected. These samples were immediately submerged in RNAlater (Ambion, Life Technologies, Inc.) and stored at -80°C until use. Total RNA was extracted using Isogen (Wako Pure Chemical Industries, Ltd.) according to the manufacturer’s instructions. The viral copy numbers were determined with a qRT-PCR as follows. The forward and reverse primers and probe were specifically designed according to the nucleotide sequence of the outer-capsid protein (OCP) region in the S4 segment of PRV-MB or to that of PRV-Samal-24. The sequences of the forward primer, the reverse primer, and the probe for the amplification of the PRV-MB genome were 5’-CATTGTCACTCCGATCATGG-3’, 5’-TGGGAGTGTGCAGAGCATAG-3’ (Eurofins Genomics, Inc.), and FAM/5’-GTAGGTATGCCACTCGTGGAATCC-3’/TAMRA (Sigma-Aldrich Co. LLC.), respectively. The sequences of the forward primer, the reverse primer, and the probe for the amplification of the PRV-Samal-24 genome were 5’-CAATTTCCACTCGTTCGTTG-3’, 5’- GATGGTGTGGAAACGGATAC -3’ (Eurofins Genomics, Inc.), and FAM/5’- GACCAGACCAGATACGTGGAATCC -3’/TAMRA (Sigma-Aldrich Co. LLC.), respectively. The qRT-PCRs were performed using a Light Cycler 96 system (Roche Diagnostics, Ltd.) with a QuantiTect Probe RT-PCR Kit (Qiagen, Ltd.). The Light Cycler experimental protocol was as follows: reverse transcription (50°C for 30 min), denaturation (95°C for 15 min), and 45 cycles of amplification and quantification (94°C for 15 s and 60°C for 60 s), followed by a final cooling step at 40°C for 30 s. In this study, the standard controls for PRV-MB and PRV-Samal-24 were 10-fold serial dilutions of the plasmid DNA containing the S4 segments of PRV-MB and PRV-Samal-24, respectively. The viral copy numbers in the samples were calculated as the ratio of the copy numbers of each standard control. The viral copy numbers were plotted using the GraphPad Prism software program, and the results were statistically analyzed by a one-way ANOVA. The viral RNA detection limits in the blood, trachea, and other tissues were determined to be 2.5 × 103 copies/ml, 1.6 × 103–5.0 × 103 copies/0.1 g, and 2.5 × 103 copies/0.1 g, respectively. One PFU was equivalent to 2.9 copies of viral RNA.
Two mice were intranasally inoculated with 1.0 × 106 PFU of the PRV-MB strain or of PRV-Samal-24 strain as described above. The mice were sacrificed on the 4th DPI, and organs (the head including the brain and the nasal cavity, trachea, lung, liver, kidney, and intestine) of the mice were collected. The infectious virus titer in each organ was determined with a plaque assay as described below. Each organ collected was immediately submerged in DMEM-2FBS, homogenized and centrifuged at 800 × g for 5 min to remove debris. The supernatant fraction was collected and stored at -80°C until use. The virus titer in the supernatant fraction was determined in a plaque assay in Vero cell monolayers as described previously [7]. The virus titers were plotted using the GraphPad Prism software program. The virus titer detection limit was determined to be 2.4 × 101 PFU/0.1 g.
The mice intranasally infected with 1.0 × 103 PFU of PRV-MB (PRV-MB-1.0×103 PFU mice) and those intranasally infected with 1.0 × 105 PFU of PRV-MB (PRV-MB-1.0×105 PFU mice) were sacrificed by exposure to excess isoflurane, and the lungs were collected on the 1st, 3rd, and 5th DPI (5 mice per group each day). The viral RNA loads in the lungs were determined with qRT-PCR. Pathological analyses of the lungs were performed by immunohistochemical (IHC) analysis as described below.
The collected tissues were stained with hematoxylin and eosin (H&E) for histopathology. An IHC analysis was performed for the detection of PRV antigen in the tissues. The IHC analysis methods were the same as those described previously except for the antigen detection antibody [19]. The sections were deparaffinized by placing them in a retrieval solution (pH 6) (Nichirei Biosciences, Inc.), followed by heat-treatment with an autoclave at 121°C for 10 min. The polyclonal antibody to the OCP (S4 segment) of PRV-MB raised in a rabbit by immunization with the antigen (OCP antibody) was used for the IHC detection of PRV antigen [20]. The OCP antibody used in the IHC analysis reacted specifically with OCP antigens of PRV [20]. To validate whether the OCP antibody reacts with the mouse lungs non-specifically, the lung tissues of 6-month-old BALB/c mice infected with severe acute respiratory syndrome coronavirus (SARS-mouse-lung), in which severe inflammation was shown, and those of the mice inoculated with mock solution (mock-mouse-lung) were tested by IHC analysis [21]. The samples showed a negative reaction in the IHC analysis (S1 Fig), indicating that the OCP antibody does not react non-specifically with the mouse lungs with inflammation and that the positive signals detected in the IHC analysis indicate the presence of the OCP of PRV. As the negative control, normal rabbit serum (NRS; Dako, Ltd.) was used in IHC analysis. After treatment, the sections were reacted with the OCP antibody or NRS and then washed with PBS. The sections were incubated with Nichirei-Histofine Simple Stain Mouse MAX PO (R) (Nichirei Biosciences, Inc.) according to the manufacturer’s instructions. The peroxidase activity was detected with 3, 3’-diaminobenzidine (Sigma-Aldrich Co. LLC.), and the sections were counterstained with hematoxylin.
Five mice were intranasally inoculated with either DMEM-2FBS containing 1.0 × 103 PFU (non-lethal dose) of PRV-MB or DMEM-2FBS (control). Serum was separated from the blood collected through the caudal vein on the 27th DPI by centrifugation. The serum was tested for the PRV-MB neutralizing antibody titers as described previously [9, 15]. In addition, mice that were pre-inoculated with PRV-MB or control were re-inoculated with 1.0 × 105 PFU (lethal dose) of PRV-MB on the 35th day after the first inoculation with PRV-MB or control. The clinical signs and body weight were monitored for 14 days. The mice that showed >25% initial body weight loss were euthanized. Their body weight changes and survival rates were plotted using the GraphPad Prism software program.
Twenty-five mice were intranasally inoculated with 1.0 × 103 PFU (non-lethal dose) of PRV-MB followed by a second intranasal inoculation with 1.0 × 105 PFU of PRV-MB 3 weeks after the first infection. The mice were then intranasally inoculated once more with 1.0 × 105 PFU of PRV-MB 3 weeks after the second infection. On the 5th day after the third inoculation, the mice were sacrificed and blood was collected by cardiac puncture. Serum was separated by centrifugation. Mouse serum, which was collected from the 25 control mice without inoculation with PRV-MB, was used as the control serum.
The serum was diluted 4-fold with PBS. Five mice per group were intranasally infected with 1.0 × 105 PFU of PRV-MB, and then the diluent of the serum (100 μL) was administered once daily until the mice showed >25% initial body weight loss for a maximum of 5 days. Serum was administered at just after 1 h after inoculation, or on the 1st, 2nd, 3rd, and 4th DPI. The diluent of the control serum (100 μL) was used for mock treatment. The mice that showed >25% initial body weight loss were euthanized. The body weight changes and survival rates were plotted using the GraphPad Prism software program.
The animal studies were carried out in strict accordance with the Guidelines for Proper Conduct of Animal Experiments of the Science Council of Japan and in strict compliance with animal husbandry and welfare regulations. All animal experiments were approved by the Committee on Experimental Animals at the National Institute of Infectious Diseases (NIID) in Japan (Approval Nos. 215016, 116086, and 116082). All of the animals infected with PRV were handled in biosafety level 3 animal facilities, in accordance with the guidelines of the NIID. The mice were inoculated with virus solution under proper anesthesia, and all efforts were made to minimize any potential pain and distress. After inoculation, the animals were monitored once a day during the study period. A humane endpoint was introduced for all mice with >25% initial body weight loss.
The PRV-MB-1.0×105 PFU mice or the PRV-MB-1.0×106 PFU mice developed symptoms (piloerection, slowness in movement, anorexia, and weight loss) from the 2nd DPI. All of the mice died by the 6th DPI (Fig 1). The severity of the symptoms in the PRV-MB-1.0×104 PFU mice was less than that in the PRV-MB-1.0×105 PFU mice or the PRV-MB-1.0×106 PFU mice, and 3 of the 5 mice died by the 8th DPI. The extent of body weight loss in the PRV-MB-1.0×103 PFU mice was greater than that in the control mice. The PRV-MB-1.0×103 PFU mice did not show any symptoms other than body weight loss. The LD50 of PRV-MB in the BALB/c mice was determined to be 6.8 × 103 PFU/head.
The level of viral RNA in the lungs (average level, 6.9 × 108 copies/0.1 g) was higher than those in the other organs (Fig 2A, left panel). Viral RNA was detected in the blood (maximum level of 7.5 × 106 copies/ml) (Fig 2A, right panel). In contrast, viral RNA was not detected in the brain, heart, liver, spleen, kidney, and intestine. The infectious virus was detected mainly in respiratory organs (Fig 2B). The titer in the lungs (average virus titer, 6.4 × 104 PFU/0.1 g) was the highest among the organs tested.
A pathological examination revealed tissue damage and inflammation (i.e., necrosis and the accumulation of inflammatory cells including lymphocytes) in the lower respiratory tract, including the bronchiole and alveoli, in which viral antigens were detected in IHC analysis by using the OCP antibody, on the 4th DPI (Fig 3A, left and middle panels). Neutrophils and type II pneumocytes infiltrated to the alveoli and alveolar walls, and tissue damage in the lungs was detected (Fig 3B). The PRV antigen-positive lesions revealed in the IHC analysis by using the OCP antibody showed negative reaction in the IHC analysis using NRS (Fig 3A, right panels). No pathological changes or viral antigens were detected in the other tissues examined.
The viral RNA in the lungs of the PRV-MB-1.0×103 PFU mice or the PRV-MB-1.0×105 PFU mice was determined throughout the course of infection (Fig 4). On the 1st DPI, the viral RNA load in the lungs of the PRV-MB-1.0×105 PFU mice was similar to that of the PRV-MB-1.0×103 PFU mice. In contrast, on the 3rd and 5th DPI, the viral RNA load in the lungs of the PRV-MB-1.0×105 PFU mice was significantly higher in comparison to the PRV-MB-1.0×103 PFU mice.
The presence of viral antigens in the lungs of the PRV-MB-1.0×103 PFU mice and the PRV-MB-1.0×105 PFU mice was investigated immunohistochemically on the 1st, 3rd, and 5th DPI. Viral antigens were detected in the bronchial epithelium of the PRV-MB-1.0×105 PFU mice on the 1st DPI (Fig 5B, upper panel), in the alveolar duct, alveoli, and bronchial epithelium on the 3rd DPI, and in the alveolar area on the 5th DPI (Fig 5B, middle and lower panels). Cellular damage characterized by positive nuclear aggregation, cellular atrophy, and cellular debris was detected in the terminal bronchioles, which was also positive for PRV-MB antigen (Fig 6). PRV-MB caused extensive and massive pulmonary infection in the PRV-MB-1.0×105 PFU mice. In contrast, few viral antigens were detected in the bronchial epithelium of the PRV-MB-1.0×103 PFU mice on the 1st and 3rd DPI (Fig 5A, upper and middle panels, respectively), and no viral antigens were detected in the bronchial epithelium or alveoli on the 5th DPI (Fig 5A, lower panel).
The serum neutralizing antibody titers induced in the mice inoculated with 1.0 × 103 PFU (non-lethal dose) of PRV-MB on the 27th DPI were between 640 and 2560. The mice were then challenged with 1.0 × 105 PFU of PRV-MB on the 35th day after the first inoculation with a non-lethal dose of PRV-MB. All of these mice survived, whereas all of the control mice died by the 6th DPI (Fig 7).
A mixture of the serum sample collected from mice infected with PRV-MB was used as an antiserum with a PRV-MB-specific serum neutralizing antibody titer of 10,240. The administration of antiserum to the mice that had been infected with 1.0 × 105 PFU of PRV-MB showed a protective effect: the survival rate of the anti-serum-treated mice was 60%, whereas all of the control mice died (Fig 8). When the antiserum treatment was initiated on the 1st or 2nd DPI, taking the day on which the mice were infected with PRV-MB as day 0, 40% of the mice survived, whereas the control mice and the mice in which the treatment was initiated on the 3rd DPI or later died by the 6th DPI. The body weight reduction in these groups was similar to that of the control group.
Nine-week-old BALB/c mice were infected with a graded dosage (1.0 × 103–1.0 × 106 PFU) of PRV-Samal-24. The intranasal inoculation of the mice with PRV-Samal-24 led to fatal outcomes. The LD50 of PRV-Samal-24 for BALB/c mice was determined to be 4.2 × 103 PFU/head.
Among the respiratory organs, viral RNA was detected in the lungs of the mice infected with PRV-Samal-24; the viral RNA copy numbers in the lungs were up to 3.7 × 108 copies/0.1 g on average (Fig 9A, left panel). Viral RNA was also detected in the blood (maximum level, 1.8 × 106 copies/ml) (Fig 9A, right panel). In addition, the infectious virus was isolated from the respiratory tract organs (Fig 9B). The infectious dose of PRV-Samal-24 was the highest in the lungs among the tissues tested with the dose being up to 9.5 × 103 PFU/0.1 g on average.
Pathological examination of the PRV-Samal-24-1.0×106 PFU mice revealed inflammatory lesions in the lungs (by H&E staining), and viral antigens were also detected, especially from the bronchioles to the alveoli (by IHC staining) as was observed in the mice infected with PRV-MB (S2 Fig).
The present study showed through virological and pathological examinations that the lung was the principle target organ of PRV replication after intranasal inoculation in BALB/c mice. PRV mainly replicated in the bronchiolar epithelium by the 3rd DPI. The bronchiolar epithelium is composed of ciliated and nonciliated cells, such as clara cells and goblet cells, which are classified as secretory cells [22]. Morphologically, the PRV antigen-positive cells were likely to be clara cells and goblet cells. PRV infection caused severe inflammation in the lungs of the mice on the 4th DPI (acute phase). Morphologically, PRV mainly replicated in the pneumocyte-like cells, and the PRV antigen-positive cells were likely to be type I pneumocytes, which are involved in the process of gas exchange between the alveoli and blood [23]. Mammalian orthoreovirus, which is classified to the genus Orthoreovirus in the family Reoviridae, was reported to replicate in type I pneumocytes and was shown to induce severe pneumonia in some rodent species, including mice and rats [24, 25]. Type I pneumocytes might be a critical replication site for PRV. It was assumed that fatal outcomes were induced in mice infected with a lethal dose of PRV due to a decrease in respiratory function that occurred as a result of the destruction of the bronchiolar epithelial cells and pneumocytes. In this study, the cell types, in which PRV replicated, were identified only by morphological observation. Further studies are needed to elucidate the primary target cells, which are infected with PRV and in which PRV replicates. The high-titer PRV genome and infectious PRV were detected in the lungs of the mice infected with a lethal dose of PRV on the 4th to 6th DPI (acute phase) (Figs 2 and 9). BALB/c mice were susceptible to PRV and developed RTIs, similarly to humans. Demonstration of infectious PRV in lungs indicates that PRV definitely replicated there (Figs 2B and 9B). Koch’s postulates (i.e., isolation of PRV from patients with RTIs, induction of RTI in mice by infection with PRV, and detection of infectious PRV in respiratory organs of mice infected with PRV) support a causal role of PRV infection in the development of respiratory tract diseases in humans [26]. Although all of the cases of PRV infection in humans showed symptoms associated with RTI, it is evident that the clinical characteristics of PRV infections in humans have not been fully elucidated. It is possible that PRV causes more severe infections than have previously been reported.
We evaluated the utility of the newly developed mouse model of PRV infection. Immunity to PRV was induced by non-lethal infection, and it protected the mice from lethal infection with PRV (Fig 7). The early initiation of antiserum treatment was effective in the treatment of lethal PRV infection (Fig 8). These results suggest that BALB/c mice may serve as a useful animal model for evaluating the efficacy of vaccines and therapeutic agents for PRV.
The pathogenicity of PRV-Samal-24 was evaluated in this mouse model. Similarly to PRV-MB, PRV-Samal-24 caused viremia and respiratory disease in the BALB/c mice. The amino acid identities (encoded by each gene segment) between PRV-Samal-24 and PRV-MB were as follows: cell attachment protein region of the S1 segment, 82%; p10 region of the S1 segment, 100%; p17 region of the S1 segment, 94%; inner-capsid protein region of the S2 segment, 97%; sigma NS region of the S3 segment, 97%; OCP region of the S4 segment, 97%; minor inner-capsid protein region of the M1 segment, 94%; major outer-capsid protein region of the M2 segment, 95%, mu NS region of the M3 segment, 91%, guanylyltransferase region of the L1 segment, 94%, RNA polymerase region of the L2 segment, 98%, and major inner-capsid protein region of the L3 segment, 98%. PRV-Samal-24 and PRV-MB were also reported to show cross-reactivity in an immunofluorescence assay [15]. As both the in vivo and in vitro characteristics of PRV-Samal-24 are similar to PRV-MB, it is possible for bat-borne PRV to cause illness in humans.
In conclusion, a BALB/c mouse model of PRV infection, in which PRV caused acute RTI, was developed. Immunocompetent BALB/c mice were sensitive to PRV, when the mice were infected with PRV intranasally. This model might be useful for analyzing the pathogenicity of PRV in mice and for evaluating the efficacy of vaccines and therapeutic agents that will be developed to prevent and treat PRV infection. This model is also useful for further studies on PRV infections in vivo.
|
10.1371/journal.pcbi.1001002 | Environments that Induce Synthetic Microbial Ecosystems | Interactions between microbial species are sometimes mediated by the exchange of small molecules, secreted by one species and metabolized by another. Both one-way (commensal) and two-way (mutualistic) interactions may contribute to complex networks of interdependencies. Understanding these interactions constitutes an open challenge in microbial ecology, with applications ranging from the human microbiome to environmental sustainability. In parallel to natural communities, it is possible to explore interactions in artificial microbial ecosystems, e.g. pairs of genetically engineered mutualistic strains. Here we computationally generate artificial microbial ecosystems without re-engineering the microbes themselves, but rather by predicting their growth on appropriately designed media. We use genome-scale stoichiometric models of metabolism to identify media that can sustain growth for a pair of species, but fail to do so for one or both individual species, thereby inducing putative symbiotic interactions. We first tested our approach on two previously studied mutualistic pairs, and on a pair of highly curated model organisms, showing that our algorithms successfully recapitulate known interactions, robustly predict new ones, and provide novel insight on exchanged molecules. We then applied our method to all possible pairs of seven microbial species, and found that it is always possible to identify putative media that induce commensalism or mutualism. Our analysis also suggests that symbiotic interactions may arise more readily through environmental fluctuations than genetic modifications. We envision that our approach will help generate microbe-microbe interaction maps useful for understanding microbial consortia dynamics and evolution, and for exploring the full potential of natural metabolic pathways for metabolic engineering applications.
| Microbial metabolism affects biogeochemical cycles and human health. In most natural environments, multiple microbial species interact with each other, forming complex ecosystems whose properties are poorly understood. In an effort to understand inter-microbial interactions, and to explore new metabolic engineering avenues, researchers have started building artificial microbial ecosystems, e.g. pairs of genetically engineered strains that require each other for survival. Here we computationally explore the possibility of creating artificial microbial ecosystems without re-engineering the microbes themselves, but rather by manipulating the environment in which they grow. Specifically, using the framework of flux balance analysis, we predict environments in which either one or both microbes in a pair would not be able to grow without the other, inducing commensal (one-way) or mutualistic (two-way) interactions, respectively. Our algorithms can successfully recapitulate known inter-microbial interactions, and predict millions of new ones across any pair amongst different microbial species. Surprisingly, we find that it is always possible to identify conditions that induce mutualistic or commensal interactions between any two species. Hence, our method should help in mapping naturally occurring microbe-microbe interactions, and in engineering new ones through a novel, environment-driven branch of synthetic ecology.
| While several aspects of microbial metabolism can be fruitfully addressed by studying individual microbial species, many contemporary challenges, including environmental remediation and infectious diseases, require a massive effort towards understanding how microbes interact with each other. In fact, in nature, most microbes do not live in isolation, but rather exist as part of complex, dynamically changing, microbial consortia [1], [2]. From a metabolic perspective, the coordinated action of multiple interacting microbes is known to enable specific metabolic processes, such as the bio-geochemical process of nitrification that occurs in soil and marine water [3], pesticide degradation in agricultural settings [4], anaerobic methanogenesis in animal rumen, fresh water ponds and sewage sludge digester [5], anaerobic oxidation of methane in marine environments [6] or degradation of xylan or complex oligosaccharides in the microbial flora of the human gut [7], [8]. Metabolic interdependencies are also thought to partially be associated with the problem of microbial unculturability [9].
Metabolic interactions between pairs of microbial species could be thought of as unidirectional or bidirectional exchanges of small molecules, which may benefit one or both species (Table 1). A commensal interaction is a one-way exchange, where one organism is dependent on the product of the other. An obligate bidirectional exchange (commonly referred to as cross-feeding, syntrophy or mutualism) is perhaps the most fascinating of all possible interactions. Such an interaction implies a mutual dependence, which seems contingent on the rise of improbable matching of resource requirements and availabilities. Metabolic syntrophy is thought to drive fundamental biogeochemical processes (Fig. 1, [10]–[13]), either through the mutual benefit of a uni-directional nutrient exchange (Fig. 1A), or through bi-directional cross-feeding [8], [10], [11], [14], [15]. In addition, engineered species can be induced to display mutualistic interactions, as shown in classical work aimed at unraveling the order of metabolic reactions in biosynthetic pathways [16]–[18], and in recent synthetic ecology experiments [19]–[21] (Fig. 1B).
In parallel to experimental studies, the rise of genome-scale constraint-based models of metabolism has the potential to help address questions that cannot be easily addressed experimentally. Constraint-based models of metabolic networks represent an efficient framework for a quantitative understanding of microbial physiology [22] (see Methods). Such models rely on the knowledge of the stoichiometry for every known metabolic reaction taking place in the cell, and focus on predicting steady state fluxes (i.e. reaction rates) rather than time-dependent metabolite concentrations. By focusing on the fluxes, one can view cellular metabolism as a resource allocation problem: given that the system has internal stoichiometric and thermodynamic constraints, and a certain amount of nutrients available, how should the flow through the network be distributed to allow the cell to achieve a given biological task, e.g. grow at maximal possible rate? This approach, also known as flux balance analysis, has been described in detail elsewhere [23]–[26], and given rise to a plethora of interesting discussions on optimality in metabolic network regulation and evolution [23], [27]–[30]. In the study of microbial ecosystems, it has been recognized that the extension of constraint-based models from individual to multiple interacting species or compartments involves novel challenges and opportunities [31]–[33]. In particular it has been shown that stoichiometric models of individual species can be combined to provide testable predictions about ecosystem-level behavior [32], [33]. The alternative method of network expansion has been used to identify putative metabolic synergy between all pairs of nearly 450 organisms in a single environmental setting [34]. Moreover, in broader context, evolutionary and functional insight was obtained through large meta-metabolism models that ignore the spatial distinctions between different organisms [35]–[37].
Here, we use constraint-based models to develop a new strategy for the study of metabolism-based symbiotic interactions in pairs of microbial species. While in most analyses of cross-feeding interactions the focus is on the properties of the organisms themselves, we take a different approach, asking whether, given two arbitrary organisms, it is possible to identify environmental conditions that induce a mutualistic or commensal interaction. We start by exploring known symbiotic pairs, to determine if available stoichiometric models seem to provide predictions that are in agreement with empirical observations. The algorithms we developed allow us not only to verify potential interactions, but also to produce lists of putatively exchanged metabolites. Then, we ask whether, given any two species whose stoichiometric models are available, it is possible to predict potential nutrient compositions that induce specific symbiotic behaviors, in particular commensalism and mutualism. Hence, taking advantage of the efficiency of constraint-based models, we explore the large space of possible media compositions, in search for nutrient combinations that sustain a co-culture of two species but do not support growth of each organism on its own. We apply our pipeline to the prediction of novel environments and interactions for a coculture of Escherichia coli and Saccharomyces cerevisiae, and for all pairwise combinations of seven bacterial species: Escherichia coli, Helicobacter pylori, Salmonella typhimurium, Bacillus subtilis, Shewanella oneidensis, Methylobacterium extorquens, and Methanosarcina barkeri. In addition to providing an algorithmic platform for synthetic ecology exploration, we envisage that our approach will help mapping and understanding interactions that occur in natural microbial consortia.
As a first step in our analysis we asked whether, using stoichiometric models, we could reproduce three metabolic interactions depicted in Figs. 2 and 1. The first, simplest interaction (Fig. 2D) is an elementary case of syntrophy in a toy model. The second interaction (Fig. 1A) is the previously modeled [33] naturally occurring interaction between a hydrogen producing bacterium and a methanogen archaeon [38], [39]. The third (Fig. 1B) is an obligate mutualism between two strains of yeast engineered to be auxotrophic for lysine (Lys-) and adenine (Ade-) respectively [19]. In each case, we built a joint model for the organism pair by combining their stoichiometric matrices into a single ecosystem-level stoichiometry (Fig. 2). This unified stoichiometry involves the creation of a new compartment (the joint environment) that can communicate with the individual species and serves as an interface for the description of environmental nutrient availability (see Fig. 2 and Methods). In the cases of Figs. 2 and 1B, upon building the joint stoichiometry, we could verify that the pairs of organisms could grow only syntrophically. Similarly, for the case of Fig. 1A, we verified that our model implementation reproduced the unidirectional flow of nutrients responsible for the symbiotic relationship between the two species.
We next asked whether the stoichiometric implementation of organism pairs could be used to generate predictions of the metabolites exchanged between the two species upon symbiotic growth. Towards this goal we developed an algorithm that allows us to predict what metabolites need to be exchanged between two species in order to survive under a given environmental condition (Fig. 3). Our search for exchanged metabolites (SEM) algorithm constitutes an extension of flux balance modeling, applied to the unified ecosystem-level stoichiometry (Fig. 2) [31]. SEM is based on a mixed integer linear programming algorithm that identifies the fewest number of metabolites exchanged between individual species, under the constraint that both organisms must still be able to produce biomass at a rate larger than a given minimal threshold. For a particular medium, SEM can recursively find multiple optimal or near-optimal solutions, though it is not guaranteed to identify all possible ones (See Methods for more details). In the toy model (Fig. 2C) and in the methanogenic pair (Fig. 1A) we recovered the expected exchange metabolites. The results were less obvious in the case of complementary yeast auxotrophs (Fig. 1B). In this case, upon applying SEM to the yeast pair, we found two feasible sets of exchanged metabolites under the glucose minimal medium used in the experiment. These sets both use lysine and then either adenosine-(3,5)-bisphosphate (PAP) or hypoxanthine (HXAN) (Fig. S2). While the observed exchange of lysine confirms the intuition, reflecting the deficiency of one of the engineered organisms (Lys-), it was somehow surprising not to observe, for analogous reasons, a predicted exchange of adenine. The reason for this discrepancy can be explained by looking at the relevant metabolic pathways (Fig. 4). Although several metabolites downstream of the ade8 reaction might restore flux through the rest of the adenine biosynthesis pathway, based on the stoichiometric model [40] only PAP and HXAN can be reversibly transported. As indicated by the metabolic pathway map, both of these metabolites are easily converted into adenosine via mechanisms that circumvent the knocked out enzyme of the Ade- strain. Thus our algorithm produced testable predictions on potentially exchanged metabolites between the two yeast strains in the engineered syntrophic pair.
So far, we have shown that joint stoichiometric models can be helpful in describing known cases of symbiotic interactions, providing novel testable predictions. This analysis, however, has been limited to a single medium composition. Can one generalize this approach, and try to identify a multitude of media that would impose mutualism or commensalism between any two species? To address this question, we developed an algorithm for the search of interaction-inducing media (SIM), aimed at finding a large number of minimal or near-minimal media that are predicted to induce interactions between two given microbial species (see Methods). As a preliminary step for the SIM algorithm, we assemble a list of metabolites that are usable by at least one of the two organisms, based on the stoichiometric models. The algorithm then starts by assigning a single minimal set of metabolites (i.e. a growth medium) that allows both organisms (in their joint pair configuration, as in Fig. 2C,F) to grow with a rate that is above a given threshold (Table S2). This medium is minimal in the sense that removal of any one metabolite makes it impossible for the pair to grow. This minimal medium is also chosen so as to avoid (if possible) nutrients that contribute more than one essential element (e.g. avoiding amino acids, which can serve both as carbon and as nitrogen sources). Next, we perturb this initial medium by removing its carbon source metabolite, causing the modified medium not to sustain growth. We then identify a substitute metabolite (or metabolite set) that restores the capacity for growth (see Methods for a more detailed description). This process is repeated by iteratively removing each possible carbon-contributing metabolite found in the previous step, resulting in a set of feasible carbon sources (black squares in carbon source array in Fig. 3A). This perturbation loop is repeated for different elements (e.g. nitrogen, Fig. 3A). The arrays of feasible nutrients contributing different elements (C and N in the example of Fig. 3; C, N, P, S in the real calculations) are then used to construct a matrix of all possible combinations (Fig. 3A). Each of these combinations constitutes a medium that can putatively sustain growth of the organism pair. The next step is to test whether each of these media indeed sustains growth of the pair, and whether it can sustain growth of each species on its own (Fig. 3B). Even if a medium allows both species to grow, this does not imply that it will necessarily induce mutualistic growth. In fact, some of these media could simply be minimal media that can be used to grow both species individually (see Methods). Other media could be supporting commensal growth, i.e. allow one species to grow, and to produce a metabolite necessary for the other species to survive under the same conditions. Finally, some of the media may sustain growth of both species, without allowing either of the two individual species to grow on its own. Within these media, therefore, the two organisms would be able to survive only by exchanging essential metabolites, in an obligate syntrophic or mutualistic interaction. Computationally, testing for growth of the joint pair and of each individual species, allows us to easily classify the type of interaction induced by each medium (Fig. 3C). The details of the algorithm are described in the Methods section. Note that here we do not take into account the specific cost that an organism would incur to produce a metabolite that can benefit another organism. Hence, we do not distinguish, for example, between commensalism and parasitism (see Table 1).
We first applied SIM to the pairs of organisms presented in Figs. 1A and 1B. For the methanogenic pair of Desulfovibrio vulgaris and Methanococcus maripaludis (Fig. 1A), SIM identified only six simple media that can sustain growth of the joint organism pair. One of these six media corresponds to an experimentally tested environmental condition, and is the one imposed in the original model. It contains lactate, ammonia and di-hydrogen sulfide. Under this condition, D. vulgaris is predicted to utilize lactate and be able to grow on its own, while M. maripaludis is not, and can only grow in presence of the H2 and formate secreted by D. vulgaris. One aspect that the model does not capture explicitly at this point is the benefit that D. vulgaris receives from its association with M. maripaludis. This benefit is due to the fact that H2 consumption by M. maripaludis reduces the partial pressure of the gas, allowing D. vulgaris to keep producing H2 in a thermodynamically advantageous way. In addition to this canonical medium composition, our approach predicts five additional media that allow for growth of both organisms. One of these is predicted to induce another commensal interaction, in which D. vulgaris reduces sulfate to sulfide, which is then also shared with M. maripaludis. This interaction, however, may not be feasible, as there is some experimental evidence that suggests that sulfate reduction does not occur alongside methane production [14]. Interestingly, the remaining four media are predicted to induce obligate (thermodynamics-independent) syntrophy (Table 2). The mutualistic interactions arise because, according to the model, M. maripaludis is capable of fixing nitrogen to ammonium and extracting ammonium from alanine. Ammonium can then be utilized by D. vulgaris, which otherwise lacks the capacity to obtain it endogenously. These nitrogen-related interaction predictions are yet to be tested, but previous work has verified that Methanococcus is both able to fix nitrogen [41] and use alanine as a nitrogen source [42]. It is important to emphasize that we are considering here a rather small number of media, which do not include certain metabolites that D. vulgaris is known to be able to metabolize, such as pyruvate, ethanol, malate and fumarate [43], [44]. This is a consequence of the fact that the specific stoichiometric models available for these organisms are not genome-scale, but rather encompass only a subset of known metabolism pathways (approximately 100 reactions each).
In the case of the engineered yeast pair (Fig. 1B), SIM led to the prediction of a total of 36212 distinct media that allow for growth of both yeast strains in the joint model (Fig. S1). Of these, a total of 12981 media (35.8%) do not support growth of the individual strains (mutualism-inducing media), 12817 media (35.4%) can sustain growth of one of the organisms but not the other (commensalism-inducing), and 10414 (28.8%) can sustain growth of each strain individually (neutralism case). The computation of these media for the synthetic yeast pair offered us the opportunity to obtain more insight into how the algorithm performs, and into the biochemical rationale for patterns of interactions observed. Specifically, we compiled a metabolite-by-condition usage matrix M whose element Mij is equal to one if metabolite j is used in condition i, and zero otherwise, as predicted by SIM (Fig. 5 and Fig. S3). By clustering the columns (i.e. metabolites, see Methods for details) of the M matrix, it is apparent that some metabolites are required under all conditions, while other metabolites are not essential, being only required occasionally, often serving as alternatives to nearby metabolites (Fig. 5). Furthermore, one can separate the M matrix into sub-matrices pertaining to the four different classes of interactions (neutralism, mutualism, and the two commensal), and detect specific patterns of interaction that can be reconciled, e.g., with the biochemistry of the syntrophic pair (Fig. 5 and Fig. S3). One may also ask whether it is possible to discriminate between different types of interactions by performing an unsupervised clustering of the different media. Upon implementing a k-means clustering of neutral and mutualistic media, we found that the two sets of media do partition significantly more than random (p-val <10−300, see Methods), but in a way that is too weak to allow for a straightforward classification.
The interaction class specificity of certain metabolites is best highlighted in Figs. 5B, D, where, for each metabolite, we enumerate the total number of media per interaction class. This provides a comprehensive picture that complements the analysis of exchanged metabolites described above. In general, this type of representation may be useful in trying to design or prioritize media for experimental testing and applications, where one may want to give top preference to media that are mostly found in one class of interactions (e.g. syntrophic), and that are common to a large number of identified media (see Text S1). Different, biologically more relevant criteria for prioritizing metabolites (and consequently media that contain them) may be envisaged, for example based on the number of reactions each metabolite participates in (see Methods, Table S4).
The detailed analysis of the three test cases described so far indicates that the SEM and SIM algorithms are helpful in identifying true interactions and providing insight into the biochemical pathways underlying experimentally observed or putative interactions. As a first step towards novel synthetic ecology predictions, we studied the spectrum of possible interactions between E. coli and S. cerevisiae. This is motivated by the fact that the individual stoichiometric models for these two organisms are possibly the best curated and thoroughly tested experimentally [30], [45]–[47]. The E. coli - S. cerevisiae pair could be seen as a reference system for future experimental testing, as well as a good benchmark for performing sensitivity analyses of our algorithms. By applying the SIM algorithm to this pair we identified ∼11.6 million media (Fig. S6), out of which 4.7% are mutualism-inducing, 3.3% and 75.3% are commensalism-inducing (S. cerevisiae - E. coli and E. coli – S. cerevisiae respectively), and 16.8% are neutralism-inducing. In Table S4 we analyze in detail two solutions selected based on the prioritization criteria described above. Fig. 6 illustrates the results of multiple types of sensitivity analysis performed on the E. coli – S. cerevisiae pair. In particular, it can be seen that the interaction class predictions are highly robust with respect to individual perturbations that remove (Fig. 6A), add (Fig. 6B) or simultaneously add and remove (Fig. 6C) individual reactions from the stoichiometric models. Furthermore, robustness does not decrease significantly, on average, for at least ten cumulative gene addition/removal perturbations (Fig. 6D).
It is in principle possible to extend our computation of interactions to any organism for which a stoichiometric reconstruction is available. Here we present a computation of all possible pair-wise interactions between seven bacteria of relevance to health (first three) or environmental (last four) applications, namely Escherichia coli, Helicobacter pylori, Salmonella typhimurium, Bacillus subtilis, Shewanella oneidensis, Methylobacterium extorquens, and Methanosarcina barkeri. These specific organisms were chosen based on a balance between the following criteria: (i) they span wide spectra of function, environment and taxonomy; (ii) most of them have well characterized laboratory strains which could be used for future experimental testing; and (iii) they have publicly available stoichiometric models. While all stoichiometric models used here have undergone manual curation and some form of experimental validation, it is important to keep in mind that different models may have different levels of agreement with experimental observations. Thus, predicted interactions between arbitrary pairs should be evaluated in light of the expected fidelity of the corresponding individual stoichiometric models.
The results of this interaction analysis are summarized in the matrix of pie charts found in Fig. 7 (see also Table S1). The number of interaction-inducing media for a pair of organisms (size of each pie chart), as well as the proportions of different types of interactions, can vary significantly between pairs, from more than three hundred million for the E.coli - S. typhimurium pair, to only one in the H. pylori - M. barkeri case. At a first glance, one can identify several trends in the patterns of predicted inter-species interactions. For example, E. coli is predicted to be able to interact with most organisms in a large number of different environments. These interactions appear to be dominated by commensalism, where E. coli acts as the provider. This may be due to a combination of E. coli's ability to survive on a variety of carbon sources and of its capacity to export a number of byproducts. In contrast, H. pylori, except for the interaction with E. coli, has very few interaction-inducing media and frequently acts as a recipient, possibly reflecting its near obligatory parasitic nature [48]. Some organisms lie between these two extremes, e.g., the archaeon M. barkeri, which is predicted to interact with other species in a moderate number of environments, most of which induce commensal interactions that have M. barkeri as the recipient. In general, pairs of organisms appear to be dominated only by a few interaction classes. In addition, specific organisms tend to have an overall specific role in interactions with all species. For example, E. coli and B. subtilis are largely on the giving end of commensal interactions, while M. barkeri and H. pylori are mainly on the receiving end, and a sizable portion of M. extorquens and S. oneidensis interactions are mutualistic. Furthermore, some pairs display a large proportion of neutral interactions, i.e. can individually grow on a lot of common minimal media. This is most prominent for the E. coli and S. typhimurium pair, for which such an outcome may be expected, based on the fact that they tend to occupy a similar environmental niche. What might be less intuitive is that, for this same pair, we found also numerous environments that induce commensalism and mutualism.
For each pair of organisms it is possible to analyze in detail the environments identified, and use the SEM algorithm to determine what nutrients might be exchanged between the two organisms. For example, a mutualistic interaction between the bacterium E. coli and the archaeon M. barkeri is explored in detail in Fig. S5. This interaction is particularly interesting as a similar bacterium/archaeon pair may have been implicated in the rise of the primordial eukaryotic cell. For certain combinations of organism pair and growth media the SEM algorithm becomes computationally very heavy, and impractical. An alternative heuristic for identifying possible minimal sets of exchanged metabolites can be implemented in these cases (see Methods). We applied this alternative method to identify the exchanged metabolites in two cases of mutualism between S. cerevisiae and E. coli and two cases for the E. coli - H. pylori pair (Table S4). The relatively simple set of exchanged metabolites of the S. cerevisiae - E. coli pair provides novel experimentally testable biological hypotheses. The more complex metabolic exchange in the E. coli - H. pylori pair may reflect the parasitic nature of H. pylori.
Interestingly, almost all pairs in Fig. 7 appear to be potentially capable of engaging in mutualistic interactions (yellow portion of pie charts), given the appropriate growth medium. This is in sharp contrast with the relative paucity of neutralism-inducing conditions (Fig. 7, green portion of the pie charts). In general, the fact that neutralism-inducing interactions are rarely observed means that, upon exploring the space of next-to-minimal media that sustain a pair of organism, it is much more difficult to find media that sustain each organism on its own than it is to find media that support lifestyles involving unidirectional or bidirectional exchange. To some extent this may be expected, given that our algorithm searches for parsimonious solutions which guarantee growth of a pair of species. However, since it was not obvious a priori whether symbiotic interactions were at all possible, one may take these results as an indication that nutrient-poor environmental conditions are expected to be dominated by symbiotic interactions. Such a view point would offer the opportunity to use our approach as a quantitative modeling framework for partially understanding and estimating microbe unculturability in the wild. Moreover, it is possible to envisage simple simulations of the long-term dynamics of symbiosis based on estimates of the probabilities of transitions between different types of interactions upon environmental fluctuations. To exemplify this idea, we computed a matrix of transition probabilities for the E. coli – S. cerevisiae pair (Fig. 8A) (see Methods). A striking feature of this graph is the high transition probability between different states, suggesting a major and dynamical role of environmental fluctuations in determining microbial community lifestyles. In this case, the mutualistic state is quite unstable, possibly a consequence of the distinct environmental niches in which the two organisms belong. The fluidity of symbiosis under environmental perturbations is especially interesting in comparison with the corresponding graph for genetic perturbations, displaying a high degree of robustness, i.e. stability of individual states (Fig. 8B, extension of the data from Fig. 6).
Understanding natural and engineered microbial ecosystems is an ongoing challenge, relevant to multiple disciplines and applications. It is a challenge that undoubtedly requires a large computational effort, and novel algorithmic approaches. Our proposed method for identifying environments that induce symbiotic interactions is based on genome-scale stoichiometric models of metabolic networks, and can be in principle scaled up to computing interactions between any pair of organisms for which individual genome-scale constraint based models are available [49], [50]. In the current search for symbiosis-inducing environments we varied only metabolites containing carbon, nitrogen, phosphorus, and sulfur. Future extensions may explore larger chemical spaces, which could help identify interactions that are based on the exchange of other essential elements, including metal cofactors such as iron and magnesium. Furthermore, future models could be made more realistic by taking into account the role of fitness cost in determining the evolutionary advantage of different metabolite-sharing strategies.
While our implementation of the SIM algorithm to identify symbiosis-inducing media is currently limited to pairs of organisms, it can be easily extended to predicting media that induce symbiosis between triplets of species, or larger combinations. This follows from the fact that the framework defined in Fig. 2 can accommodate any number of different species, and that the algorithm would only need to check a longer list of options (e.g. all possible three-way interactions) relative to what presented in Fig. 3C. Extending the SEM algorithm for searching minimal sets of exchanged metabolites to three or more organism, however, would be more challenging, and one may need to further develop heuristics such as the one we used for the data in Table S4. In going from pairs to more complex communities of organisms known to populate a given environment, our algorithms would provide putative sets of interaction networks viable for such ecosystem. As of now the computation of the media for a given pair of microbial species takes on the order of 10 CPU hours per 10,000 media tested. This implies that systematic calculations for pairs or triplets between tens or hundreds of species will require high performance computing platforms and optimized algorithms. Given the current high pace of developments in these research areas, we envisage that this will be possible in the near future.
While increasingly detailed genome-scale stoichiometric models are being built and validated experimentally for several species [49], [50], one should keep in mind that these models constitute only coarse approximations of real biochemical complexity, which lack several layers, such as regulatory feedback and many thermodynamic constraints, and are often limited by our knowledge of gene function. Missing pathways, or wrongly annotated ones, as well as different levels of knowledge available for different organisms could bias our predictions, and could give rise to false positive or negative predictions of interactions. As seen in the engineered yeast strain test (Fig. 4, and Fig. S2), the direction of metabolite transport can play a role in defining possible exchanges between species, as can the number and specificity of annotated transporters and listing of non-enzymatic metabolite diffusion capabilities. However, examination of the relative number of transport reactions and their direction across all the species used in this study does not show any major bias (data not shown), suggesting that the different patterns of interactions observed in different organism pairs are not merely a consequence of major discrepancies in level of detail for different species.
In addition to transporters, any individual reaction within an organism's network could have an impact on possible interactions. Hence it would be important to know how sensitive our overall classification of media is to potential annotation errors in the models. We addressed this point by performing multiple types of sensitivity analyses on two different pairs of species (Fig. 6 and Fig. S4). The results of this analysis indicate that our interaction classification is largely robust to missing reactions in the model, suggesting that the patterns depicted in Fig. 7 roughly reflect true biological expectations. Even though our predictions are computationally robust, the ultimate way to determine the predictive capacity of our approach will be experimental testing. Similar to the cycles of refinement for individual genome scale stoichiometric models [49]–[51], we expect that the results from biological experimentation can feed back into refinement of models and algorithms for predicting interactions. For example, if a medium predicted to induce a mutualistic interaction is found experimentally to induce a commensal relationship instead, this would provide useful knowledge about metabolite uptake properties not accounted for in the model (Table S3).
Despite these potential limitations, our community models were successfully used to identify media that allow for mutualistic growth and identify metabolites that are being exchanged in three test cases. Running these processes on the toy system showed us that these algorithms give results that match our intuition and can be manually validated. By applying the algorithm to more complicated scenarios, such as an experimentally proven syntrophic yeast pair, we could recapitulate biological observations and predict additional experimentally testable environments under which similar mutualistic communities should arise.
Bridging across levels of description, one can draw an analogy between epistatic interactions in genetic networks and symbiotic interactions between microbes. The molecular composition of a sterile medium could be seen as an ecological unperturbed phenotype. It could be the case that a given microbial species is unable to grow on such medium. Hence, “perturbing” the environment by inoculating this microbe would give no observable change in the medium composition, i.e. no phenotype. This could happen equally for a second microbial species. However, if the two species are able to grow syntrophically in this environment, inoculation of both microbes (i.e. a double perturbation) will produce a major change in the medium, i.e. consumption of resources used for growth, and generation of novel byproducts. Thus, the combined effect of the microbes is highly synergistic relative to what is expected by observing the two microbes alone. This phenomenon is formally analogous to an extreme epistatic interaction between gene deletions, such as synthetic lethality [45]. Given the increasing amount of data and mathematical expertise generated in the study of genetic interaction networks, we envisage that valuable cross-fertilization will be possible between the field of synthetic ecology and the study of genetic networks.
Our results offer new insight on some evolutionary aspects of microbial ecosystems. Cryptic metabolic interactions may be the source of the unculturablity of many organisms. Since next-generation sequencing technologies can now produce the sequenced genomes of organisms we can not culture, it is possible to construct the metabolic network model for a species that has never been grown in a lab as pure culture. From this model one could infer what metabolites such organism would require from an interaction partner, paving the way for novel experimental testing. The mechanisms responsible for the evolutionary emergence of mutualistic interactions are an unresolved puzzle, though there is evidence that gene loss or horizontal gene transfer may drive some of these processes [52], [53]. Our results suggest an alternative mechanism, driven by environmental changes. Two organisms, initially growing independently of each other in a given environment, may be forced to become a commensal pair upon environmental depletion of a metabolite required by one of the species and producible by the other. A bidirectional co-dependency could ensue from a subsequent environmental change forcing a similar interaction in the opposite direction. Based on our prediction of nutrient sets that support symbiotic interactions, it may be possible to estimate the chances that a random walk in the space of environments will hit a mutualism-inducing or a commensalism-inducing one.
Another aspect that should be stressed is that many of the predicted symbiosis-inducing media may be theoretically feasible, but still not practically viable in a straightforward way, e.g. because the relevant metabolic pathways or necessary transporters may be down-regulated, allosterically inhibited or kinetically unfavorable. Such limitations, as well as potential limitations in metabolite transportability, might be overcome by implementing targeted (e.g. regulatory) mutations, or rounds of experimental evolution [52], [54], [55].
Finally, we envision that our proposed approach of a computationally driven synthetic ecology based on re-designing environments rather than organisms could have several applications. First of all, in analogy with synthetic biology, and as explored already with some artificial synthrophic species, the payback will be partially in terms of understanding interactions by building them. This may be seen as a first step towards building a stoichiometry-based microbial interactome, to help in the interpretation of metagenomic sequencing and microbial ecosystem data. Moreover, in terms of metabolic engineering, using the enzymatic potential of multiple interacting species can greatly expand the space of process optimization possibilities. Generating novel pathways by inducing interactions between different organisms rather than (or in addition to) genetically engineering the genomes of individual species has several benefits. First, one could use the metabolic potential of organisms that may be hard to genetically manipulate. Second, communities may be inherently more stable than individual modified species, in which specific mutations could potentially revert. We anticipate that methods like the ones we propose will be important in developing and analyzing synthetic communities of organisms. Our algorithms can be extended to simulate communities containing more than two organisms, predict gene knockouts that would give rise to mutualistic interactions and eventually entire consortia of microorganisms. Furthermore, the algorithms and methods we developed could be extended to study human health related problems. In addition to understanding interactions in the human microbiome, similar approaches could be used to ask how different cell types interact within a specific tissue and how a pathogen interacts metabolically with the host it infects.
Our algorithms use the framework of stoichiometric constraint-based models of metabolic networks, which have been described in detail elsewhere [56], [24], [57]. A stoichiometric matrix (S) is used to encode all the information about the topology and mass balance in a metabolic network, including the complete set of enzymatic and transport reactions in the system. Transport reactions, inferred from genome annotations, specialized prediction tools or literature curation, include both protein-catalyzed transport, e.g. ATP-driven transport, or ion-coupled symport/antiport, as well as free diffusion of small molecules (e.g. O2, CO2, etc.) through the cell boundaries. Element Sij represents the number of molecules of metabolite i participating in reaction j (with i = 1,…,m, and j = 1,…,n). The stoichiometric matrix S can be used as the starting point for efficiently generating predictions of metabolic rates (fluxes, vj, with units mM·(g dry mass)−1·(hr)−1) at a genome scale, e.g. using flux balance analysis (FBA) [56], [24]. FBA is generally based on two main simplifying assumptions. The first is a steady state assumption, which in matrix form can be expressed as Sv = 0. This assumption generates a large number of equality constraints that define the space of feasible metabolic states for the system. Further constraints (e.g. associated with reaction irreversibility of individual reactions) are imposed through inequalities of the type , where LBi and UBi constitute vectors of lower and upper bounds of reaction i respectively. These constraints will be later written concisely as . The second step of FBA is an optimization step, in which Linear Programming (LP) can be used to determine feasible flux distributions for some presumed cellular objective (c), subject to the previously described constraints. Typical objectives include maximization of biomass or ATP production, though different optimization approaches are used throughout this work, as illustrated below. Our implementation of flux balance models uses the GNU Linear Programming Kit (GLPK, http://www.gnu.org/software/glpk/) called from a Matlab shell through GLPKmex (http://glpkmex.sourceforge.net/).
The toy model is composed of two simple organism models (Fig. 2, and Fig. S1). Each sub-model contains transporters for three metabolites (X, Y and Z), one biochemical reaction (X→Y or X→Z) and one growth reaction (X+Y+Z→Biomass). Individually organism 1 (red in Fig. 2) can grow on metabolites X and Z, and produce metabolite Y. Organism 2 (blue in Fig. 2) can grow on metabolites X and Y, and produce metabolite Z. If organisms 1 and 2 where grown as a co-culture (i.e. sharing the same environment) they would only need metabolite X to be both able to grow.
We implemented a stoichiometric model for the methanogenic syntrophic pair constituted by D. vulgaris and M. maripaludis, as described in [33], and originally implemented on the FluxAnalyzer platform [58]. The original model was shown to produce valuable quantitative predictions about metabolic interactions between the two species. During growth on lactate, D. vulgaris produces H2, acetate and formate, all byproducts which M. maripaludis can utilize to grow and produce methane. Stolyar et al. [33] created biochemical models for each organism and joined them through an intermediate extracellular space. In transferring the model to our simulation platform we performed slight updates to the stoichiometric matrices, as detailed in the supplementary data file on our website, http://synthetic-ecology.bu.edu/. Our flux balance models involve 108 reactions for D. vulgaris and 103 for M. maripaludis. It is important to mention that these specific stoichiometric reconstructions are not genome-scale (as opposed to the ones used subsequently in our work), and only account for carbon, nitrogen, sulfur and hydrogen atoms.
To create the joint yeast model, we began with two copies of the newest yeast genome scale metabolic reconstruction iMM904 [40], and modified them to match the biological strains used in [19]. In one model, we identified and disabled the reactions associated with the gene Lys2, by setting the upper and lower flux bounds to zero. This modeled strain corresponds to the experimentally constructed Lys- strain, and is unable to grow on glucose minimal media without lysine as a supplement. Similarly, to model the Ade- biological strain we identified and disabled the reactions associated with Ade8 gene. This resulted in a model that required an adenine supplement to grow on a minimal medium. The strains in [19] had additional mutations that disabled the allosteric regulation of lysine and adenine pathways. As regulation is not represented in our constraint-based models, this aspect was left out. Unless otherwise noted, in the simulated minimal media we limit the amounts of glucose and O2, and allow free use of ammonia, sulfate, phosphate and salts.
The genome scale metabolic models for 5 of the organisms used have been published and are publicly available (Escherichia coli [59], Bacillus subtilis [60], Helicobacter pylori [61], Salmonella typhimurium [62], Methanosarcina barkeri [63]). For Shewanella oneidensis we used an early version of the recently published model [64], provided to us by Jennifer Reed. For Methylobacterium extorquens we used a genome scale extension of a previous reconstruction [65], provided to us by Steven Van Dien. Models were imported into Matlab from the XML files using the Cobra toolbox [25]. The metabolites of each model were manually checked for consistency across models.
Our approach for generating multi-species model extends the multi-species model employed by Stolyar et al., by introducing a fictitious compartment that represents the extracellular environment shared by both species, in addition to the original extracellular spaces for individual models. Our formulation, for two species 1 and 2 (assumed for this explanation not to have a periplasm) uses the following compartments (represented in square brackets, as in the standard notation used in [59], [40], [66], [67]): [CYT1] = cytoplasm for species 1, [EX1] = extracellular space of species 1, [CYT2] = cytoplasm of species 2, [EX2] = extracellular space of species 2, [ENV] = environment shared by species 1 and 2. For a metabolite Xi that can be exchanged between the environment and a given species (say species 1), we define the following reactions (where Y and Z are potential cofactors involved in transport across the cell membrane):
In previous formulations of joint models for different species or organelles [31]–[33] the extracellular spaces for the two interacting organisms (i.e. [EX1] and [EX2]) was collapsed into a single compartment, with no need for [ENV]). Here, by introducing an extra layer (and the extra shuttle reactions) we make it much easier to monitor what metabolites are being transported through the membrane. For example, if metabolite Xi is transported in and out of the cell through several different transporters, in order to know whether there is a net influx or efflux of Xi we would have to add up all the transporter fluxes. In our formulation, this is achieved simply by looking at the shuttle reaction flux. In a single species model, this would have been easily achieved by observing the exchange flux, but the extra degrees of freedom entailed by the multi-species model requires this extra layer of description. In addition, this formulation makes it much easier to implement our search algorithm for exchanged metabolites (SEM), in which we want to minimize the number of exchanged metabolites irrespective of the number of transporters available for each metabolite. The other important aspect of this distinction between exchange and shuttle fluxes is that, in terms of constraints, we have independent control on what molecules are environmentally available versus what molecules we want to make available for individual species. While this feature has not been used in the current work, it may be useful in future developments. The proposed formulation could serve as a standard way of building ecosystem-level stoichiometric models.
We have developed a mixed integer linear programming algorithm to identify a minimal set of possible exchanged metabolites between two organisms 1 and 2 that can grow simultaneously under a specified condition. Solving this problem requires imposing additional constraints to the regular mass balance and capacity constraints. First, since we require growth of both organisms, we fix the minimal growth rate of both organisms ( and ) to an arbitrary minimal amount ( = 0.1 in our simulations):A second constraint derives from the need to identify the set of metabolites that can be exchanged between the two species. This can be done by finding the intersection TM(1 and 2) between all the metabolites that are potentially transportable in the first and in the second model (metabolite sets TM(1) and TM(2) respectively). Each interchangeable metabolite i in TM(1 and 2) is associated with two shuttle reactions and importing the metabolite into individual species, and one exchange reaction , mediating its transport to the common environment. The condition of mutual exchange of metabolite i can then be expressed as the following constraint:where L is a large number (“infinite”), and θi is a binary variable which assumes a value of 1 if metabolite i is transported between the two species, and 0 otherwise.
Identifying a minimal set of exchanged metabolites amounts then to minimizing the sum of the θi variables over all metabolites. Overall, the optimization problem can be expressed as follows:In most cases, the joint flux balance model for two interacting species in a given medium can have multiple feasible flux solutions. Correspondingly, under a given growth condition there may be multiple equivalently minimal sets of exchanged metabolites. To address this degeneracy, we developed an algorithm that systematically identifies a large number of exchange metabolite sets. We reasoned that a set of exchanged metabolites will likely be a minimal set that allows for growth of both organisms. Hence, if we remove any one metabolite from the exchanged set, growth of both organisms is not possible without adding in at least one other metabolite into the set. Applying this algorithm in an iterative way allows us to identify multiple alternate exchange sets. At any given step, one metabolite is removed from the last solution, forcing the solver to find a substitute exchanged metabolite at the next iteration. This process is repeated until no more feasible solutions can be found.
For very large pairs of stoichiometric models, the SEM algorithm described above may be impractical. Therefore, we have devised a heuristic 2-steps alternative method that is more easily scalable to large systems. In the first step of this approach we solve a modified FBA problem for the joint pair of organisms using Linear Programming. Specifically, we minimize the sum of shuttle reactions fluxes that do not involve metabolites found in the current medium. This means that the search space is defined bywhere EM is the set of metabolites contained in the current growth medium. All constraints are the same as described for the SEM algorithm. As opposed to SEM, in this optimization problem we minimize the sum of the absolute values of the fluxes, hence removing any non productive (e.g. cycles) exchange of metabolites:This first step does not necessarily find a minimal set of metabolites that mediate the interactions, but rather one of many possible feasible set. As a second step, we can then apply the SEM algorithm where we limit TM to the metabolites found in the first step. The final set identified will be minimal (in the sense that removal of any metabolite will lead to infeasibility), but may not have a globally minimal count of exchanged metabolites, due to the intermediate step before SEM.
Here we describe the heuristic for identifying the set of media that support growth of multi-species co-cultures, and predicting the class of interaction they induce (see Fig. 3 for more details). After building a joint stoichiometric model as previously defined (Fig. 2), we identify an initial minimal medium (MM) that allows for positive growth rate of both organisms. In this work, we choose this initial medium manually, so as to select nutrients that are common to most organisms, and that constitute single element sources (e.g. do not contain both C and N). Our MM contained, for all pairs, succinate, ammonium, inorganic phosphate and sulfate, as well as oxygen and minerals. Then, in individual pairs, we included a minimal number of additional secondary metabolites (e.g., co-factors) as needed (see Table S2).
The core of the SIM algorithm is a function that identifies all possible metabolites (or sets of metabolites) that can substitute in the medium an initially available source for a given atom. In the current work we focus on identifying different sources for carbon, nitrogen, sulfur, and phosphate only. The analysis could be in principle extended to other atomic contributions, including cofactor metals, such as iron. The core function in SIM is recursive, and is best described, for a specific atom A, through the following pseudo-code (where CM is the Current Medium being evaluated) (see also Fig. 3):
<Initialize CM = MM
Initialize PM = {All environmental metabolites}\CM
Function Find_Replacement(A, CM, PM) {
Temporary Medium (TM) = CM\{molecules that contain A}
Replacement Sets (RS) = empty set of replacement metabolites.
X = Find X∈PM such that Vgrowth>0 on union(TM, {X}) (SMM algorithm, see below)
if ( X is not empty ) {
TM = union(TM, {X})
Remove X from PM
RS = Find_Replacement(A, TM, PM)
RS = [[X], RS ];
}
return RS
}
Note that the real algorithm (see Matlab scripts at http://synthetic-ecology.bu.edu/) takes into account the fact that it may not be possible for a single metabolite to substitute a previous one. In such case the function will continue searching for an additional metabolite that would allow nonzero growth. Hence, it may be the case that at a particular iteration, two molecules are compensating for the initially removed one. In the subsequent step, the algorithm will bifurcate, and try to remove each of these molecules individually.
In the next step, we generate a large set of possible media by determining all possible combinations of replacement metabolites for different atoms (Fig. 3A). We then check each predicted medium for growth of the joint and two individual models by applying FBA (Fig. 3B). Those media that allow for growth of the two organisms in the joint model, but not the individual models induce mutualistic growth through the exchange of metabolites. Media that sustain growth of only one individual organism, in addition to the pair, induce commensal interactions. Finally, those media that sustain growth of both individual organisms constitute cases of a neutral interaction (Fig. 3C, Table 1).
Here, we implement SIM based on a single initial MM. However, the specific choice of MM may influence the composition of the media predicted by SIM. To address this question, we implemented a sensitivity analysis, by focusing on a specific pair of organisms (yeast syntrophic pair), and recalculating the interaction-inducing media based on different choices of the initial carbon sources (glucose, pyruvate, acetate, fructose). Regardless of which initial medium was used, the same metabolites were identified as being viable carbon sources in the different trials. The identified media (i.e. combinations of the above metabolites) were almost identical (average 97.4% overlap±2%), regardless of the initial medium used.
A minimal medium is defined here as a set of metabolites that allows for a feasible solution with positive growth rate, and such that removal of any metabolite from the set would force the system to have no solution, or solutions with zero growth. To find the metabolites that belong to a minimal media, we implemented a mixed integer linear programing algorithm similar to what has been previously used in [68], [69]. As a first step we identify the set {vEX} of exchange reactions (labeled as (1) in the section Multi-species stoichiometric models above). We then solve a minimization problem which uses, in addition to the usual FBA constraints: (i) a constraint on minimal growth rates, as described for SEM () and (ii) a constraint expressing whether or not metabolite i is utilized (). Here, the binary variable θi assumes a value of 1 if metabolite i is transported between the two species, and 0 otherwise.
Identifying a minimal set of metabolites in a medium then amounts to minimizing the sum of the θi variables over all metabolites in {vEX}. Overall, the optimization problem can be expressed as follows:
To cluster the metabolites of all the media identified in the syntrophic yeast pair (Fig. 5 and Fig. S3), we compiled a metabolite-by-condition usage matrix M whose element Mij is equal to one if metabolite j is used in condition i, and zero otherwise. We clustered the columns (i.e. metabolites) of the M matrix, by implementing an average linkage hierarchical clustering using the Jaccard distance as a metric in Matlab. Alternate clustering methods gave equivalent results. Rows were clustered with the same method, but, for Figs. 5A, C, and Fig. S3A, C, we built separate clustering trees for each class of interactions. In addition, to determine whether media that induce different types of interactions tend to spontaneously segregate, we applied the same clustering algorithm to the combined set of neutralism and mutualism-inducing media. We next counted the number of interactions of each type that were called correctly using the clustering. We obtained: TP = 10513; TN = 7964; FP = 2450; FN = 2468; Hypergeometric p-val = 0; accuracy (TP+TN)/(P+N) = 0.790 (T = true; F = false; P = positive (in this case, mutualism); N = negative (in this case, neutralism). The high accuracy suggests that it is possible to roughly discriminate mutualism and neutralism cases. However, this accuracy does not extend to the case of all four interaction types (data not shown).
For the S. cerevisiae and E. coli pair, 1000 media were chosen for each interaction class at random. For each of these media the set of metabolites was perturbed and the pair retested for interaction class 100 times. This was done by selecting a carbon containing metabolite, and replacing it with another carbon containing metabolite at random (but still allowing growth of the organisms in the joint model). The transition probabilities fore each interaction class were then calculated as the mean fraction of times a given interaction class is transformed into another interaction class upon the perturbation (Fig. 8A).
For the S. cerevisiae and E. coli pair, 1000 media were chosen at random. For each of these media, three types of reaction perturbations were performed 100 times. The first type of reaction perturbation consists of the deletion of a reaction at random from the joint model. The second type of reaction perturbation is implemented by adding a reaction at random to the joint model. The third type of perturbation corresponds to the simultaneous deletion of a reaction and addition of another reaction at random. The transition probabilities between any two interaction classes A and B were calculated as the mean fraction of times interaction class A became interaction class B after perturbation (Fig. 6A,B,C). In order to study the effects of multiple (k, ranging from 1 to 10) insults, we extended the perturbation analysis by randomly selecting and applying k random insertions or deletions 1000 times, and counting the number of times the interaction class change. The randomly added reactions were taken from the subset of KEGG reactions [70] involving metabolites present in the joint model.
|
10.1371/journal.pcbi.1002742 | Modeling Protective Anti-Tumor Immunity via Preventative Cancer Vaccines Using a Hybrid Agent-based and Delay Differential Equation Approach | A next generation approach to cancer envisions developing preventative vaccinations to stimulate a person's immune cells, particularly cytotoxic T lymphocytes (CTLs), to eliminate incipient tumors before clinical detection. The purpose of our study is to quantitatively assess whether such an approach would be feasible, and if so, how many anti-cancer CTLs would have to be primed against tumor antigen to provide significant protection. To understand the relevant dynamics, we develop a two-compartment model of tumor-immune interactions at the tumor site and the draining lymph node. We model interactions at the tumor site using an agent-based model (ABM) and dynamics in the lymph node using a system of delay differential equations (DDEs). We combine the models into a hybrid ABM-DDE system and investigate dynamics over a wide range of parameters, including cell proliferation rates, tumor antigenicity, CTL recruitment times, and initial memory CTL populations. Our results indicate that an anti-cancer memory CTL pool of 3% or less can successfully eradicate a tumor population over a wide range of model parameters, implying that a vaccination approach is feasible. In addition, sensitivity analysis of our model reveals conditions that will result in rapid tumor destruction, oscillation, and polynomial rather than exponential decline in the tumor population due to tumor geometry.
| An innovative approach to treating cancer envisions developing preventative anti-cancer vaccines to train a person's immune cells to eliminate early-stage tumors close to genesis. The design of such a treatment strategy requires an understanding of the tumor and immune interactions leading to a successful anti-cancer immune response. To engage this problem, we formulate a mathematical model of the immune response against incipient tumours consisting of as low as hundreds to thousands of cancer cells, which is far below the clinical detection threshold of over 100,000 cells. The model considers the initial stimulation of the immune response and the resulting immune attack on the tumor mass and is formulated as a hybrid agent-based and delay differential equation model. We apply the model to test dynamics over a wide range of dynamic parameters, including immune and tumor cell growth rates and the size of the initial anti-cancer immune population. Our results show that an anti-cancer memory immune cell population of 3% or less can successfully eradicate an incipient tumor population over a wide range of dynamic parameters, indicating that a vaccination approach is feasible.
| The most effective way to treat a disease is to prevent its development in the first place. Consequently, a next generation approach to cancer treatment envisions developing preventative cancer vaccines that would train a person's immune response to eliminate tumors near inception by stimulating a person's immune system, especially cytotoxic T lymphocytes (CTLs), to attack cancer cells expressing tumor-associated antigens [1]. Such an immune response would destroy developing tumors close to genesis, before tumor cells have acquired the ability to suppress immune responses or metastasize to other tissues. A successful preventative cancer vaccine would revolutionize the approach to cancer treatment, and several experimental studies have successfully induced CTL responses against different types of tumor cells [2]–[5].
A number of important questions need to be addressed. In particular, is it a realistic goal to immunize a person against cancer, and if so, how many anti-cancer CTLs would be required to provide significant protection against cancer development? There are several conceivable obstacles that could hinder a memory anti-tumor CTL response from being effective. Since cancers develop from colonies of several cells and grow much more gradually than most infectious diseases, developing tumors will only produce a weak antigenic signal, resulting in the activation of only a small fraction of antigen-specific CTLs. Furthermore, activated CTLs will have to encounter the incipient tumor mass in the midst of a large volume of surrounding tissue. It is conceivable that these effects could render an anti-tumor CTL response ineffective. Consequently, the aim of this paper is to assess the feasibility of preventative cancer vaccines from a quantitative perspective.
A challenge to designing effective vaccines will be to understand the quantitative dynamics of the protective anti-tumor CTL response that initiates in the lymph node and proceeds to the tissue containing the tumor. CTL responses almost always begin in lymph nodes rather than the affected tissue. In particular, unactivated CTLs spend most of the time circulating through lymph nodes until they are stimulated by antigen-presenting cells, at which point they proliferate and migrate to the affected tissue [6].
To model this system, we synthesize current experimental research of CTL dynamics into a hybrid mathematical model consisting of a system of delay differential equations (DDEs) and an agent-based model (ABM). Using this hybrid framework, our model connects the fast-timescale dynamics of immune interactions within lymph nodes with the probabilistic, slow-timescale dynamics of immune surveillance in the tumor microenvironment. We then apply the model to investigate rates of tumor elimination under a wide range of parameters, including tumor and CTL proliferation rates, tumor antigenicity, CTL recruitment rates, and initial CTL populations. In addition, the model sheds light on the scale and nature of the dynamics relevant to an immune response against a clinically undetectable, localized microtumor.
Mathematical modeling of tumor growth and tumor immunology has grown rapidly in recent years and several modeling approaches have been applied to understanding these phenomena. For example, a large body of tumor-immune models have been developed using ordinary differential equations (ODEs) [7]–[14] and partial differential equations (PDEs) [15]. (See also [16] for a review of ODE models of tumor-immune interactions and [17], [18] for reviews of ODE and PDE models of tumor growth.) Another approach has focused on agent-based (or cellular automata) models, sometimes coupled with differential equations, to simulate tumor growth [19], [20], tumor growth with angiogenesis [21], [22], and tumor growth in the presence of an immune response [23]–[25]. These models focus primarily on chemotherapy, immunotherapy, and other treatments operating against existing tumors following clinical detection than on protective immunity against undetectable, developing tumors. In our model, we focus on protective anti-tumor immunity by anti-tumor memory CTLs that would be generated by a preventative cancer vaccine.
The ABM-DDE system we develop in the following sections is most similar in formulation to the hybrid cellular automata-PDE model of [24], which also considers host immune responses against growing tumors. However, our model differs from that of [24] in that it simulates interactions in three dimensions rather than two, cell motion and cell contacts take place in Euclidean 3-space rather than on a lattice, and the model simulates two compartments that account for the communication between the tumor site and the lymph node. On the other hand, since the agent-based and cellular automata components of the two models are comparable and simulate tumor populations of similar orders of magnitude (fewer than 100,000 cells), we can readily compare the results and parameter sensitivity analysis of our model with those of [24] as we do in the Results section.
Although our model is formulated in a way that could apply to multiple types of tumors by modifying parameters, a large body of experimental research has been directed toward developing treatment strategies for breast cancer, particularly by identifying potential antigens that could be targeted by preventative breast cancer vaccines [2], [5]. In addition, key model parameters, such as tumor growth rates, are readily available for breast cancer, e.g., [26]–[29], so for the purposes of focusing the scope of our model formulation and parameter sensitivity analysis, we estimate tumor parameters using breast cancer data.
The paper is organized as follows. In the Results, we discuss the results of model simulations. In particular, we show plots of example simulations, conduct a parameter sensitivity analysis, and discuss the conditions under which the tumor population in the ABM could exhibit a polynomial rate of decline, rather than exponential, due to killing by CTLs. In the Discussion, we discuss several natural extensions of the model and directions for future work. In the Models, we present a two-compartment model, consisting of the ABM of the tumor site and the DDE model of the lymph node. We also justify the parameter estimates.
The hybrid ABM-DDE model was simulated using Matlab R2011b. Results from an example simulation are shown in Figures 1 and 2.
As shown in Figure 1(right), the tumor begins growing in the periphery. As the tumor size reaches approximately 1,000 cells, more and more immature APCs in the periphery become mature, begin presenting tumor antigen, and migrate to the lymph node (see Figure 1(left)). The presence of mature, tumor-antigen-bearing APCs in the lymph node causes memory CTLs to activate into effector CTLs. These effector CTLs proliferate and migrate to the periphery, leading to an anti-tumor CTL response at the tumor site (see Figure 1(right)). Note that only a small fraction of immature APCs and memory CTLs become stimulated into mature APCs and effector CTLs, respectively, so the populations and remain almost constant throughout the simulation.
Figure 2 shows snapshots of the ABM simulation at various time points of the CTL response. The tumor begins to grow from one cell at time 0. By day 14, the tumor has grown to 1,714 cells and the anti-tumor CTL response has increased enough so that anti-tumor CTLs begin to circulate around the tumor site at a concentration of . By day 20, several CTLs have engaged the tumor, giving rise to recruitment of additional CTLs. By day 22, the anti-tumor CTL response has overcome tumor growth causing the tumor cell population to decline. By day 36, the tumor has shrunken to 191 cells, and anti-tumor CTLs eliminate all tumor cells on day 42.
Since the system is probabilistic, each simulation produces different results even when the underlying parameters are kept constant. For example, a CTL response will not always eliminate a tumor in one attempt. Indeed, when tumors decrease to tens of cells or fewer, a moderate chance exists that all the CTLs in the vicinity of the shrinking tumor mass may die or migrate away, allowing the residual tumor to relapse. This phenomenon can happen under any set of parameters, but happens more frequently when the average time for CTL recruitment, , is high (see Figure 3). In Figure 3, the slow rate of CTL recruitment to the tumor site allows the tumor to survive and relapse 11 times. Nonetheless, the memory CTL response keeps the tumor population below 5,700 cells.
These results are corroborated by the cellular automata results of [24], in which Mallet and de Pillis observe that a relatively high CTL recruitment rate leads to few oscillations in the tumor population and early tumor elimination, whereas a lower CTL recruitment rate gives rise to ongoing oscillations, during which the tumor is nearly eliminated at several points, but manages to relapse.
To obtain a broader view of the influence of parameter values on the behavior of the system, we analyze the sensitivity of the model to the following eight parameters: , , , , , , , , . We conduct our sensitivity analysis by varying each parameter individually over the ranges shown in Table 1, while holding all other parameters constant at their base values. For each set of parameters, we conduct 5 simulations. Due to the computational cost of the ABM, we do not conduct more simulations, but even with such few repetitions, we can observe key trends in the influence of the parameters on the model. To assess the influence of the parameters, we calculate the Spearman rank-order correlation of each parameter versus the time to tumor extinction and the maximum number of tumor cells. Table 2 shows Spearman rank-order correlations, , and -values for each parameter.
One parameter that stands out as being remarkably insignificant to the final outcome of the simulations is the time-delay parameter, , representing the duration of one CTL division. This parameter has almost no correlation to both the time of tumor extinction and the maximum tumor population. One reason for this lack of significance is that over the entire range of , the duration of the CTL division program, , varies from 2.5 to 10 days, while typical tumor extinction times are on the order of 100 days. Hence, a variable delay of several days hardly impacts the final outcome. Similarly, the maximum tumor population will only be minimally affected by a slight delay in the initiation of the CTL response.
Although a natural extension of DDE system is to consider a distributed instead of a discrete time delay for the duration of CTL division, the sensitivity results above imply that a variable delay over the range 4 to 24 hours will probably hardly affect the final outcomes. Moreover, the vast majority of CTLs are likely to have division times within this range [6], [30]
From Table 2, we see that the outcomes of the simulations are most significantly influenced by the average tumor division time, ; average CTL recruitment time, ; initial number of CTL divisions upon activation, ; and the antigenicity of the tumor, . Figure 4 plots the outcomes of the simulations with respect to , , and .
In column 1 of Figure 4, we see that the time to tumor extinction grows almost linearly with respect to the average tumor division time, . On the other hand, for tumor division times of greater than 10 days, the maximum tumor population hardly changes and nearly all tumors are destroyed at populations of fewer than 1,000 cells. The reason is that the CTL response begins to respond to tumors once they reach a certain size (approximately several hundred cells). A more slowly growing tumor will take proportionally longer to reach this critical size at which CTLs respond. Interestingly, even when an incipient tumor divides at a very rapid rate of once per day, the CTL response destroys the tumor in under 100 days at populations of on the order of 10,000 cells. These results suggest that the immune system responds more effectively to quickly growing tumors. However, very quickly growing tumors can attain orders of magnitude higher populations before being destroyed (see Figure 4 (column 1, bottom row)). As a result, these tumors may grow large and diverse enough to develop immune evasion and metastatic capabilities before the CTL response can eliminate them. On the other hand, tumors that grow very slowly could persist for several years before being detected by CTLs. Consequently, it seems that incipient tumors that grow very quickly or very slowly could cause the most difficulty for an anti-tumor CTL response.
In column 2 of Figure 4, we see that both the time to tumor extinction and the maximum tumor population increase steadily as the average CTL recruitment time, , increases. Nonetheless, the maximum tumor population increases by less than an order of magnitude over the range . On the other hand, the variance of tumor extinction times seems to increase suddenly once passes 18 hours. This sudden shift is probably due to the increased chance of tumor survival and relapse leading to oscillations when CTL recruitment become sufficiently slow (for example, see Figure 3). This phenomenon may be akin to a Hopf bifurcation in dynamical systems. A useful future direction would be to devise an analogous version of the model as a dynamical system to analyze whether Hopf bifurcations could underlie this and other shifts in the behavior of the ABM-DDE system.
In Table 2, we also see that although the system is sensitive to the CTL recruitment time, it is much less sensitive to the time for CTLs to kill tumor cells. This result is reasonable, since CTLs that ineffectively recruit additional CTLs are unlikely to eliminate the tumor during their lifespans regardless of their killing rate.
In column 3 of Figure 4, we see that tumor antigenicity, , influences the behavior of the system the most. Indeed, a tumor that is 10 times less antigenic than another would require a tenfold higher tumor cell population to elicit a CTL response of the same magnitude. Nonetheless, over the entire simulated range of antigenicities, the CTL response succeeds in destroying the tumor in under than 300 days and at populations below 30,000 cells, corresponding to tumors of less than 0.35 mm in diameter, which is still under the typical clinical detection limit of a few millimeters or greater [31], [32]. Therefore, although it is difficult to estimate the level of antigenicity of an incipient tumor, it appears that an anti-tumor memory CTL response could be reasonably effective for a wide range of tumor antigenicities.
In Table 2, we see that the outcomes of the simulations also depend significantly on , the number of divisions of memory CTLs upon activation. The plots for simulation outcomes versus resemble those for in column 3 of Figure 4, so we do not show them here. The strong dependence of the system on is expected, because each increase or decrease in coincides with a twofold increase or decrease in the magnitude of the CTL response. Nonetheless, over the range , corresponding to a 100 to 100,000-fold CTL expansion upon activation, the maximum tumor populations remain between 5,000 and 100 cells, and extinction times remain under 200 days for all simulations.
We also consider the sensitivity of the system to the size of the memory CTL pool, since this parameter will inform the development of a preventative breast-cancer vaccination strategy. Figure 5 shows simulation outcomes against the steady-state frequency of anti-tumor memory CTLs. From the figure, we see that the time to tumor extinction only decreases slightly as the memory CTL population increases. On the other hand, the maximum tumor population decreases approximately threefold as the memory CTL population rises from 1 to 3% and then stabilizes for memory CTL populations from 3 to 10%. Based on this result, a preventative vaccination strategy would maximize its potential efficacy by generating a memory CTL pool of around 2 to 3% of the steady state CTL population. A target memory pool of that size may be attainable by a strategic use of cellular vaccines and adjuvants [33].
The results above indicate that over the range of parameter values considered in Table 1, the presence of anti-tumor memory CTLs effectively restricts the maximum growth and longevity of an incipient tumor, perhaps even to the point of preventing it from diversifying and exhibiting immunosuppressive or metastatic behaviors. As we see in Table 1 most of the varied parameters were considered over a range of at least 1/2 to 2 times the estimated value. The only parameters that were not varied over this wide of a range were the CTL diffusion parameter, , and the minimum number of CTL divisions, , which vary over a range of approximately of the estimated value. The ranges of these parameters were tightened, since estimates were based on direct experimental measurements of the required quantities [34]–[36]. In addition, the mass-action coefficient, , was varied from 1/10 to 1 times the estimated value, since higher values of would only make the CTL response stronger and more effective. Consequently, we decide to only consider lower values of this parameter to see whether weaker CTL responses can result in a favorable outcome. From these results, the model suggests that pursuing the development of breast cancer vaccines that would boost immune defenses against incipient tumors may be a feasible preventative treatment strategy over a wide range of parameter values.
In the results above, we do not discuss the probability of tumor survival, because the tumor is always eventually eliminated in our simulations. In fact, for the model as formulated, eventual tumor elimination is highly likely and perhaps guaranteed on an infinite-time horizon.
Several reasons why the model formulation makes eventual tumor elimination very likely are as follows. The modeled tumor site is a finite volume of , and tumor cells cannot grow beyond this region. CTLs are continually supplied from a regenerating memory population, so CTLs never go extinct. On the other hand, the tumor-free state is an absorbing state from which no new tumor cells can be generated. Over the range of considered model parameters, CTLs proliferate and recruit additional CTLs at a faster rate () than tumor cells divide (), so once CTLs engage tumor cells, the CTL population can always exceed or keep abreast of tumor proliferation. These parameter assumptions seem reasonable, since the expansion rate of proliferating memory CTLs will most likely exceed the growth rate of tumor cells (see the discussion and references in the Parameter Estimates section).
In addition to these considerations, the probabilistic nature of the model implies that there is always a nonzero chance that even a few CTLs can kill a large number of tumor cells rapidly, meaning that a chain of strongly cytotoxic events could lead to complete tumor elimination even in unlikely circumstances. Consequently, over an infinite-time horizon, tumor elimination becomes more likely and perhaps even inevitable. For example, as we see in Figure 3, the underlying dynamics appear to be oscillatory. However, every oscillation increases the chance that the tumor could be eliminated, which occurs on day 1,600 in the displayed simulation.
Due to these limitations in the model, we choose to assess the simulations based on time to tumor elimination and maximum tumor population rather than probability of elimination. Our focus is whether an anti-tumor CTL response can eradicate an incipient tumor quickly and below a certain size, instead of whether it can eventually eliminate the tumor. The reason we are interested in a quick and decisive immune response is that tumors that grow for a long time or to a large size most likely have a higher probability of avoiding immune elimination by either metastasis and migration away from the primary tumor site or by mutating to develop immunosuppressive or immune evasive capabilities. At this point, we do not explicitly model tumor metastasis or adaptive mutation, so these aspects remain a key direction for future work.
As a substitute to directly measuring the probability of tumor elimination, one can set a criterion for failure of the immune response and reinterpret the results. For example, a possible criterion could be that the incipient tumor must be eliminated in fewer than 10,000 tumor cells ( diameter) and in less than 2 years. However, the probability of tumor elimination still follows the same trends shown in Figures 4 and 5, so we do not display additional results under this criterion.
In the simulations above, the tumor decline to extinction appears to follow a curve of the form , rather than an exponential decay, where approximately coincides with the simulated extinction time and is constant. An interesting observation is that, unlike an exponential decay, the cubic curve reaches 0 in finite time, meaning that the descent to tumor elimination proceeds almost deterministically, even though the model is probabilistic.
Figure 6(left) shows the time plots of the tumor and CTL populations from a simulation of the ABM-DDE system, where the tumor antigenicity and all other parameters are taken from the base values shown in Table 1. Figure 6(right) shows plots of and from day 190 to extinction, where is the tumor population corresponding to the agent-based simulation of the tumor site.
From Figure 6(right), we see that declines almost linearly to the extinction time, whereas is far from linear, indicating that the tumor decline much more closely follows a cubic function than an exponential decay. The linear regression for is . (The plot of the fit is not shown, because it overlaps the curve very closely.)
The cubic curve predicts a deterministic finite time extinction of the tumor at time , which is very close to the simulated extinction time of 213 days. The cubic decline can be explained by considering the geometry of the system. If we assume that the growing tumor mass is approximately spherical, most CTLs will engage tumor cells on the surface of the sphere. This is not to say that some CTLs will not penetrate the tumor, even causing the tumor to fragment and lose its spherical shape. Indeed, fragmentation happens more frequently as the CTL recruitment rate decreases, causing the model to depart from a strictly cubic decline. However, if for the most part, the majority of CTLs encounter and engage the tumor near its surface, the rate that tumor cells are killed by CTLs will be proportional to . A system that will yield a cubic solution like the one above is a differential equation of the form , where the coefficient is proportional to how thoroughly CTLs cover the tumor surface and is a function of the total CTL population, .
This observation implies that tumor-CTL dynamics, at least during the decline phase could be modeled by a system of differential equations that predicts deterministic extinction in finite time. In fact, in a different study, a deterministic ODE model of cancer virotherapy was formulated that predicts cancer elimination in finite time [37]. At this point, we leave a more thorough development of a deterministic differential equation model for a future work, but for now, we present the following simple ODE model:(1)where is the number of tumor cells and is the number CTLs that are close enough to engage the tumor at its surface. In addition, is the tumor volume, where is the cell radius given in Table 1, is the tumor surface area, is the maximum number of CTLs that can be in contact with the tumor surface at the current time, and is the density-dependent CTL immigration term.
The first equation in (1) pertains to the number of tumor cells. The first term is the growth rate of the tumor mass. We assume that the growth rate is proportional to the surface area of the tumor, since nearly all growth will happen at or near the surface of the tumor. The second term is the rate that tumor cells are killed by CTLs. The rate of tumor death is proportional to the number of CTLs in contact with the tumor. We assume that all CTLs in the close vicinity of the tumor are in contact with the tumor up to a maximum number . This maximum is the ratio of the surface area of the tumor divided by the cross-sectional area of a CTL.
The second equation in (1) pertains to CTLs in close vicinity of the tumor. The first term is the rate at which CTLs in the periphery come into the close vicinity of the tumor. Since the ODE does not account for CTL diffusion, we assume all CTLs are evenly distributed throughout the periphery at concentration , where is the CTL concentration in the periphery given by the DDE model (3). Since is in units of thousands of cells per mm3, the factor is the number of CTLs that would occupy a region of volume . The density-dependent term ensures that the rate CTLs come into the vicinity of the tumor decreases to 0 as the population approaches capacity . The second term is the death rate of CTLs and the parameter is the same as the one in Table 1. The third term is the rate at which CTLs in the vicinity of the tumor recruit additional CTLs to the vicinity of the tumor. This term is also modified by the density-dependent factor to ensure that the CTL recruitment rate decreases to 0 as the population approaches capacity.
As with the ABM, the ODE system (1) for the tumor site is coupled with the DDE system (3) for the lymph node. We simulated the combined system using ‘dde23’ in Matlab R2011b. Figure 7(left) shows numerical simulations of the tumor and CTL populations given by (1). Figure7(right) shows a time plot of the cube root of the tumor population obtained from the numerical simulation.
From Figure 7(left), we see that the rise and fall curves of the tumor and CTL populations exhibit similar shapes as those of the ABM-DDE simulation in Figure 6(left). In addition, from Figure 7(right), we see that the final decline of the cube root of the tumor population, , closely follows a linear decline and deterministically reaches extinction in finite time on day 19.04.
These observations suggest that the rate of CTL killing of the tumor resembles a differential equation of the form for , rather than a mass-action model given by . The reason the tumor death rate is proportional to for is that not all tumor cells are equally accessed by CTLs due to the geometric structure of the tumor. Because , a differential equation model of CTL-tumor dynamics could predict deterministic finite-time extinction of the tumor, and as we see in Figure 6, the dynamics of the ABM could closely follow this deterministic decline to extinction.
We formulate a model of an anti-tumor memory CTL response elicited by vaccination that will act against an incipient tumor. The primary goal of the model is to assess whether it is realistic for a person's immune system to have a sufficient pool of anti-cancer memory CTLs to significantly reduce the chances of developing cancer. We focus our investigation on breast cancer, since extensive experimental research has been done on growth parameters and tumor sizes, e.g., in [26]–[29] and clinical detection limits, e.g., in [31], [32].
Our model suggests that protective immunity against the development of breast cancer could be feasible, because an anti-tumor memory CTL pool of 3% of CTLs could eliminate a developing tumor before it reaches an average size of 1,000 cells, and an anti-tumor memory CTL pool of only 1% of CTLs could eliminate a growing tumor in fewer than 30,000 cells (a diameter of approximately 0.35 mm). These predictions are corroborated by experimental results. In one mouse study, vaccination with telomerase led to telomerase-specific T cell responses of no more than 3% in different mouse strains and had a protective effect against tumor growth [38], and another mouse model showed that a 2% threshold for a vaccine-elicited T cell response predicted efficacy in limiting tumor growth and survival [39].
In our simulations, the effectiveness of the anti-tumor CTL response depends largely on how quickly CTLs can locate and then eliminate an incipient tumor. The key challenges to locating the tumor are that the incipient tumor expresses a very low antigenic signal to the draining lymph node and it takes up a tiny volume in the tissue. The rapidity of this phase depends primarily on the number of CTLs that become activated and migrate to the periphery. These dynamics are governed mostly by , the equilibrium memory CTL population; , the number of divisions undertaken by an activated CTL; and , the antigenicity of the tumor. Once the tumor has been located and CTLs begin to engage tumor cells, the survival of the tumor depends mostly on , the rate additional CTLs are recruited to the tumor site and, to a slightly lesser degree, , the rate at which CTLs kill tumor cells. A future step for experimental and modeling research will be to understand how to design an optimal vaccination strategy that would elicit a sufficient CTL response to seed an adequate anti-tumor memory CTL pool [33].
An additional observation from our simulations is that faster growing tumors are often destroyed faster than more slowly growing ones. This result agrees with experimental observations that CTL responses react more effectively to rapidly increasing sources of antigen than to constant or slowly increasing stimuli [40]. In other words, for a protective immune response, a population of rapidly growing tumor cells might not be more difficult for the immune response to eliminate than a very slowly growing population.
In our current study, we are interested conditions that allow the CTL response to eliminate a tumor before it reaches a sufficient size or diversity to effectively suppress the immune response, metastasize, or induce angiogenesis. The question remains: At what size or in what time frame is a tumor likely to develop these capabilities, and how would this development impact the immune response. Thus, a direction for future work would be to incorporate the mutation of tumors cells to model the competition between the CTL response and the evolution of the tumor cell population.
Another extension of the model is to explicitly incorporate the chemotaxis of CTLs up a signal gradient to the tumor site. We currently model CTL recruitment using an approach analogous to that of [24], in which new CTLs arrive at a probabilistic rate in the vicinity of recruiting CTLs. In reality, CTLs migrate up a signal gradient toward a region of high cytotoxic activity. However, this process appears to happen much more quickly than the time scale of the CTL response simulated in the model [3]. As a result, in this study, we do not explicitly model the trajectory of recruited CTLs toward the tumor mass. Indeed, if CTLs move at an average rate of , and the radius of the simulated tumor site is , a migrating CTL could travel from the boundary of the region to the center in less than an hour.
As discussed at the beginning of this paper, various models have recently been developed for immune interactions with solid tumors, using both probabilistic agent-based (or cellular automata), deterministic differential equation, and hybrid approaches, e.g., [7], [8], [12], [23], [24]. A future step will be to bridge these frameworks. In the case of our ABM, we noticed that the decline in the tumor population could closely follow a cubic rather than an exponential curve. Consequently, a deterministic differential equation version of the ABM would have to account for the tumor geometry as well as cell localization around the tumor. Developing a differential equation version of the ABM will provide a means of analyzing the stability of the system, particularly around the tumor-free fixed point and determining what conditions allow the tumor to be eliminated in finite time, see [37]. In addition, a differential equation model will shed light on whether a stability bifurcation underlies the rapid increase in the amplitude of oscillations that occur as the CTL recruitment time increases (see Figure 4(column 2)).
Characterizing tumor-immune dynamics using different modeling perspectives will provide a means of assessing whether it would be feasible to prevent breast cancer using preventative vaccines. Since nearly all relevant cell interactions for protective anti-tumor immunity occur at a level below clinical detection, insights provided by models of immune responses against developing tumors will inform further modeling and experimental directions and aid the advancement of next-generation therapeutic strategies.
Our model considers two compartments of immune activity: the site of the incipient tumor in the tissue and a tumor-draining lymph node. We model dynamics of the tumor compartment using a probabilistic ABM. The advantage of the ABM is that it allows us to capture the probabilistic nature and spatial structure of CTL-tumor interactions. In our simulations, cell populations fall under 100,000, making an ABM computationally practical.
On the other hand, we model dynamics in the lymph node using a system of DDEs. The advantage of the DDE system is that it allows us to capture the dynamics of an arbitrary number of cells efficiently. In the lymph node, immune cells interact at a faster time scale and exist at orders of magnitude higher concentrations than in the periphery, making an ABM formulation computationally impractical. As a result, we devise a hybrid model connecting an ABM and a DDE system for the tumor site and lymph node.
The ABM simulates tumor cells and CTLs at the tumor site. All cells are modeled as spheres of radius in Euclidean 3-space, and no two cells can overlap the same space. The system is updated according to algorithmic rules at discrete time steps . The rules for each type of cell are described below.
To investigate the possible strength of a secondary anti-cancer CTL response, we simulate the anti-cancer immune dynamics in a vaccinated host. The anti-tumor CTL response begins when antigen presenting cells (APCs) bearing tumor antigen mature and migrate to the draining lymph node, where they activate memory CTLs that begin to proliferate and emigrate to the site of infection. We model this process in five steps illustrated in Figure 13:
The model is formulated as the following system of DDEs:(3)where is the tumor cell population at the tumor site modeled by the ABM, is the concentration of APCs in the periphery, is the concentration of APCs that have matured, started to present tumor antigen, and migrated to the lymph node, is the concentration of memory CTLs in the lymph node, is the concentration of effector CTLs in the lymph node, and is the concentration of effector CTLs in the periphery. (The concentration is the value used by the ABM to generate CTLs in the CTL cloud.) Concentrations are measured in units of (thousands of cells per cubic millimeter). Note that = (microliter).
The first equation in (3) pertains to APCs waiting in the periphery. These cells are supplied at a constant rate, , and die at a proportional rate, . Thus, without stimulation, the population remains at its equilibrium level, . The factor is the proportional rate that APCs take up tumor antigen, mature, and migrate to the lymph node. Rather than explicitly modeling antigen generation, we assume that the rate of APC stimulation is proportional to the tumor population, , where is a constant related to the antigenicity of the tumor.
The second equation in (3) pertains to APCs that have matured, started to present tumor antigen, and migrated to the lymph node. The model accounts for APC maturation, antigen presentation, and migration as one collective event, because APCs that only undergo one or two of the three processes are not pertinent to the dynamics of the model, since they cannot stimulate tumor-specific CTLs. The first term of the equation corresponds to the rate at which these APCs enter the lymph node. The factor is the ratio between the volumes of the tissue and the draining lymph node. Since we measure populations in terms of concentration, this factor is necessary to account for the change in concentration due to traveling between regions of different volume. The second term is the natural death rate of population.
The third equation in (3) pertains to memory CTLs in the lymph node. The population is replenished up to an equilibrium capacity, , according to a logistic growth model with rate . The second term is the rate of stimulation by mature APCs. The bilinear form of this term follows the law of mass action where is the proportionality constant, or mass-action coefficient.
The fourth equation in (3) pertains to effector CTLs that have finished the division program of divisions. The first term gives the rate at which activated memory CTLs enter the effector state, . This term corresponds to the final term of the previous equation for , except that it has an additional coefficient of and it depends on cell concentrations at time . The coefficient accounts for the increase in population of memory CTLs after divisions, and the time delay, , is the duration of the division program. This term accounts for newly proliferated effector CTLs that appear in the population time units after activation from . The second term is the rate at which cells are stimulated by mature APCs for further division and the third term is the rate at which dividing cells reenter the system time units later after undergoing one cell division. The time delay is the duration of one cell division. The fourth term corresponds to the death of cells at rate . The last term is the rate at which effector CTLs flow out of the lymph node to the tissue at rate .
The last equation in (3) pertains to effector CTLs in the tissue. The first term is the rate at which effector CTLs in the lymph node flow out to the tissue. As with the inflow rate of APCs into the lymph node, this term is scaled by the volume ratio . As shown in the last term, effector CTLs in the tissue die at the same rate at effector CTLs in the lymph node.
To incorporate the model (3) with the ABM, we translate the DDE system (3) into a system of difference equations evaluated at time steps of length , the same time step for the ABM. Our derivation of a system of difference equations from the continuous system is comparable to the reverse process of that used in [45] to translate an agent-based model to a partial differential equation system. More precisely, we translate the system from DDEs to difference equations by assuming that the population variables are constant over intervals of length and that the rates of state transitions across time steps are governed by Poisson processes.
However, since the lymph node contains orders of magnitude higher concentrations of immune cells than the tissue [46], [47] and hence interactions occur orders of magnitude more rapidly [48], we additionally assume that (1) immune populations in lymph node are continuous and (2) transition rates governed by Poisson processes closely follow the mean field rates. In other words, instead of using Poisson random variables , we model transition rates using deterministic factors of the form . As a result, we do not consider stochasticity or discrete populations in the lymph node.
Furthermore, we account for the time-delay terms by incorporating population values from earlier time steps into the difference equation system. In other words, the difference equations for populations at time may depend not only on population values from the immediately preceding time step , but also on population values from earlier time steps for . The system of difference equations that we obtain is given below.
Let , , , , and . Then we rewrite (3) as the following analogous difference equation system:(4)where , , and . Here, we assume that and are positive integers.
The first equation in (4) pertains, as before, to APCs waiting in the periphery. The first term of the equation is the rate at which new APCs are supplied into the system during one time step . The coefficient of the second term is the probability that a cell survives the next time step. Hence, the second term is the concentration of APCs that survive from time to . The coefficient of the third term is the probability that an APC survives the next time step and is stimulated to become a mature, antigen-bearing APC. The coefficients in the the first three equations in (4) similarly express other transition probabilities.
As before, the fourth equation in (4) pertains to effector CTLs in the lymph node. The first term is the rate at which activated memory CTLs enter the effector state after completing divisions. As in (3), this term depends on mature APC and memory CTL concentrations and from time steps earlier. The factor in the second term is the probability that a cell survives the next time step. The factor in the third term is the probability that a cell survives the next time step and gets stimulated by a mature APC to undergo further division. The fourth term is the rate at which dividing CTLs reenter the system time steps later. The factor in the final term is the probability that a cell survives the next time step, does not get stimulated to divide, and flows out of the lymph node to the periphery. The terms in the final equation of (4) are similar to those already discussed.
The advantages of rewriting the DDEs (3) as the difference equations (4) are that the difference equation can be updated in parallel with the ABM with time steps of length , rewriting transition rates from the DDEs in terms of probabilities for the difference equations is consistent with the probabilistic treatment of cell behavior in the ABM, and the numerical values of the difference equations are guaranteed to remain nonnegative. Since the time step that we use is relatively small, numerical solutions of (3) using the Matlab function ‘dde23’ and numerical evaluations of (4) are nearly indistinguishable. Using the ABM algorithm and the difference equation system (4), we simulate the combined system in the following steps:
Parameter estimates for the ABM are shown in Table 1. We discuss how we obtained the estimates below.
For our simulations, we set the time step to , because 1 minute is the timescale of the fastest dynamic simulated in the model, i.e. CTL motion. For the cell radius, we estimate , since typical diameters of CTLs and tumor cells fall around [12], [21], [24], [46].
By fitting a growth model to experimental breast tumor data, Spratt et al. estimate that the initial tumor cell doubling time is between 30 and 4800 days ( years) [28]. In another study, Weedon-Fekjær et al. obtain similar doubling times of 1.2 months to 6.3 years [29]. Other experimental studies report long-term doubling times of around 100 days [26], [27]. However, some mathematical models consider the possibility of aggressive early-stage tumors with division times of under 10 days [8], [24]. To model a relatively fast-growing tumor, we estimate the tumor division time, , to be 7 days, but consider a range from 1 to 400 days. As we see in the Results, this range is sufficient to clarify how this parameter influences the model.
We model CTL motion using a Wiener process, so it is difficult to speak of velocity. Instead, we use the standard deviation of distance displaced per unit time as a substitute measure. Friedl and Gunzer estimate that CTLs migrate at mean velocities of , and Catron et al. choose an estimate of [35], [46]. Therefore, we set the maximum unit standard deviation, , of CTL motion to be , but consider a range from .
It is difficult to estimate how long it takes a CTL to accelerate from stationary to its maximum diffusion rate, so we suppose that the acceleration time, , takes approximately 5 hours and consider a wide range from 0 to 24 hours.
The experimentally measured half-life of effector CTLs during contraction is 41 hours, so in the model, we set [30]. We do not have clear estimates of the average times for CTL recruitment, , and CTL killing, . However, experimental studies show that anti-tumor CTLs can effectively recruit additional CTLs [3], [4] and rapidly kill target cells, sometimes even killing multiple target cells simultaneously [49]. To consider a wide range, we assume that the average CTL recruitment time, , is 8 hours, but we consider a range of 2 to 24 hours. Since killing target cells may require a a long recovery period, we assume that the average CTL killing time, , is 24 hours, but we consider a range of 4 to 48 hours. We consider this an adequate range since the CTL half-life is 41 hours. We do not consider rates of 0 hours, because that would mean CTLs can recruit or kill infinitely fast.
The requirement for the region of interest is that it is large enough to contain the relevant tumor-immune dynamics without inducing too many effects from dynamics occurring too close to the boundary. As a result, we set the radius of the region of interest to be , since such a region can adequately simulate a spherical tumor of over 50,000 cells with ample surrounding space, and the volume of the region conveniently comes out to .
Similarly, we require the CTL cloud to be wide enough that any CTLs that may be beyond the cloud have a very low chance of migrating across the cloud and into the region of interest during one time step. Since CTLs move according to a Wiener process with unit standard deviation , the distance a CTL will move orthogonally toward the surface of the region of interest in one time step is given by the normal distribution . Hence, if we set the width of the CTL cloud to be , the probability that a CTL could pass from outside the cloud into the region of interest is 0.001.
A list of parameters with estimated values for the DDE is shown in Table 1. We discuss how we obtained the estimates below.
An experimental study measuring the volumes of head and neck lymph nodes in men and women estimate lymph node volumes ranging from 0.1 to 1 mL, depending on the location of the lymph node [50]. If we assume our lymph node compartment is approximately 1 mL and that the breast tissue is approximately 1 L, we obtain a volume ratio of .
Cell concentrations are obtained from a study by Catron et al. in which they simulated a hypothetical, spherical, skin-draining lymph node of radius 1 mm [46]. In their paper, they considered a slice of about 1/500 of the total volume and estimated that the slice contains about 1600 CTLs (CD8+ T cells) and 100 dendritic cells (DCs) [46]. Such a slice would have a volume of , yielding T cell and DC concentrations of approximately and , respectively.
We assume that the lymph node contains a population anti-tumor memory CTLs, which were previously induced by a preventative anti-tumor vaccine. For a base estimate, we assume that the equilibrium memory CTL concentration, , in the lymph node is 2% of , and we consider a range of 1 to 10% of . Since we are setting initial conditions for DDEs, we are interested in the history of cell concentrations on the time interval , so we assume that the system was at steady state before time 0 and set for . For the logistic growth rate, we estimate that memory CTLs replenish at rate , which corresponds to a minimum doubling time of 1 day. As seen in the Results, only a very small fraction (less than 1%) of memory CTLs becomes activated by the incipient tumor, so variations in the replenishment rate of memory CTLs does not significantly influence the outcome of the simulations (results not shown), so we do not consider it worthwhile to vary this parameter along with the others.
Since DCs are the primary APCs that stimulate T cells [6, p. 319], we assume that our estimate of the DC concentration is also a good estimate of the APC concentration. We do not know how many APCs reside in a tissue that drains into a particular lymph node, but we assume that it is of the same order of magnitude as the number of APCs in the lymph node. Hence, we estimate that the initial concentration of APCs in the tissue before time 0 is , i.e., for . We assume that all other cell concentrations start at 0.
Next, we estimate the death and supply rates of immature APCs. Since we are dealing with a closed system, we recognize that cells may leave the system due to random circulation or emigration, but for convenience, we incorporate these cases into the death rates. Not having explicit references for the turnover rates of immature APCs in tissue, we assume they are similar to those of naïve T cells, which is estimated to be around 3% per day [51]. Hence, we set the immature APC death rate, , to be and calculate the steady state supply rate to be .
The half-life during T cell contraction is 41 h, so we estimate an effector CTL death rate of [30]. Furthermore, the level of antigen presentation following the third day after infection decays with a half-life of around 19 h and 20.4 h [52]. Hence, using a half-life of 20 h, we obtain a mature APC death rate of . We note that these APCs might not actually be dying. Instead, they might be turning over surface molecules, but for our purposes, these APCs can be considered eliminated.
For the minimal CTL division program, various studies estimate that newly activated CTLs (from a naïve state) undergo between 7 and 10 initial divisions [36], [53], and that a responding CTL population could expand up to five orders of magnitude [34]. This range corresponds to between 7 and 17 cell divisions. Since activated memory CTLs probably undergo more divisions than newly activated naïve CTLs, we assume a base estimate of divisions upon activation and consider a range from 7 to 17 divisions.
To calculate the mass-action coefficient, , we use the estimate that in the lymph node slice of Catron et al., one T cell and one DC will have interactions per hour, or interactions per day [46]. Assuming that DCs represent the majority of APCs that stimulate T cells, we obtain an estimate of the mass-action coefficient [46]. Recalling that the lymph node slice has a volume of , we obtain the unit conversionIt is unlikely that every antigen-specific CTL-APC interaction leads to CTL stimulation, so we set the probability of successful stimulation to 0.5 as a base estimate and consider probabilities from 0.05 to 0.5. These estimates translate to a base estimate of for the mass-action coefficient and a range of .
For the time delays, the duration of one division is between 6 to 12 hours (i.e., 2 to 4 times per day) [6, p. 19]. In addition, the T cell population doubles approximately every 8 hours during expansion [30]. We use the intermediate value of , or 1/3 day, as a base estimate and consider a range from 4 to 24 hours. The CTL division program consists of divisions, but the first division does not occur until 24 hours after stimulation [54], [55]. Hence, we set the duration of the division program to be to account for the fact that the first division takes one day while subsequent divisions take days.
We do not have a good estimate of the antigenicity, , of the tumor, so we assume a base value of and consider a range from to . The parameter can be understood to represent the reciprocal of the rate at which one APCs will encounter and take up antigen from one tumor cell in the tissue. In other words, if we assume that the APC concentration is (i.e., ) in the tissue and that the tissue has a volume of 1 L, then there are APCs circulating in the tissue. As a result, if , it will take an average of for one circulating APC to encounter antigen from a single tumor cell in the tissue. A range of to corresponds to average discovery times of a single APC from 10,000 days () to a couple of hours.
For the flow rate of effector CTLs out of the lymph node to the tissue, we assume that effector CTLs that are not being stimulated to divide emigrate from the lymph node at a half life of 1 day, so that the flow rate .
To derive (2), we take advantage of the connection between random walks on a lattice and the Wiener process. Suppose that at each time step , a CTL has equal probability of moving distance in any of the six cardinal directions on a 3-D square lattice. If we let and go to zero in such a way that remains constant, the random walk approaches a Wiener process corresponding to a diffusion rate and unit standard deviation [41], [42].
Suppose a stationary CTL accelerates at a constant rate and reaches the maximum velocity at time . Hence, at time after beginning acceleration, the CTL has a velocity that is of the maximum, so we suppose that the CTL conducts a random walk of steps size instead of the maximum step size. The associated diffusion rate for this random walk is , which yields a unit standard deviation of . Since a CTL's motion cannot exceed the maximum rate given by , we obtain the expression given in (2).
|
10.1371/journal.pcbi.1004653 | MiR-192-Mediated Positive Feedback Loop Controls the Robustness of Stress-Induced p53 Oscillations in Breast Cancer Cells | The p53 tumor suppressor protein plays a critical role in cellular stress and cancer prevention. A number of post-transcriptional regulators, termed microRNAs, are closely connected with the p53-mediated cellular networks. While the molecular interactions among p53 and microRNAs have emerged, a systems-level understanding of the regulatory mechanism and the role of microRNAs-forming feedback loops with the p53 core remains elusive. Here we have identified from literature that there exist three classes of microRNA-mediated feedback loops revolving around p53, all with the nature of positive feedback coincidentally. To explore the relationship between the cellular performance of p53 with the microRNA feedback pathways, we developed a mathematical model of the core p53-MDM2 module coupled with three microRNA-mediated positive feedback loops involving miR-192, miR-34a, and miR-29a. Simulations and bifurcation analysis in relationship to extrinsic noise reproduce the oscillatory behavior of p53 under DNA damage in single cells, and notably show that specific microRNA abrogation can disrupt the wild-type cellular phenotype when the ubiquitous cell-to-cell variability is taken into account. To assess these in silico results we conducted microRNA-perturbation experiments in MCF7 breast cancer cells. Time-lapse microscopy of cell-population behavior in response to DNA double-strand breaks, together with image classification of single-cell phenotypes across a population, confirmed that the cellular p53 oscillations are compromised after miR-192 perturbations, matching well with the model predictions. Our study via modeling in combination with quantitative experiments provides new evidence on the role of microRNA-mediated positive feedback loops in conferring robustness to the system performance of stress-induced response of p53.
| DNA damage triggered activities of the tumor suppressor protein p53 could be significantly dynamical. The functional role of p53 oscillations in cellular decision making during cancer development has been appreciated. A set of recent studies have revealed extensive crosstalk between the p53 network and microRNAs, but the specifics of the participation of microRNAs in the regulation of the p53 signaling pathway remains largely elusive. Here we investigated microRNAs that form feedback regulation with p53. We enumerated the molecular interactions among these microRNAs and the p53 core and developed a mathematical model to reproduce the DNA damage induced p53 oscillations in single cells. We performed computer simulations and system analysis in combination with experimental assessment to probe the behavior of p53 under microRNA-inhibited conditions. We show that the robust cellular performance of the stress response of p53 in a breast cancer cell line is controlled by miR-192, which forms positive feedback loops with p53.
| Cells depend on complex intracellular signaling to process and react to external stimuli. One prominent type of dynamic response is the periodic accumulation of key transcription factors in the nucleus, where they elicit temporally controlled gene activation [1–4]. The tumor suppressor protein p53, a pivotal player involved in cancer initiation and prevention [5], undergoes oscillations in response to extracellular stress signals. Experiments show that transient DNA lesion of double-strand breaks, induced by acute application of γ-irradiation, trigger oscillatory response of the p53 protein and its negative regulator MDM2 [6–8]. At a single-cell level, the oscillation of p53 is undamped and the mean period of the pulses are constant and independent on the damage level [7]. While the cellular function of the oscillatory dynamics of these transcription factors is unclear, investigations have started to probe the significance of the p53 oscillations in inducing downstream effect such as apoptosis. For instance, recent results demonstrate that the dynamical pattern and not the absolute level of p53 protein controls the life-or-death fate decision in response to DNA damage at cellular level, highlighting the role of p53 oscillations in cellular decision making in cancer [9, 10].
Negative feedback has the potential to generate limit-cycle oscillations and is viewed as a necessary structure for biochemical oscillators [11, 12]. Indeed there exists a consensus in the literature that the p53-MDM2 negative autoregulatory loop is essential for the stress-induced p53 oscillations [3, 13]. A number of mathematical models, that typically assume an explicit time delay in the core p53-MDM2 autoregulatory loop, can reproduce the undamped p53 oscillations [14–16]. More generally, coupled negative and positive feedback loops can give rise to oscillatory phenotypes [11, 17]. The architecture of positive feedback loops, on top of a negative feedback loop, can endow performance properties such as the tunability of frequency, entrainability to cycles, and robustness under molecular noise [17–19]. Indeed, mathematical models can predict sustained oscillations under auxiliary positive feedback loops on p53 [20], but the general role of positive feedback loops in p53 oscillations remains largely elusive.
MicroRNAs are small noncoding RNAs, approximately 22 nucleotides in length that serve as post-transcriptional regulators, and have been shown to regulate the activity of nearly 30% of all protein-coding genes. Intriguingly, a set of recent studies revealed extensive crosstalk between the p53 network and microRNAs [21, 22]. We have identified that, with respect to the dynamical behavior of the system, several microRNAs form positive feedback loops with p53, typically through direct or indirect inhibition of MDM2.
In this work, we investigate the role of microRNA-mediated positive feedback loops in regulating the performance of p53 stress network. We first developed a mathematical model of a microRNA-p53-MDM2 network that involves three different microRNAs that form positive feedback loops. The core p53-MDM2 model is based on our previously published work [14, 16]. We performed simulations and studied the robustness of p53 oscillations under abrogation of microRNA-mediated positive feedback loops. Furthermore, we adopted bifurcation diagrams in order to explore the system behavior under parametric variability in relationship to cellular noise. To experimentally evaluate our in silico predictions, we introduced microRNA inhibitors in a modified breast cancer cell line MCF7, and performed time-lapse microscopy tracking single-cell p53 dynamics under induced DNA double-stranded breaks. Our experimental quantification, in agreement with modeling analysis, reveal that the three microRNA-mediated positive feedback loops confer different level of control to the robust performance of stress-induced p53 oscillations within a population of cells.
In this work, we seek to elucidate the role of miRNAs in the regulation of the p53 oscillation elicited by the stress signal of cellular DNA damage. Among the signaling regulations of p53 induced by miRNAs, we focus on feedback pathways. Intriguingly, three groups of miRNAs that are identified to be a part of feedback regulations of p53, form positive feedback loops with the p53 pathway. The three microRNA-mediated positive feedback networks and the associated molecular interactions are described as follows.
The miR-192 family, including miR-192, miR-194 and miR-215 [23], is directly correlated with p53 protein upregulation [24] and overexpression of these microRNAs elicited dramatic down-regulation of MDM2 at protein and mRNA levels. These findings indicate that, on top of the core autoregulation of p53 through MDM2, there is a microRNA-mediated autoregulatory loop of p53, where miR-192 is activated by p53 [25] and in turn inhibits the antagonizing effect of MDM2 [24]. This autoregulatory loop is a positive feedback loop with the feature of double-negative regulation (Fig 1a).
The miR-34 family, including miR-34a, miR-34b, and miR-34c, is upregulated by p53 [26, 27]. Two positive feedback loops between p53 and miR-34a have been reported (Fig 1b). The first loop is through the regulation by the protein named silent information regulator (SIRT1). Specifically, miR-34a inhibits SIRT1 mRNA translation [26, 27]. In irradiated cells the SIRT1 protein acts as an antagonist of the post-translational modification of p53 and thus repressing the transcriptional activity of the p53 protein [28]. The second feedback loop is mediated by Yin Yang 1 (YY1), a ubiquitous transcription factor that negatively regulates p53, and is directly repressed by miR-34a [29]. The YY1 protein can enhance the degradation of p53 promoted by MDM2 [30], thereby closing a feedback loop composed of p53, miR-34a, YY1 and MDM2. Both of the two regulatory loops are positive feedback with double-negative regulations (Fig 1b).
The miR-29 family, including miR-29a, miR-29b, and miR-29c, is upregulated by p53 [31]. The miR-29 family members in turn can enhance p53 activity. For instance, all three miR-29 family members can successfully elevate the phosphorylation level of p53 by repression of Wip1 [31], a phosphatase of p53 [13]. In addition, miR-29 microRNAs directly suppress CDC42 [32], a Rho family CTPase, which directly inhibits the protein activity of p53. Intriguingly, these feedback regulatory pathways are again positive feedback loops in the form of double negative feedback (Fig 1c), and they are closely interlinked with the core p53-MDM2 autoregulation in that Wip1 upregulates MDM2 via inhibiting its degradation [33] (Fig 1c).
We first developed a mass-action model that accounts for the core p53-MDM2 autoregulatory network [14, 16] coupled with all the three families of microRNA-based positive feedback loops at single-cell level (Fig 2). Each microRNA is modeled by accounting for the mediating microRNA component and its associated target proteins (i.e. SIRT1 and YY1 for miR-34a, and CDC42 and Wip1 for miR-29a) (Fig 3a). We assume that the microRNA binds quickly with its target mRNA molecule and dispose the microRNA-mRNA complex into degradation [34]. The active form of ATM, a protein kinase that detects DNA damage, is induced by transient DNA damage signal, following the mathematical formula used in Batchelor et al [35]. The assumptions on the interactions among p53, MDM2, microRNAs and intermediate proteins, as well as the ordinary differential equations and parameters of the deterministic single-cell model are included in the Supporting Information (S1 Text, S1 and S2 Tables). Note that our model includes a second negative feedback loop formed by ATM, p53 and Wip1 (Fig 2 and S1 Text), a network structure proposed in previous experimental and computational studies [35, 36]. As shown in a recent study of NF-kB, another oscillatory transcription factor, a longer negative feedback loop in addition to the core faster negative feedback could provide further system properties, such as better tracking of duration of input signal as well as potential induction of damped oscillations [37]. These behaviors potentially allow for more sophisticated signal coding and processing patterns in cellular stress response than that could be achieved by single negative feedback structure.
A simulation of the deterministic model of microRNA-p53-MDM2 network at wild-type condition shows that this system yields oscillations of p53 and MDM2 with period of approximately 5 hours under constant DNA damage stimulus (Fig 3b), equivalent to single-cell response to γ-irradiation or radiomimetic drug observed in our in-house experiments and published experiments [6, 7, 38]. We next probed the behavior of p53 in response to DNA damage and the associated role of the microRNAs. More specifically we introduced to the mathematical model inhibitors that modulate the three microRNAs by complexation reactions. We assume that the microRNAs are repressed by ~6-fold after addition of the inhibitors. Simulations of the deterministic model show that, when miR-192 is inhibited, the stress-induced oscillation of p53 is abolished, but when miR-34a or miR-29a is inhibited the oscillation of p53 persists (Fig 3c). These simulations indicate that at single-cell level the p53 oscillatory behavior is more sensitive to the down-regulation of miR-192 than miR-34a and miR-29a. In light of the deterministic simulation results and considering that a prominent feature of the dynamics of a cell population is the cell-to-cell variability, we decided to investigate further the single-cell system behavior under cellular noise.
The heterogeneity in a cell population has been widely observed in the experiments of the stress response of p53 and other cellular processes [6, 7, 39]. Cellular noise, broadly defined as stochastic fluctuations of molecular processes within and between cells, can be divided into intrinsic and extrinsic noise [40, 41]. Intrinsic noise refers to random deviation of the molecular processes from their average deterministic kinetics within a cell, mostly due to probabilistic biochemical reactions associated with low copy numbers. Several attempts have been made to use stochastic models to study the noisy single-cell p53 dynamics under the influence of intrinsic noise due to low copy number of reactants [42, 43]. Nevertheless, the high molecule numbers measured in the p53 network (104−105) [14, 44] suggest that the role played by intrinsic noise may not be critical especially in the variable induction of oscillatory and non-oscillatory phenotypes in single cells [6], as intrinsic noise mostly just results in irregular profiles of a trajectory with high copy numbers [42]. Indeed, a previous study shows that oscillation produced by limit cycle seems to be very robust under intrinsic stochasticity, where the simulated stochastic oscillations persist when the maximum molecule numbers are in the order of hundreds [45]. On the other hand, extrinsic noise generally dominates cellular stochasticity, especially in eukaryotic systems [46–48], and arises from global factors that impact cell-to-cell variation [49]. Therefore, in this study we focus on analyzing the impact of extrinsic noise on the sustainability of p53 oscillation at single cell level. To this end, deterministic single-cell model with varying model parameters can be used to compute the impact of extrinsic noise [50–52].
In our experimental setup, the perturbations are performed using the inherently “noisy” transient transfections of microRNA inhibitors, resulting to variable down-regulation levels between cells. To probe the effect of microRNA inhibitor copy-number variability we first performed parametric simulations for a wide range of inhibitor concentrations. As illustrated in Fig 3d, the effect of the miR-192 inhibitor is more prominent, leading to gradual collapse of the oscillating p53 behavior at high concentrations.
To further investigate the effect of microRNA abrogation under cellular extrinsic noise we performed bifurcation analysis. We assayed model parameters along the microRNA-mediated positive feedback loops that directly affect the transduction of the microRNA perturbation through the network. In particular, we probed 14 parameters under wild-type and the three microRNA-repressed conditions (S3 Table). First, we calculated the bifurcation diagrams of the steady-state p53 concentration versus the 5 association rates between the microRNAs and their target mRNAs for the wild-type case as well as the three perturbed cases with microRNA inhibitors (Fig 4). The paired dots represent the bounds of p53 oscillation amplitude at steady state and the solid line represents stationary steady state. According to Fig 4, the miR-192 inhibitor either reduces the range of the oscillatory p53 response, or abolishes p53 oscillation, across different affinities of miR-192 to its target. Note that the p53 response at the nominal parameter set is still within oscillatory region when miR-34a and miR-29a are inhibited, while it falls out of the oscillatory region when miR-192 is repressed. This is consistent with the persistent oscillations in the former two cases and the stationary steady state in the latter case, as shown in Fig 3b. The bifurcation diagrams versus the rest 9 parameters are shown in S1 Fig.
It is noteworthy that the experiments by Geva-Zatorsky et al showed that the amplitude of p53 in individual cells is highly variable with a variation up to ~70%. There have been attempts to implement the high variation of p53’s amplitude by theoretical modeling. For instance, Jolma et al assumed that certain rate parameters were allowed to vary randomly and rapidly within a certain range to achieve the variable p53 amplitude [53]. Such method essentially implements the extrinsic noise computationally as explained above. In our model, the variable p53 amplitude can also be induced by extrinsic noise via allowing variations in parameters, whereby the impact of varying parameters on p53 amplitude is demonstrated by the bifurcation plots (Fig 4 and S1 Fig). For instance, varying the value of kon1 alone between [10, 40] achieves ~65% variation of the p53 amplitude (see the wild-type case in the plot with respect to kon1 in Fig 4). Also, varying the value of kw alone between [1, 3] achieves ~53% variation of the p53 amplitude (see the wild-type case in the plot with respect to kw in S1 Fig). Therefore, significant variation in p53 amplitude in our model is attainable by assuming considerable extrinsic noise in the model parameters.
The bifurcation diagrams with respect to each of the 14 parameters embedded in the microRNA-based positive feedback loops parameters (S3 Table) reveal that for 8 out of the 14 parameters the repression of miR-192 leads to the smallest regions of oscillation compared to those of miR-34a and miR-29, while for the other 6 parameters the repression of miR-192 completely abolishes the p53 oscillation over the varying ranges (Fig 4 and S1 Fig). These plots indicate that the non-oscillating phenotype of an individual cell can be yielded when certain parameter, due to extrinsic fluctuation, is pushed out of the bounds of its oscillatory regime, thus providing a plausible mechanism underlying the observed heterogeneous behavior of p53 in a cell population. Moreover, if the region of p53 oscillation significantly shrinks, it is more likely for the oscillatory behavior to be ruined by extrinsic noise, and thus the probability of observing non-oscillating single-cell phonotype in a stochastic population should increase.
To further quantify the impact of different types of microRNAs on regulating the p53 network, we measured the system’s robustness performance, which is the capability to maintain the oscillatory behavior of p53 in response to DNA damage (S2 Text). A large robustness index defined in S2 Text predicts that the system’s probability of sustaining stable oscillation under stochastic perturbations is relatively high. As a result, for the model under a particular condition the value of its robustness index is positively correlated with the fraction of oscillating cells in a population. The robustness indices confirm that when miR-192 is repressed the oscillatory phenotype is the least robust among the three microRNAs (S4 Table). Based on the model predictions, we infer that the inhibition of miR-192-mediated positive feedback loop would lead to the highest probability of non-oscillating cells across a population due to extrinsic noise.
As a summary, the theoretical modeling and analysis show that the robust performance of the p53 stress network is subject to the control of specific microRNA-feedback regulation.
To experimentally probe the effect of microRNA abrogation on the p53-MDM2 oscillator we used a breast cancer cell line MCF7 [38] that contains a stably integrated fluorescent reporter Venus fused to the cDNA of p53 under the expression of the metallothionein promoter (Fig 5a). To down-regulate the desired microRNAs we introduced to MCF7 cells synthesized single-stranded RNA molecules that are complementary to the mature microRNA sequence.
Prior to performing the microRNA perturbation experiments, we verified the expression of the three selected microRNAs using quantitative PCR (qPCR). We confirmed the expression of miR-29a, miR-192, and miR-34a as well as efficacy of their inhibitors (Fig 5b); results from qPCR show between 70% to 80% down-regulation for each microRNA targeted.
We then performed two independent time-lapse experiments to test the following five conditions: wild type (WT) untransfected cells, negative control with transfection using a synthetic microRNA that does not target the p53-mdm2 core, and the three selected microRNA inhibitors. For the negative control case we transfected the MCF7 cells with a synthetic microRNA (FF4) [54] that does not interfere with the p53-MDM2 core.
The oscillations of p53 protein after addition of the microRNA inhibitors and neocarzinostatin (NCS) [38] were captured using time-lapse microscopy. The details of the preparation of the wild type and microRNA-perturbed MCF7 cells and the subsequent time-lapse microscopy are described in the Materials and Methods Section. The image data acquired from the time-lapse microscopy were processed using ImageJ [55] and MATLAB. First, the fluorescence signal of the nuclear-localized p53 cells in the image stack was tracked and the average fluorescent intensity of p53 in individual cells was recorded for each cell at 10-minute intervals for 20 hours.
To evaluate the impact of the microRNA perturbations on the stress-induced p53 oscillations, we analyzed one hundred single-cell trajectories of p53 fluorescence profiles, after artificially induced DNA damage. The raw trajectory data was denoised using the stationary wavelet transform (SWT) in MATLAB [56] (S2 Fig). After smoothing out the raw signal intensities to reduce noise, we classified each denoised p53 trajectory data by extracting relevant parameters of the intensity profile to help us determine whether the observed fluorescence profile possesses qualities consistent with typical p53 oscillation or not (Materials and Methods).
Specifically, we located relevant peaks of the time-series data and recorded their locations to identify cells that have abnormally long or short periods. If two consecutive peaks occur within 50 minutes or at least 12 hours apart, the cell was eliminated from being classified as oscillatory. For a better look at the overall trend of the dynamic fluorescence signal, we also calculated the instantaneous slope of the fluorescence profile and the frequency of the time that it is below zero during the 20 hour span. If the slope was zero more than 70% of the 20 hour period, we exclude the cell from being classified oscillating and vice versa. The entire classification process was automated in MATLAB, and the classified cell-population results of the duplicate experiments under WT, negative control, as well as the three microRNA-inhibited conditions are illustrated in S3–S7 Figs.
After sorting our time-lapse data for individual cells showing oscillation, we found that the targeted microRNA suppression seems to affect the stress-induced p53 oscillation quantitatively in terms of the number of cells with oscillating p53 expression, but have a little qualitative effect on the period or amplitude of the oscillations. Specifically, we found the mean oscillation period among the selected cells to be approximately 5 hours, and this mean period remained stable after suppression of the three targeted microRNAs. As expected, we found that there was non-negligible level of cell-to-cell variability in the oscillation period based on the coefficient of variation. More importantly, we found that the microRNA suppression seems to have little effect on this variability (Fig 5c).
We then calculated the percentage of oscillating cells after each microRNA abrogation. In our time-lapse experiment, fluorescence signal from 100 cells were captured every 10 minutes over a 20-hour time period in 5 different conditions, giving us over 60,000 data points per experiment. To present this data efficiently, we employ heatmaps composed of pixels that each represents a single data point. Each row of the graph represents the p53 signal intensity trajectory of a single-cell, and each column represents a single time point. Color at each pixel is indicative of the relative intensity of the p53 signal. To highlight the differences between the fluorescence profiles of oscillating and non-oscillating cells, we re-organize the time-series heatmaps into two populations and re-order them based on the location of the first observed peak. We found that the cells transfected with the inhibitor of miR-192 show markedly decreased number of p53 oscillating cells comparing to wild type, while the population affected by inhibitors of microRNA 29a and 34a show similar occurrence rate of oscillation as the wild type (Fig 5d). Note that the results from duplicate experiments show the same trend of reduced number of oscillating cells in a population only when miR-192 is repressed, although the absolute value differs (S5 Table).
Here we use an approach of theoretical modeling combined with quantitative experiments to elucidate the role of microRNAs on the cellular performance of oscillatory p53 induced by DNA double-strand breaks. Our results show that the microRNA-mediated positive feedback loops influence the robust manifestation of stress-induced p53 oscillations in stochastic cellular systems. Specifically, the repression of miR-192 led to widespread collapse of the sustained p53 oscillations across a population of variable cells while the repression of miR-34a and miR-29a mildly affected the phenotype under double-stranded DNA damage.
A functional role of microRNAs has been proposed in that they confer robustness to biological processes [57], including cellular differentiation in development or tumorigenesis [58]. Notably, the microRNA-mediated functional network motifs that previously have been discovered to bestow the function of robust maintenance of cell fate are all positive feedbacks, consisting of a transcription factor and a microRNA, either with or without intermediate signaling components [58]. Our findings add new evidence of microRNA-mediated positive feedback loops that function as a mechanism that reinforces the robustness of a system phenotype.
Bifurcation analysis, a method widely used in engineering to evaluate system robustness, provides effective means to delineate the variable behavior of single cells subject to extrinsic noise in parameters. Our bifurcation diagrams of the single-cell model of microRNA-p53-MDM2 network show substantially higher reduction of oscillation regions under the inhibitions of miR-192 comparing to the other two microRNAs. Theoretical studies of the biochemical oscillators arising from a core negative feedback loop plus an additional positive feedback loops have been performed recently [59, 60]. Specifically, a modified Goodwin model consisting of a three-component negative feedback loop interlinked with different positive feedback motifs was studied for the performance of the oscillator with regard to the benefits acquired by the auxiliary positive feedback regulations. This is analogous to our wild-type model, where the core mechanism for the p53 oscillator is the p53 protein-MDM2 mRNA-MDM2 protein negative feedback loop, and it is coupled with positive feedback loops. Besides the advantages that may be gained for the oscillator by positive feedback loops, their results show that a positive feedback loop is the most beneficial for the robust performance of the oscillator if its pathway components have the fastest dynamic, such as the fastest degradation rate. In other words, the finding predicts that a positive feedback loops with faster information transduction on top of the core negative feedback is a more favorable structure to stabilize oscillation. We can apply this finding qualitatively to interpret our model behavior. For the positive feedback regulations in our model, miR-192 directly regulates MDM2, while miR-34a and miR-29a regulate intermediate nodes prior to reaching MDM2, before the MDM2 information is eventually fed back to p53 protein to close the loop. Consequently the miR-192 mediated double-negative feedback loop contains shorter signal-processing path than that of the miR-34a- and miR-29a-mediated feedback loops. Although miR-34-a and miR-29a each also forms a short positive feedback loop together with a mediating protein with the same length as the miR-192-MDM2 feedback loop, the MDM2 protein is degraded at a faster rate upon DNA damage and thus the latter loop would still contribute the most to the system robustness. We therefore infer that among the three groups of microRNAs forming positive feedback loops with p53, miR-192 and its mediated feedback pathway plausibly exert the greatest impact on maintaining the robustness of p53 oscillations in response to DNA damage.
A previous modeling study has proposed the role of a different positive feedback loop in enhancing the robust stability of p53 oscillation [61], whereby the cytosolic MDM2 protein translocates into nucleus to interact with the mRNA of p53 through direct binding and promote the translation of p53 mRNA to close the loop [62, 63]. Such positive feedback is a relatively long and slow loop involving steps of the compartmental trafficking of MDM2 protein from cytosol into nucleus, the protein-mRNA binding between MDM2 protein and p53 mRNA, and the final translation of p53 protein induced by MDM2 protein. The microRNA-mediated positive feedback loop, on the other hand, is more efficient. For instance, the positive feedback mediated by miR-192 is only composed of fast binding of microRNA to MDM2 mRNA and the post-translational degradation of p53 protein promoted by MDM2 protein. The core negative feedback is also a fast and efficient loop containing the post-translational degradation of p53 protein promoted by MDM2 protein. Note that the same type of processes in the positive feedback loops, such as the induction of mRNA or microRNA by p53 and the translation of MDM2 protein, is not enumerated in the above comparison. In addition, we note that translational process is in general much slower than post-translational regulation. In sum, the positive feedback facilitated by MDM2-enhanced translation of p53 occurs in a much lower efficiency than the core p53-MDM2 negative feedback and the microRNA-mediated positive feedback loops. It thus is reasonable to assume that the positive feedback loop through the MDM2-enhanced translation of p53 has relatively weak impact on the p53 oscillation. This probably is the reason why the recent major theoretical studies of the p53 oscillator do not account for the positive feedback loop through the MDM2-enhanced translation of p53 [36, 42, 64–66]. Indeed, if we add an MDM2-dependent translation term into the ordinary differential equation of the p53 mRNA with a translational rate half that of the basal translational rate to approximate the slow processes due to MDM2 translocation and MDM2-p53mRNA interaction, the simulations of the p53 oscillation are very similar to the model without the positive feedback through the MDM2-enhanced translation of p53, indicating that this positive feedback does not have much impact on the p53 oscillation.
In conclusion, our modeling and experimental results provide new evidence on the relationship between microRNAs and p53 function [23] with implications to cancer initiation and progression. Importantly, understanding the mechanisms underlying the abnormal p53 behavior due to microRNA depletion [67] may lead to innovative microRNA-based therapeutics.
The MCF7 breast cancer cell line, consisting of the fluorescent protein Venus fused to p53, is the same as previously described [9, 38], a gift from Galit Lahav, Harvard Medical School. Cells were maintained in 95% humidity at 37 degrees Celsius and cultured in RPMI (Invitrogen) media with 10% FBS (Invitrogen), 1% PenStrep (Invitrogen 0.045 units/mL Penicillin, 0.045 units/mL Streptomycin). After the first splitting following resurrection from liquid nitrogen, stably integrated MCF7 cells were maintained at 20 mL volume in petri dishes with 400ug/mL G 418 disulfate salt (400ng/mL, Sigma). In a 12 well plate (Griener), 80,000 MCF7 p53-Venus cells were plated on the afternoon before transfection. The following morning, the cells were treated with Neocarzinostatin (Sigma) and transfected with 3ul JetPRIME (Polyplus) mix with 25nM of the microRNA inhibitors 192, 29a and 34a (Qiagen) according to the manufacture’s protocol. To stimulate the activity of p53 we added the radiomimetic drug Neocarzinostatin (NCS), which induces the particular lesion of DNA double-strand breaks and elicits p53 to oscillate [35, 38]. Following the addition of 0.8 μl NCS per well from 0.5 mg/mL stock (Sigma) and either with or without the transfection of microRNA inhibitors we commenced a time-lapse microscopy at 37 degrees Celsius with humidified 5% CO2.
Images were collected every 10 minutes for the bright field and fluorescent intensity of p53-Venus using a Hamamatsu camera attached to the Olympus IX81 microscope at 10x magnification. The time lapse ran for 24 hours and used exposure times of 10ms for Bright Field and 500ms for YFP. We chose three positions for each well, ensuring that the imaging field did not overlap between positions. The filter used for capturing Venus fluorescence is excitation ET500/20x and emission ET535/30m (Chroma). Cells were incubated within the microscope at 37 degrees Celsius with approximately 5% CO2.
The image stacks were first processed by an ImageJ plug-in CGE to measure and track the average intensity of nuclear p53 in single cells [55]. For a specific cellular condition, we tracked an average of 33–34 cells within each of the three locations, and a total number of 100 cells, for 20 hrs (see the tracked cell trajectories in S6 Table). The raw time trajectories of p53 intensity for 100 cells then underwent a de-noising step implemented by the Stationary Wavelet Transform De-Noising 1-D Tool of MATLAB to remove the high-frequency noise and extract the low-frequency p53 oscillation [56] (S2 Fig). Finally, each trajectory was subject to a classification algorithm for the purpose of determining its phenotype, oscillating or non-oscillating. Specifically, to obtain additional characteristics from the resulting p53 trajectories, instantaneous slope at each point was calculated using MATLAB code diff, and the peaks were detected using MATLAB code mspeaks. Instantaneous slope of each trajectory, along with the locations of its peaks, were used to determine the oscillating and non-oscillating phenotypes of individual cells. The single-cell trajectories were classified as non-oscillating (S3–S7 Figs) if: (a) less than 3 peaks were detected during the time-lapse, (b) there were 2 consecutive peaks that occur within 5 time units (100 min), (c) there were 2 consecutive peaks that occur more than 70 time units (1400 min) apart, and (d) the slope of the trajectory was negative (or positive) more than 75% of the time. Otherwise, the trajectory was classified as oscillating.
|
10.1371/journal.ppat.1000185 | Immunity to HIV-1 Is Influenced by Continued Natural Exposure to Exogenous Virus | Unprotected sexual intercourse between individuals who are both infected with HIV-1 can lead to exposure to their partner's virus, and potentially to super-infection. However, the immunological consequences of continued exposure to HIV-1 by individuals already infected, has to our knowledge never been reported. We measured T cell responses in 49 HIV-1 infected individuals who were on antiretroviral therapy with suppressed viral loads. All the individuals were in a long-term sexual partnership with another HIV-1 infected individual, who was either also on HAART and suppressing their viral loads, or viremic (>9000 copies/ml). T cell responses to HIV-1 epitopes were measured directly ex-vivo by the IFN-γ enzyme linked immuno-spot assay and by cytokine flow cytometry. Sexual exposure data was generated from questionnaires given to both individuals within each partnership. Individuals who continued to have regular sexual contact with a HIV-1 infected viremic partner had significantly higher frequencies of HIV-1-specific T cell responses, compared to individuals with aviremic partners. Strikingly, the magnitude of the HIV-1-specific T cell response correlated strongly with the level and route of exposure. Responses consisted of both CD4+ and CD8+ T cell subsets. Longitudinally, decreases in exposure were mirrored by a lower T cell response. However, no evidence for systemic super-infection was found in any of the individuals. Continued sexual exposure to exogenous HIV-1 was associated with increased HIV-1-specific T cell responses, in the absence of systemic super-infection, and correlated with the level and type of exposure.
| Serosorting, the practice of seeking to engage in unprotected sexual activities only with partners who are of the same HIV-1 status, is a growing trend. Unprotected sexual intercourse between two HIV-1 infected individuals can lead to consequences such as HIV-1 super-infection. However, continued exposure to HIV-1 may also have an important influence on the immune response. Here, we explored this influence in a cohort of HIV-1 infected individuals who were in long-term partnerships with other HIV-1 infected individuals. We found that individuals, who regularly engaged in unprotected receptive sexual intercourse with an HIV-1 infected viremic partner, displayed higher T cell responses to HIV proteins compared to those who were not regularly exposed to a viremic partner. None of the individuals within this study showed evidence of systemic super-infection. Exposure had limited impact on general activation or poly-functionality. These results are clearly of importance for HIV-1 infected individuals who chose to engage in unprotected sexual activity with other HIV-1 infected individuals. These data also reveal a more general mechanism that occurs in infectious diseases: immune responses to chronic viruses are influenced not only by the virus within the host, but also by exposure to the virus from without.
| Immune responses seen during chronic viral infections are thought to be driven only by ‘endogenous’ virus. However, continued exposure to ‘exogenous’ virus could boost anti-viral immunity. HIV-1 infection provides a model to test this hypothesis. Continued sexual intercourse between two HIV-1 infected individuals leads to exposure to exogenous HIV-1. Recent reports have shown that there is a growing trend for serosorting, the practice of seeking to only engage in unprotected sexual activities with partners who are of the same HIV-1 status [1],[2]. This provides a model system in which to investigate the influence of exogenous versus endogenous viral exposure on immunity to a chronic virus infection. Moreover, given the potential risks of super-infection upon re-exposure to HIV-1, understanding the immune responses involved is particularly important [3],[4].
CD8+ T cells are thought to play an important role in controlling HIV-1. Model viral infections, such as lymphocytic choriomeningitis (LCMV) and simian immunodeficiency virus (SIV), have shown that the CD8+ T cell response is a crucial component in the control or elimination of viral infections [5],[6],[7]. Moreover, the power of CD8+ T cell responses have been elegantly shown to be a driving force in the selection of escape variants in SIV [8].
The immune response towards HIV-1 is complex; we set out to determine if T cell mediated responses in an HIV-1 infected individual can be stimulated through exposure to exogenous virus. However, even this simple question has its complexities, as responses to exogenous virus are indistinguishable from responses directed toward the primary infecting virus. To address this we selected individuals with suppressed viral loads while on antiretroviral therapy. Previous studies have shown that viral suppression by highly active antiretroviral therapy (HAART) to below 50 copies/ml leads to the subsequent waning of anti-HIV-1 T cell responses, due to the reduction in viral antigen [9],[10],[11]. Therefore, we hypothesized that individuals with suppressed viral loads while on HAART, who are regularly exposed to an HIV-1 viremic partner, would have greater anti-HIV-1 specific T cell responses compared to individuals who are exposed to a non-viremic partner.
We studied 49 individuals from the San Francisco Positive Partners prospective couples cohort who were suppressing their virus while on HAART. The subjects were divided into two groups depending on the viral status of their partner, the viremic partner (VP) group, and the non-viremic partner (NVP) group. These groups did not differ in terms of the clinical parameters, such as age, time on therapy, CD4+ T cell count and viral loads (Table 1), or the level of sexual activity, as defined by the exposure score or the average number of insertive or receptive exposures (Table 2). However, there was a trend for higher levels of exposure in the subjects from the NVP group.
Analysis of T cell IFN-γ responses revealed significantly more individuals from the VP group made responses (as defined in the materials and methods section) to HIV-1 Protease and Integrase peptides, compared to individuals with the NVP group (Figure 1A). However, there was no difference in the percentage of individuals from each group making responses to HIV-1 Gag, Reverse Transcriptase (RT), Nef, or CEF. Further analysis of the individuals who made responses revealed that there was no significant difference in the magnitude of the responses made to either of these proteins (Figure 1B). However, Protease, RT, and Integrase (p<0.001) responses were significantly higher in the VP group compared to the NVP group. Overall, there was a trend towards a higher magnitude of response to HIV-1 Gag, Protease, Integrase and Nef by individuals with viremic partners (VP group). Moreover, cumulatively there was a significantly higher anti-HIV-1 T cell response made by the VP group (p = 0.0274).
We further explored the behavioral influence of exposure on the T cell response by utilizing data on the frequency of unprotected sexual intercourse in the partnerships. The exposure scores were plotted against the magnitudes of IFN-γ responses for individuals within each group (Figure 2). Individuals from the VP group showed strong correlations between their level of exposure and their T cell responses to peptide pools corresponding to HIV-1 Protease, RT, Integrase and Nef (Figure 2A). Gag and CEF peptide pools showed no significant correlation, although Gag responses did follow the same trend as the other HIV-1 proteins. In contrast, no correlations were found between the exposure scores and IFN-γ responses made by individuals whose partners were suppressing their virus (NVP group) (Figure 2B). Thus, not only is the level of exposure important in determining the magnitude of the T cell response, but critically it is exposure to a viremic partner that is associated with these responses.
The type of exposure was further analyzed in the VP group by assessing the average number of unprotected receptive (Figure 3A) versus insertive (Figure 3B) anal intercourse episodes. All the individuals in this study engaged in both activities. Significant correlations were only seen between the IFN-γ responses and the number of receptive exposures. Again, responses against HIV-1 Protease, RT, Integrase, and Nef derived peptides correlated with the immune response. Gag followed the trend but did not reach significance, while CEF responses showed no correlation. There were no significant correlations between the average number of insertive exposures and IFN-γ responses towards HIV-1 Protease, RT, Integrase and Nef. As expected, data from the NVP group did not show a correlation between the IFN-γ responses and the direction of exposure (data not shown).
To determine which T cell subset was responsible for the IFN-γ response detected in the ELISpot assay, individuals from the viremic partner group (n = 7) were screened by flow cytometry. Individuals were tested using the same 15mer peptide pools used in the ELISpot assay. Both CD4+ and CD8+ T cell subsets produced IFN-γ in response to all the HIV-1 peptide pools (Figure 4). However, no significant difference between the two subsets was seen, although there was a trend towards the CD8+ T cell subset.
In recent studies poly-functional T cells have been associated with slower disease progression [12]. PBMC from 8 individuals from the viremic partner group and 10 from the individuals from the non-viremic partner group were stimulated with 15mer peptide pools covering HIV-1 Gag and RT. CD4+ and CD8+ T cell expression of IFN-γ and IL-2 were measured by multi-parametric flow cytometry (Figure 5). No significant difference in the level of poly-functionality was seen between the two groups, although there was a trend for greater responses in the viremic partner group. IL-2 expression by CD8+ T cells was significantly higher (p = 0.0059) in the viremic partner group compared to the non-viremic partner group. In the positive control, SEB stimulation, revealed poly-functional T cells (data not shown).
The total level of T cell activation was measured by CD38 and HLA-DR expression. No significant difference was seen in the level of expression of the activation marker CD38 on either CD4+ or CD8+ T cells (Figure 6 A and B). However, there was a significantly higher level of CD38+ and HLA-DR+ double positive CD8+ T cells in the viremic partner group (p = 0.0312), although this was not reflected in the CD4+ T cell subset (p = 0.5148) (Figure 6 C and D).
Seven individuals from the viremic partner group had PBMC samples available from a one year follow up time point. All the individuals continued to suppress their viral load below the level of detection (<50 copies/ml) and their CD4+ T cell count remained constant (Table 3). The exposure scores generated from the second time point showed little change in exposure levels in the majority of individuals (Figure 7A). However, two individuals from the viremic partner group had partners who started anti-retroviral therapy and were suppressing their viral loads at the second time point, one year after the first (Figure 7A filled squares). Both individuals also showed a drop in exposure, which was mirrored by a drop in the response to RT, although responses to Gag did not change (Figure 7B and C). A third individual whose exposure score dropped dramatically at the second time point also showed a drop in their RT response (Figure 7D). One individual had a modest increased the level of exposure to their partner compared to the pervious time point. However, no responses to either Gag or RT were detectable at either time point (Figure 7E).
T cell responses can be influenced by a number of other factors in addition to direct viral antigen stimulation. In order to account for this, we analysed the responses measured in the VP group against a number of clinical parameters (Table 4). Only significant correlations could be found for the length of time the individuals were on therapy and responses to Protease and RT. However, this association was confounded by exposure. In a least squares linear regression model where time on treatment and exposure effects was independently controlled, time on treatment provided no additional explanatory power. In fact, the effect of treatment on both RT and Protease T cell responses completely washed out (r = 0.014, p = 0.94; and r = 0.14, p = 0.51 respectively), while the effect of exposure remained high and significant in the model overall (r2 = 0.56, p = 0.001; r2 = 0.551, p = 0.001 respectively). We found no correlation between any of the other parameters and the magnitude of the T cell IFN-γ responses.
The most likely mechanism, by which these responses are maintained or boosted in the individuals with viremic partners, is through the infection of host cells. Thus, potentially these individuals could be super-infected with their partner's HIV-1. To address this concern we used phylogenetical analysis to assess all the patients within this study at both Gag (data not shown) and Pol (Figure 8). Population analysis of the individual's cellular DNA and partner's plasma RNA revealed no evidence of systemic super-infection in any of the individuals studied within either group.
The consequence of continued exposure to HIV-1 in individuals already infected, to our knowledge has never been reported. This is becoming increasingly relevant with the advent of a growing trend for serosorting, the practice of seeking to only engage in unprotected sexual activities with partners who are of the same HIV-1 status [1],[2]. The study set out to measure the impact of continued exposure to HIV-1 on anti-HIV-1 T cell responses in individuals already infected.
Here, we have shown that HIV-1+ individuals who are regularly exposed to an HIV-1+ viremic long-term partner display greater HIV-1-specific CD4+ and CD8+ T cells responses, than infected subjects with aviremic partners. Furthermore, in these individuals the magnitude of the T cell IFN-γ response towards Pol and Nef significantly correlated with the level of exposure. Longitudinally, RT responses mirrored the level of exposures.
These observations are further strengthened by the fact that only the number of receptive events correlated with the T cell IFN-γ responses and not the number of insertive events. This is consistent with the observation that receptive intercourse represents a greater risk of acquiring HIV infection compared to insertive intercourse [13].
It is interesting to note that responses were predominantly directed towards HIV-1 Pol proteins rather than to HIV-1 Gag proteins. This is in agreement with Karlsson et al, who showed responses switch from predominantly Gag to Pol in individuals on antiretroviral therapy [14]. Moreover, it was the RT responses that mirrored decreases in exposure longitudinally, suggesting that RT-specific responses are more susceptible to antigen levels compared to Gag.
The responses observed reflect the engagement of host T cells with cells infected by exogenous HIV-1. These responses could be driven by three possible sources of viral antigen: via antigen presented on partner-derived cells within the seminal fluids, super-infection of host cells, or virion derived proteins.
The presentation of viral antigens on a partner's HLA can occur on partner-derived cells or cell-free HLA molecules within the seminal fluids [15],[16]. However, this is unlikely to be the sole mechanism for stimulating the host immune response, as this could only occur through HLA alleles that are shared by both partners. A more likely explanation involves the infection of host cells with a partner's virus.
Viral suppression within the peripheral blood by HAART has been shown not to be mirrored within the gut mucosal layer, where limited viral replication has been measured [17],[18]. Therefore, it is possible that the alleviation of drug pressure within the mucosal layer could allow a limited super-infection. Phylogenetic analysis of all the patients within this study revealed no evidence of systemic super-infection. However, this does not exclude limited or localized super-infections within the gut.
Free virus within the semen could provide enough antigenic stimuli to act as a “natural” immunogen. Proteins present in the viral particle have been shown in vitro to be sufficient to induce both cytotoxicity and IFN-γ secretion by CD8+ T cells, prior to viral replication within an infected CD4+ T cells [19],[20]. This suggests that a response could be stimulated in vivo in the absence of viral integration.
The ability of HIV-1 to induce an immune response, but remain undetectable at the peripheral level, has been shown in exposed uninfected individuals [21],[22],[23]. In particular this has been shown in health care workers who only received very limited HIV-1 exposure [24]. Furthermore, it is well established that discordant partners and sex workers, regularly exposed to HIV-1, can generate an anti-HIV-1 T cell response while remaining uninfected. However, it is noteworthy that this apparent protection can be lost after just temporarily ceasing exposure [25],[26],[27]. This suggests that HIV-1 can act as a potent immunogen, capable of generating immune responses either in the absence of infection or at levels below the current level of detection. Indeed, the site of inoculation appears to play a role too. A low dose of X4 SHIVSF33A induces potent cellular immunity via the vaginal route of inoculation [28]. Furthermore, HIV-1 chronically infected individuals have been shown to exhibit robust multifunctional CD8+ T cell responses within the rectal tissue [29],[30]. Macaques initially infected with live attenuated SIV, show protection against subsequent intrarectal challenge with a more virulent strain [31],[32],[33],[34]. This protection from super-infection was thought to involve local CD8+ T cell immunity. However, in this study only limited differences between the level of poly-functionality and total activation were observed between the VP and NVP groups. This suggests that although exposure to exogenous HIV-1 can maintain responses, it cannot alter the functionality of the T cells, although further studies are needed.
The ability of exogenous HIV-1 to shape an immune response has important implications for HIV-1 infected individuals who choose to engage in unprotected sex with other HIV-1 infected individuals. The goal of therapeutic immunization strategies in HIV-1 infected people is to maintain strong anti-HIV-1 immune responses in individuals on HAART (reviewed in [35],[36]). Here, we have observed a “natural immunization” with live infectious virus delivered to the rectal mucosa. While the maintenance of an anti-HIV-1 response may be considered a positive outcome, individuals engaging in unprotected sexual intercourse with other HIV-1 infected individuals could be at risk of super-infection, which would be particularly detrimental if the super-infecting virus carried drug resistant mutations. Furthermore, these maintained responses remained limited in functionality, suggesting the “quality” of the responses was not improved.
Although none of the study participants showed evidence for super-infection at the systemic level (in which a new virus overgrows the resident strain), we cannot rule out potential compartmentalized super-infections within the rectal tissues, or super-infections that were cleared locally. Moreover, all of the study individuals were on successful HAART (VL<50 copies/ml), which could also limit infection with exogenous HIV-1.
We propose that HIV-1 exposure can lead to infection in susceptible individuals, but also act as a potent “natural” immunogen in individuals already infected. The mechanism/s behind the immunogenicity of exposure to a partner's HIV-1 remains unclear. However, we speculate that the maintenance of the anti-HIV-1-specific T cell response most likely reflects limited super-infection within the rectal tissues. However, catching a partner's HIV-1 in the act of super-infecting an exposed partner's tissues will be a technically challenging task.
This data also reveals a more general mechanism that occurs in infectious diseases; immune responses to chronic virus infections in a host are influenced not only by the chronic virus within the host, but also by exposure to exogenous virus from without.
We selected 49 individuals from the San Francisco Positive Partners prospective couples cohort study based on specific criteria: they had been on highly active antiretroviral therapy (HAART) for over 3 months and had viral loads under the limit of quantification on a sensitive assay (<50 copies/ml), they had a co-enrolled partner who was also HIV-1 positive and who was either viremic (the lowest viral load was >8000 copies/ml) or had a viral load <50 copies/ml. All subjects were men who have sex with men (MSM).
From these study subjects two groups were created based on the partner's viral load. The first group comprised of individuals with viremic partners, viremic partner (VP) group, and the second included subjects with virologically suppressed partners (<50 copies/ml), non-viremic partner (NVP) group. The two groups did not differ in terms of age, CD4 count, time on therapy, length of time infected, or time with partner (Table 1).
Sexual exposure data were derived from self-administered questionnaires completed by both partners independently regarding sexual practices during the past three months. The instrumentation was based upon one of the few questionnaires ever developed to measure sexual behavior that has been extensively validated [37],[38]. It was adapted to the study of HIV-positive seroconcordant partnerships and extensively piloted and revised prior to the enrollment of the couples in this sub study. An exposure score was calculated from the number of times the subject reported they had had unprotected receptive anal intercourse with their partner, and the number of times their partner reported they had unprotected insertive anal intercourse with the subject. This gave an average receptive exposure score, which was multiplied by the associated risk of infection per receptive exposure for HIV-1 negative individuals through this type of exposure [13]. The insertive exposure score was calculated in a similar manner and multiplied by the associated risk of infection per exposure for unprotected insertive anal intercourse [13]. The sum of these averages gave a final exposure score. All immunological studies were performed blinded to the exposure data.
In addition to sexual exposure to an enrolled partner the same sexual exposure measures were asked of all other sexual partners in the past three months. There was no significant difference in exposure with non-enrolled partners between the VP and NVP groups (Table 2). Although the HIV-status of most non-enrolled partners was known, knowledge of treatment status was irregular, and neither self-reports nor laboratory values of viral load were available for these partners. Therefore, outside partnerships provided no additional data useful in this analysis.
The study included 49 study subjects from the San Francisco Positive Partners prospective couples study. HIV-1-positive seroconcordant sexual partnerships were enrolled in this study if they reported unprotected intercourse. All subjects, in this sub-sample were MSM, who had been HIV-1 positive for over 2 years and were currently suppressing their virus, below 50 copies/ml, while on HAART.
T cell responses were determined by IFN-γ ELIspot assay as previously described [39]. Overlapping peptides of 15–18 amino acids in length were employed, which encompassed HIV-1 consensus B (NIH) Gag, Protease, Reverse Transcription (RT), Integrase, and Nef. In addition, a non-HIV-1 viral peptide pool comprising of peptides from Cytomegalovirus, Epstein-Barr virus, and Influenza (CEF) was used as an additional control. Cryopreserved PBMCs were thawed and plated at 105 cells were well, with a final peptide concentration of 1 µg/ml. All spot numbers were normalized to numbers of IFN-γ spot-forming units (SFU) per 106 PBMCs. Spot values from medium control wells were subtracted to determine responses to each peptide. Responses were determined as either greater than two times background, or greater than 50 SFU/106 PBMCs, which ever was the higher. All experiments were conducted blinded to the individuals exposure score.
Cryopreserved PBMCs were thawed and washed with complete media. For functional assays PBMCs were incubated with 2 µg/ml peptide pools plus anti-CD28 and anti-CD49d (BD Biosciences, San Diego) at 1 µg/ml, as describe above and after 30 min Brefeldin A was added to the cultures. Cultures were left for a further 7 hours and 30 minutes at 37°C, washed and stained. Flourochrome conjugated antibodies directed against cellular molecules: CD3 (Beckman coulter), CD4, CD8, IFN-γ, and IL-2 (all BD Biosciences), were used to stain cells. In addition the activation markers CD38 and HLA-DR (both BD Biosciences), were also stained for ex vivo in conjunction with CD3 (Beckman coulter), CD4 and CD8 (both BD Biosciences). Positive controls were Staphylococcal enterotoxin B (SEB) (Sigma). Data was acquired with a LSRII (BD Biosciences), and analyzed using FlowJo software (TreeStar).
Statistical significance and graphical presentations were completed using GraphPad Prism version 4.00 for Windows, GraphPad Software, San Diego California USA, www.graphpad.com, or Microsoft Excel version 2003, Microsoft Corporation. Differences in the proportion of individuals responding to the various antigens were analyzed by Fishers Exact test (two-tailed). Statistical analysis of the response magnitudes was completed using a Mann-Whitey test. Correlations were determined by two-tailed nonparametric Spearman correlation, Spearman r-values are also given. Linear regression analysis was plotted on each correlation as a straight line. Least squares linear regression models were executed using SPSS statistical package version 11.5.
The sequence of HIV-1 reverse transcriptase and protease was performed using the TRUGENE HIV-1 RNA genotyping kit and OpenGene system software for sequence analysis (Siemens Medical Solutions Diagnostics) [40],[41]. Mixtures are designated with standard ambiguity codes when representing 30% or greater minor variant at any particular base. Viral RNA from blood plasma was extracted using the Qia-Amp viral RNA kit (Qiagen), reverse transcribed, amplified and sequenced in subjects with sufficient viremia (>100 copies/mL). Proviral DNA from PBMC was amplified after extraction using the DNeasy tissue kit (Qiagen) and sequenced in samples from subjects with low (<100 copies/mL) or undetectable (<50 copies/mL) plasma viremia. Codons 1–99 of protease and 40 through 247 of reverse transcriptase were sequenced and analyzed in all samples.
HIV-1 sequences were obtained from the HIV-1 Pol region and sequenced. All sequences were assembled using BioEdit v.7.0.4.1 and aligned with the Clustal X v1.83 sequence alignment tool. Phylogenetic analysis was done using neighbor-joining trees with bootstraps generated using Clustal X. We used 1000 random samples of sites from the alignment, drawing 1000 trees (1 from each sample) and counted how many times each grouping from the original tree occurs in the sample trees.
|
10.1371/journal.pcbi.1005894 | Modeling visual-based pitch, lift and speed control strategies in hoverflies | To avoid crashing onto the floor, a free falling fly needs to trigger its wingbeats quickly and control the orientation of its thrust accurately and swiftly to stabilize its pitch and hence its speed. Behavioural data have suggested that the vertical optic flow produced by the fall and crossing the visual field plays a key role in this anti-crash response. Free fall behavior analyses have also suggested that flying insect may not rely on graviception to stabilize their flight. Based on these two assumptions, we have developed a model which accounts for hoverflies´ position and pitch orientation recorded in 3D with a fast stereo camera during experimental free falls. Our dynamic model shows that optic flow-based control combined with closed-loop control of the pitch suffice to stabilize the flight properly. In addition, our model sheds a new light on the visual-based feedback control of fly´s pitch, lift and thrust. Since graviceptive cues are possibly not used by flying insects, the use of a vertical reference to control the pitch is discussed, based on the results obtained on a complete dynamic model of a virtual fly falling in a textured corridor. This model would provide a useful tool for understanding more clearly how insects may or not estimate their absolute attitude.
| On the basis of vision-based feedback control of optic flow occurring during insects’ flight, we developed a dynamic model that accounts for the pitch orientation and speed in plummeting flies. We compared the hoverflies’ responses with our model and showed that an optic-flow based control strategy can be used to correct the initial pitch misorientation caused by the free fall situation. To complete the model, we combined the closed-loop control of the vertical optic flow with an additional feedback control loop based on the value of the absolute pitch orientation. The need for this measurement to stabilize the pitch orientation raises the question as whether this is also the case in dipterans. After ruling out the possibility that insects may use gravity acceleration cues to control their flight, for which no experimental evidence has been found so far, we discussed the three main sensory processes possibly involved in in their ability to control their attitude. Our model provides a useful tool for studying the various sensory processes possibly involved in dipterans’ flight stabilization abilities as well as the interactions between these processes.
| Flying insects are subjected to a broad range of disturbances, for which fast, robust sensorimotor reflexes compensate. The flight stabilization performance of flies are even more impressive in view of the intrinsic aerodynamic instability of their flapping flight [1–4]. Compensating for this passive instability requires an active inner control of the wing kinematic in addition to an outer-loop system which responds to specific sensory cues (looming objects, odours and navigational cues). Discovering insects’ abilities to sense movement via optic flow crossing the compound eye or inertially via the halteres (for dipteran) is still of great interest. However, very few studies have focused so far on how flying insect sense their absolute body orientation in the three-dimensional space (attitude) with respect to a vertical reference. Oppositely, several studies have suggested that flies may lack the ability to perceive the vertical (via graviceptive cues) in order to stabilize their flight [5–7]. To address this point, we used an already designed free-fall procedure [7] with which insects can be briefly exposed to near-weightless conditions in a box lined with horizontal black and white stripes. The present study focused on the following questions:
The control of flight speed based on optic flow cues have been confirmed by ethological studies [8–10]. In the same time, in flying insects, as in helicopter, lift vector and body orientation are fixed in time and consequently flight speed and pitch orientation [1, 3, 11], and the idea that insects’ attitude may be stabilized on the basis of the optic flow has been tested successfully on a 2 degree-of-freedom flying robot [12]. A pitch rate control process has also been proposed previously to model the drosophila’s forward velocity during flight [13, 14]. Based on the existence of coupling between pitch control and optic flow regulation, we challenged the suitability of such closed-loop control compared with hoverflies subjected to an unsteady free fall situation.
It has been previously established that the fly’s auto-stabilizer involves several sensory modalities, which interact during flight. First, insect vision is based on two physical structures, compound eyes and ocelli. The fly’s photoreceptors feature a high temporal resolution giving them a great ability to detect fast motion based on contrast changes [6]. Optic flow measurement have shown that motion vision is involved in many visually guided tasks such as flight speed and altitude control, wall following, odometry and optomotor response [9]. Most of the optic flow processing is performed by compound eyes, and local contrast motion measurements are fused by lobula plate tangential cells (LPTC) responsible for detecting large field motion [15, 16]. In addition, it has been established that several groups of interneurons, including VSTCs (Vertical Sensitive Tangential Cells) [17] and HSTC (Horizontal Sensitive Tangential Cells) [18], process the various components of visual motion and in particular that they distinguish between the rotational and translational components of the optic flow with respect to the fly’s reference frame [19]. In addition, the ocelli, which are usually composed of three simple unfocused eyes forming a triangle at the top of the head [20], may be involved in the visuo-motor stabilization reflexes that maintain postural equilibrium by detecting the head’s rotational speed [21–25].
Dipteran also possess two minute dumbbell-shaped organs called halteres, which have evolved from hind-wings and beat simultaneously in anti-phase with wings. This active beating along with the campaniform sensilla provide flies with sensitivity to Coriolis forces and consequently to their own body’s angular speed [26–28]. The halteres enable the fly’s autopilot to respond to extremely abrupt changes in attitude with a latency as short as 5ms [29, 30]. In addition, insect’s hairs and antennae are sensitive to airflow during flight. Airflow sensing by the Johnston’s organs present in the antennae is known to be involved in flight speed regulation complementary to optic flow regulation [31, 32].
All in all, these sensorimotor units are mainly characterized in flies by their high temporal resolution and their low latency response [29]. Flies’ sensors are indeed highly tuned to detecting and quickly counteracting any change in their environment [6]. The combination of various sensory modalities with different bandwidth allows them to cover a wide range of dynamic perturbations.
In this study, an insect flight control model was developed, based simply on the closed-loop control of the pitch rate and the regulation of the horizontal component of the optic flow. In a first step, our model was devoid of any kind of absolute reference. The results obtained with this model simulating the fly’s response in unsteady free fall situations are compared with experimental data obtained on plummeting hoverflies in a box lined with horizontal black & white stripes. The model simulated data matched what occurred during the first few milliseconds of the insects flight, but the pitch and speed responses became highly unstable after around 0.4s. In the second step, the accuracy of the model’s predictions was greatly improved by including two additional feedback loops: one controlling the pitch rate on the basis of the absolute estimation of the pitch orientation and one controlling the lift and thrust forces on the basis of the vertical optic flow. The ability of the fly to measure its pitch orientation with respect to an absolute reference value is discussed in term of the existence of visually mediated responses such as the dorsal light response (DLR).
In a previous study on flight stabilization in plummeting hoverflies [7], we established that the flies’ crash avoidance performance depended more on visuo-motor reflexes than on gravity perception. In order to understand those reflexes more deeply, we modeled a fly’s pitch rate control system based on optic flow cues (see Figs 1 and 2) and compared the results obtained during model simulations with experimental free falling hoverflies. First, we focused on the pitch because we observed that during the period elapsing between the onset of the fall and wingbeat initiation, flies pitched down smoothly, probably because of the pin glued onto their thorax. Therefore, pitch was taken to be the main state to be controlled by the fly’s stabilizer to avoid crash. Secondly, since gravity cues do not seem to be involved in insect flight control [6, 7, 33], we assumed that fly’s flight control does not rely on any absolute vertical reference of the environment but that it is based rather on visual and inertial motion perception and compensation. We therefore based our model on previous studies on insects’ flight behaviour providing clear-cut evidence that optic flow-based control are involved during several tasks (for a review see [9]). We considered here that the forward speed was controlled by pitching-down from the nose the body and then orienting the force vector produced by flapping wings [34], as occurred in the case of the helicopter analogy [11]. The pitch rate is set so as to keep the forward optic flow constant, as found to occur in bees traveling in a textured corridor [10].
In order to parametrize the gains in the PD controller in charge of the visual optic flow process in the model (see Fig 2C), we conducted a series of experiments with hoverflies.
The parameters of the visual Proportional-Derivative controller (PDV), Kp and Kd, were estimated directly from experimental data. We first selected only the trials in which flies triggered their wingbeats in less than 150ms after the onset of the fall and were able to compensate for the fall by reaching a positive vertical speed (i.e., a lift force superior to their weight), amounting 44 experimental trials. A simulated falls was then achieved and compared with each of the selected falls as described above with several combinations of Kp, ranging from 0 to 20, and Kd, ranging from 0 to 2. A likelihood estimation (MLE) map was obtained for each fall, giving 44 maps in all, from which we extracted the average map shown in S2A Fig.
As shown in Fig 4A (top view), in the box lined up with stripes only on the lateral walls (X = −20/20), the hoverflies did not seem to express any kind of preference for a specific wall. As expected from our previous study, it can be seen from Fig 4B that no crash occured in presence of visual cues (horizontal periodic stripes) and that most of the trajectories ended with a rising flight, which confirm the ability of hoverflies to control their flight in the free fall tests. The initiation times of the wingbeats, around 100ms in average (see Fig 4C), are also coherent with our previous findings [7]. In this study, we selected only trials featuring a time to wingbeat triggering inferior to 150ms to keep a sufficient margin from the 200ms time limit, after which it is impossible for the fly to stop its fall and avoid crash onto the ground [7].
Fig 5A (dark lines) shows the time course of the mean pitch orientation around the onset of the flies’ wingbeats. Hoverflies pitched down (i.e. head downward) when falling freely, but soon after initiating their wingbeats, they were able to compensate for the misalignment of their body tilt with respect to the horizontal within 150ms. Despite the existence of significant differences in pitch orientation at flight initiation, no difference were observed in terms of the final pitch orientation or correction times between late initiation (125–150ms), medium initiation (100–125ms) or early initiation (75–100ms), which shows the robustness of the reflex response involved.
In Fig 5B (dark lines), the mean theoretical optic flow was calculated versus time during free fall and flight recovery phases. The vertical component of ω, ωzRfly, increased to around 0.04rad.s−1 (for the latest initiation group) during the actual free fall and decreased quickly to zero after initiation of the wingbeats, whereas the horizontal component ωxRFly decreased before the wingbeats was triggered and continue to decreased slightly after the initiation of the wingbeats and reached a mean steady state value of about −0.04rad.s−1 regardless of the initial conditions. This result supports the idea that hoverflies may control the optic flow in closed-loop so as to keep it constant during flight.
The results of the parameters identification, from which the parameters used during the simulations were selected, are presented in supporting information (S2 Fig). Fig 5 shows the results of 150 simulated free falls into the virtual 40cm width corridor and the parameters used.
The initial values used in simulation were determined by randomly setting a wingbeat triggering time ranging between 75 and 150ms to fit the data range (Fig 4C). The initial state of the system (i.e. the wingbeat triggering state), θPi, θ ˙ P i, Zi and VZi, was obtained by simulating a free fall without any friction and adding a passive rotation of the body to the pitch dynamics before the onset of wingbeat triggering which was modeled by a third order transfer function (see supplementary materials, S2 Fig). The model accounted successfully for the dynamics of pitch orientation and optic flow observed experimentally during the 0.2s after the wingbeats initiation (see Fig 5).
Fig 6A shows the time course of the mean acceleration produced by the hoverflies, estimated from experimental data after subtracting gravity acceleration. After wingbeat initiation, the acceleration increased immediately to a value around 10m.s−2, which is equal to gravity acceleration absolute, during about 0.1s. After this initial phase, the acceleration increased within approximately 0.1s to a value of 25–30m.s−2, representing 2.5-3 times the absolute value of gravity acceleration, followed by a slight descent phase to around 20m.s−2 at 0.4s. The force produced by flapping wings in the model was adjusted to these dynamics as shown in Fig 6A (green line). However, the acceleration estimated from 3-D trajectory data is really noisy and could result in some discrepancies between simulated and experimental data. As it can be seen from Fig 5B, the average Z position observed during the simulations shows that the model was able to counteract the fall but the values obtained did not completely match the experimental data on the hoverflies. Nor did the heave and surge speeds match experimental data: they rather showed the occurrence of instability after around 200ms (Fig 5C and 5D).
In this study, control theory was used to model the pitch stabilization process at work in hoverflies placed in free fall situation, using simple rules based on optic flow measurements previously described in navigational tasks context [9]. We proposed a model (Fig 2) based on a virtual fly falling within a textured corridor accounting for the fly’s pitch and speed during about 200ms after the onset of the insect’s wingbeats. As in previous studies [9, 10, 13, 14], we assumed that the pitch control and hence the lift and thrust force control processes rely on the closed-loop control of the pitch rate via the halteres combined with an OF-based feedback loop. In line with [35] and [13, 14], we implemented in our model a pitch rate feedback-loop mimicking the halteres via a proportional-integrator (PI) controller. Recent results have suggested that sensory cues delivered by the eyes, halteres and antennae may interact via specific actions and coupling arrangements [37–39]. The present model involves then two nested feedback-loops, one featuring fast dynamics thanks to the halteres and one featuring much slower dynamics due to the presence of a double integrator between the pitch rate and the speed of the fly (see Fig 2). The OF was defined as the ratio between the fly’s speed and its distance to the wall which was constant during the fall and did not vary conspicuously during the 0.2s analyzed during insects’ flights. The simulated OF measurements are therefore very similar to the airspeeds apart from a different scaling due to dwall. As shown in Fig 1 and described by equation of T →, the forward OF can be controlled directly by adjusting the fly’s speed VRI and thus by controlling the pitch. As shown in Fig 5, a non-null forward OF component was observed during the fall due to a passive pitching of the fly.
We simplified the model of hoverflies´ flight dynamics by neglecting any coupling between the pitch control and the other two rotational axes (roll and yaw). There were two main reasons for focusing only on the hoverflies´ pitch attitude control:
The main characteristic of the model (see Fig 2) presented here is the total absence of any kind of vertical reference for controlling the pitch in the closed-loop system. This idea was based on previous data showing the absence of graviception in dipteran’s flight control [7]. This model accounts for the fly’s transient response during a period of up to approximately 0.2s from the onset of the flapping flight. However, as shown in Fig 4B, most of the stabilizing manoeuvres in response to the free fall situation occurred within 0.2s. The ability of the optic flow model to counteract the fall without requiring any information about the insect’s absolute orientation confirms that optic flow regulation, in addition to navigation processes, may play a stabilizing role [12]. Indeed, a slight tilting of the body and hence of the lift quickly led to a involuntary translation in the environment that results in generating optic flow. Thus, actuating the wingbeats motor system to cancel the generated OF would lead to correcting the attitude. In particular, during an instable flight, the gravity acceleration would induces a permanent increase in the speed toward the ground and a simple strategy such as maintaining a constant forward optic flow will therefore intrinsically induces the pitch to decrease with respect to the horizontal and therefore a restabilization. Indeed, as shown in, the drag force experienced during free fall can be neglected, at least in the range of our experimental paradigm, reinforcing the detection of any heave acceleration by the mean of optic flow variation. In addition, previous studies [8–10] on speed regulation based on optic flow strategy validate the implementation of such feedback loops in flight control system. Future experiments would allow to better describe these sensorimotor regulation by using moving gratings on the walls of the box or a virtual reality setup [40, 41] for example.
However, some instability in the model´s responses can be clearly seen to have occured after 300-400ms in Fig 6C. This means that a closed-loop pitch control system based only on the optic flow regulation does not suffice to maintain stable flight. Despite the presence of a fast closed-loop control of the pitch rate based on the halteres, a simple PD controller cannot be fast enough to stabilize a system featuring three integrators between the measured forward optic flow and the required pitch (see Fig 2 and eqs 6 and 7). Instead of increasing the complexity of the controller COF(s), we decided to improve the model shown in Fig 2 by adding two biologically plausible feedback loops. First, we added a feedback loop controlling the lift based on the vertical flow ω z R I through a proportional-integrator controller. The integrator cancels the vertical optic flow while keeping a non-null steady lift force, thus simulating the altitude control process observed in dipteran [42, 43]. In addition, based on the existence of sensory mechanisms involved in the estimating pitch orientation such as the Dorsal Light Response and that based on an integration of the halteres’ and/or compound eyes’ signal, we added another pitch rate control loop including a Proportional-Derivative controller based on the absolute pitch orientation (Fig 7). Both additional PI and PD controllers has been set manually, gains are given in Fig 7 and all model parameters are summarized in S1 Table (supporting information). These two optic flow and pitch feedback loops combined made it possible to stabilize the simulated pitch and height in steady state (Fig 8). The ability of the improved model to stabilize the hoverflies´ attitude thanks to the addition of a closed-loop control of the pitch orientation suggests therefore a complementary control strategy involving pitch rate, pitch and optic flow measurements. With the model parameters presented here, a single pitch feedback loop would be too slow to stabilize the fly within 200ms as required by the 40cm-high box. Although, the ability of flying insects to estimate their absolute orientation (on the pitch and roll axis) still gives rise to some controversy, the model developed here would certainly provide a basis for studying these sensorimotor reflexes and the coupling that may exists between the sensory modalities involved. It is worth noting that [12] have established that the pitch of an aerial robot can be stabilized without any need for absolute reference value by regulating the dorsal and ventral OF of a 2 degree-of-freedom flying robot. Still our simulation gives controversial results in an unsteady situation such as free-fall recovering. Probably because in their study the rotor forces are adjusted by another control based on ventral optic flow in regard to experimental observation in bees [44]. In addition to visual motion and attitude perception, we can also assume that the fly could probably relies on others sensorimotor reflexes such as those based, for example, on the expansion of the OF [45, 46] that we did not challenge in this paper.
The hoverflies rely probably on specific sensory channels to estimate its absolute attitude with respect to its environment (i.e. a vertical reference) and to control their attitude as shown by the comparison made here between the two versions of the present model. Although our setup did not include any salient cues such an artificial horizon, the light from above may stimulate the DLR [22, 47, 48] which could help insects to estimate their attitude. However, a significant improvement in the hoverflies’ ability to stop falling before crashing have been previously observed when the insects were placed in a striped box rather than uniform white environment [7]. The reflex controlling the lift force orientation may not therefore relies solely on a pitch feedback based on the position of the brightest part of the visual field. A combination between optic flow based and DLR-based control may possibly be involved. However, the exact role of the DLR and its contribution to the visual-driven stabilization of insects’ flight is still an open question.
It is proposed in the future to investigate more closely how a pitch orientation could be estimated by hoverflies’ sensory system. An argument supporting the idea that pitch estimation is involved in the hoverflies’ response to free fall situations is the relative independence of the responses in regard to the initial conditions. In contrary, our model was found to be over-shooting in short initiation times (75-100ms) and under-shooting in long initiation times (125-150ms). To study the processes that can underlie the insects’ pitch orientation estimation, we can start with some hypotheses:
The accuracy of these four hypotheses still remains to be determined, along with the question as to whether any vertical information is really carried by one or more of these processes combined. The comparison with the optic flow strategy made here should help to determine how these various channels combined may serve to maintain a stable attitude during flight.
|
10.1371/journal.pcbi.1005929 | Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening | This work introduces a number of algebraic topology approaches, including multi-component persistent homology, multi-level persistent homology, and electrostatic persistence for the representation, characterization, and description of small molecules and biomolecular complexes. In contrast to the conventional persistent homology, multi-component persistent homology retains critical chemical and biological information during the topological simplification of biomolecular geometric complexity. Multi-level persistent homology enables a tailored topological description of inter- and/or intra-molecular interactions of interest. Electrostatic persistence incorporates partial charge information into topological invariants. These topological methods are paired with Wasserstein distance to characterize similarities between molecules and are further integrated with a variety of machine learning algorithms, including k-nearest neighbors, ensemble of trees, and deep convolutional neural networks, to manifest their descriptive and predictive powers for protein-ligand binding analysis and virtual screening of small molecules. Extensive numerical experiments involving 4,414 protein-ligand complexes from the PDBBind database and 128,374 ligand-target and decoy-target pairs in the DUD database are performed to test respectively the scoring power and the discriminatory power of the proposed topological learning strategies. It is demonstrated that the present topological learning outperforms other existing methods in protein-ligand binding affinity prediction and ligand-decoy discrimination.
| Conventional persistent homology neglects chemical and biological information during the topological abstraction and thus has limited representational power for complex chemical and biological systems. In terms of methodological development, we introduce advanced persistent homology approaches for the characterization of small molecular structures which can capture subtle structural difference. We also introduce electrostatic persistent homology to embed physics in topological invariants. These approaches encipher physics, chemistry and biology, such as hydrogen bonds, electrostatics, van der Waals interactions, hydrophobicity and hydrophilicity, into topological fingerprints which, although cannot literally recast into physical interpretations, are ideally suitable for machine learning, particularly deep learning, rendering topological learning algorithms. In terms of applications, we construct a structure-based virtual screening model which outperforms other existing methods. This competitive model on the DUD database is derived by assessing the performance of a comprehensive collection of topological approaches proposed in this work and introduced in our earlier work, on the PDBBind database. The topological features constructed in this work can readily be applied to other biomolecular problems where the characterization of proteins or small molecules is needed.
| Arguably, machine learning has become one of the most important developments in data science and artificial intelligence. With its ability to extract features of various levels hierarchically, deep convolutional neural networks (CNNs) have made breakthroughs in image processing, video, audio, and computer vision [1, 2], whereas recurrent neural networks have found success in analyzing sequential data, such as text and speech [3–6]. Deep learning algorithms are able to automatically extract high-level features and discover intricate patterns in large data sets. In general, one of the major advantages of machine learning algorithms is their ability to deal with large and diverse data sets and uncover complicated relationships.
Recently, machine learning has become an indispensable tool in biomolecular data analysis and structural bioinformatics. Almost every computational problem in molecular biophysics and biology, such as the predictions of solvation free energy, solubility, partition coefficient, protein-ligand binding affinities, mutation induced protein stability change, molecular multipolar electrostatics, virtual screening, etc., has machine learning based approaches that are either parallel or complementary to their physics based counterparts. The success of deep learning has fueled the rapid growth in several areas of biological science [3, 5, 6], including bioactivity of small-molecule drugs [7–10] and genetics [11, 12], where large data sets are available.
A key component of a learning machine based on biomolecular structures is featurization, that is translating the 3D structures of biomolecules to features. While the degrees of freedom of the original biomolecular structures are large and vary among different molecules, it is almost inevitable that information loss happens with dimension reduction during featurization. Besides the choice of learning models, the performance of a predictor heavily depends on how the features are extracted. Although deep learning has been known to be powerful for the automatic extraction of features from original inputs such as images, deep learning based models directly taking biomolecules as inputs are not as competitive as the state-of-art machine learning models with carefully designed features, due to the intrinsic complexity of biomolecules [13].
Biomolecules can be characterized by geometric features, electrostatic features, high-level (residue and global level) features, and amino-acid sequence features based on physical, chemical, and biological understandings [14]. Geometric features, such as coordinates, distances, angles, surface areas [15–17] and curvatures [18–21], are important descriptors of biomolecules [22–24]. However, geometric features often involve too much structural detail and are frequently computationally intractable for large biomolecular data sets. Electrostatic features include atomic partial charges, Coulomb potentials, atomic electrostatic solvation energies, and polarizable multipolar electrostatics [25]. These descriptors become essential for highly charged biomolecular systems, such as nucleic acid polymers and some protein-ligand complexes. High-level features refer to pKa values of ionizable groups and neighborhood amino acid compositions, such as the involvement of hydrophobic, polar, positively charged, negatively charged, and special case residues. Sequence features consist of secondary structures, position-specific scoring matrix (PSSM), and co-evolution information. Sequence features and annotations provide a rich resource for bioinformatics analysis of biomolecular systems. Topology offers a new unconventional representation of biomolecules. Topology can describe biomolecules in a variety of ways [26]. Some of the most powerful topological features are obtained from multi-component persistent homology or element specific persistent homology (ESPH) [14, 27]. Recently, we carried out a comprehensive comparison of the performance of geometric features, electrostatic features, high-level features, sequence features and topological features, for the prediction of mutation induced protein folding free energy changes of four mutation data sets [14]. Surprisingly, topological features outperform all the other features [14].
Unlike geometry, topology is well known for its power of simplification to geometric complexity [28–35]. The global description generated by classical topology is based on the concept of neighborhood and connectedness. If a space can be continuously deformed to another, they are considered to have the same topological features. In this sense, topology can not distinguish between a folded protein and its unfolded form if only covalent bonds are considered. Such property prevents the use of classical topology for the characterization of biomolecular structures. Instead of using topology to describe a single configuration of connectivity, persistent homology scans over a sequence of configurations induced by a filtration parameter and renders a sequence of topological invariants, which partially captures part of geometric features. Persistent homology has been applied to biomolecular systems in our earlier works [26].
In mathematics, persistent homology is a relatively new branch of algebraic topology [29, 36]. When dealing with proteins and small molecules, it is conventional to consider atoms as point clouds. For a given point cloud data set, one type of persistent homology turns each point into a sphere with their radii systematically increasing. The corresponding topological invariants and their persistence over the varying radius values can be computed. Therefore, this method embeds multiscale geometric information in topological invariants to achieve an interplay between geometry and topology. Consequently, persistent homology captures topological structures continuously over a range of spatial scales. It is called persistent homology because at each given radius, topological invariants, i.e., Betti numbers, are practically calculated by means of homology groups. In the past decade, much theoretical formulation [37–46] and many computational algorithms [47–52] have been developed. One-dimensional (1D) topological invariants generated from persistent homology is often visualized by persistence barcodes [53, 54] and persistence diagrams [55]. In recent years, multidimensional persistence has attracted much attention [43, 56] in hope that it can better characterize the data shape when there are multiple measurements of interest.
Persistent homology has been applied to various fields, including image/signal analysis [57–62], chaotic dynamics verification [63, 64], sensor networks [65], complex networks [66, 67], data analysis [68–72], shape recognition [73–75], and computational biology [76–79]. Compared with traditional computational topology [80–82] and/or computational homology, persistent homology inherently adds an additional dimension, i.e., the filtration parameter. The filtration parameter can be used to embed important geometric or quantitative information into topological invariants. As such, the importance of retaining geometric information in topological analysis has been recognized [83], and persistent homology has been advocated as a new approach for handling big and high dimensional data sets [54, 68, 84–86]. Recently, we have introduced persistent homology for mathematical modeling and/or prediction of nano-particles, protein unfolding, and other aspects of biomolecules [26, 87]. We proposed the molecular topological fingerprint (TF) to reveal topology-function relationships in protein folding and protein flexibility [26]. We established some of the first quantitative topological analyses in our persistent homology based predictions of the curvature energy of fullerene isomers [87, 88]. We have also shown correlation between persistence barcodes and energies computed with physical models during molecular dynamics experiments [26]. Moreover, we have introduced the first differential geometry based persistent homology that utilizes partial differential equations (PDEs) in filtration [88]. Most recently, we have developed a topological representation to address additional measurements of interest, by stacking the persistent homology outputs from a sequence of frames in molecular dynamics or a sequence of different resolutions [89, 90]. We have also introduced one of the first uses of topological fingerprints for resolving ill-posed inverse problems in cryo-EM structure determination [91]. In 2015, we constructed one of the first integrations of topology and machine-learning and applied it to protein classification involving tens of thousands of proteins and hundreds of tasks [92]. We also developed persistent-homology based software for the automatic detection of protein cavities and binding pockets [93].
Despite much success, it was found that persistent homology has a limited characterization power for proteins and protein complexes, when applied directly to biomolecules [92]. Essentially, biomolecules are not only complex in their geometric constitution, but also intricate in biological constitution. In fact, the biological constitution is essential to biomolecular structure and function. Persistent homology that is designed to reduce the geometric complexity of a biomolecule neglects biological information. To overcome this difficulty, we have introduced multi-component persistent homology or element specific persistent homology (ESPH) to recognize the chemical constitution during the topological simplification of biomolecular geometric complexity [14, 27, 94]. In ESPH, the atoms of a specific set of element types in a biomolecule are selected so that specific chemical information, such as hydrophobicity or hydrophilicity, is emphasized in each selection. Our ESPH is not only able to outperform other geometric and electrostatic representations in large and diverse data sets, but is also able to shed light on the molecular mechanism of protein-ligand binding, such as the relative importance of hydrogen bond, hydrophilicity and hydrophobicity at various spatial ranges [27].
The objective of the present work is to further explore the representability and reduction power of multi-component persistent homology for biomolecules and small molecules. To this end, we take a combinatorial approach to scan a variety of element combinations and examine the characterization power of these components. Additionally, we also propose a multi-level persistence to study the topological properties of non-covalent bond interactions. This approach enables us to devise persistent homology to describe the interactions of interest between atoms that are connected by weak non-covalent bonds and delivers richer representation especially for small molecules. Moreover, realizing that electrostatics are of paramount importance in biomolecules and to enhance the power of our topological representation, we introduce electrostatic persistence, which embeds charge information in topological invariants, as a new class of features in multi-component persistent homology. The aforementioned approaches can be realized via the modification of the distance matrix with a more abstract setting, for example, Vietoris-Rips complex. The complexity reduction is guaranteed in the 1D topological representation of 3D biomolecular structures. Obviously, the multi-component persistent homology representation of biomolecule leads to a higher machine learning dimensionality compared to the original single component persistent homology for a biomolecule. Therefore, it is subject to overfitting or overlearning problem in machine learning theory. Fortunately, gradient boosting trees (GBT) method is relatively insensitive to redundant high dimensional topological features [14]. Finally, since the components can be arranged as a new dimension ordered by their feature importance, multi-component persistent homology barcodes are naturally a two-dimensional (2D) representation of biomolecules. Such a 2D representation can be easily used as image-like input data in a deep CNN architecture, with different topological dimensions, i.e., 0, 1, and 2, being treated as channels. Such a topological deep learning approach addresses the nonlinear interactions among important element combinations while keeping the information from less important ones. Barcode space metrics, such as bottleneck distance and more generally, Wasserstein distance [95, 96], offer a direct description of similarity between molecules and can be readily used with nearest neighbor regression or kernel based methods. The performance of Wasserstein distance for protein-ligand binding affinity predictions is examined in this work.
After assessing the new method’s ability to represent small molecules and protein-compound complexes, the derived model is used for virtual screening. Virtual screening computationally screens a collection of small molecules to identify those who can potentially bind to the protein target. There are mainly two types of virtual screening which are ligand-based and structure-based. Ligand-based approaches depend on a measurement of similarity among small molecules using either 2D or 3D structural information of small molecules. Structure-based approaches attempt to dock the small molecule candidate to the protein target and determine if the candidate is a potential ligand based on the top docking poses. The performance of structure-based virtual screening methods heavily depends on the quality of the docking method and the accuracy of the post-docking scoring method. Our effort focuses on the development of a topology based method for the latter part. It has been shown that using machine learning or deep learning based methods to rescore the docking poses can significantly boost the performance [97, 98]. For the models such as ensemble of trees and classical neural networks, carefully constructed features are needed. For example, a neural network based method NNScore uses a collection of derived features such as the count of hydrogen bonds and electrostatics of close contacts to describe the protein-compound complex [97]. Another class of deep learning based methods feed lower level features to deep neural networks and relies on the neural networks to automatically extract higher-level features. For example, DeepVS first computes features on each atom involved in the docking interface and feed this information to a deep neural network starting with convolution layers to hierarchically extract higher-level features [98].
The rest of this manuscript is organized as follows. Section Methods is devoted to introducing methods and algorithms. We present multi-component persistent homology, multi-level interactive persistent homology, vectorized persistent homology representation and electrostatic persistence. These formulations are crucial for the representability of persistent homology for biomolecules. Machine learning algorithms associated with the present topological data analysis are briefly discussed. Results are presented in Section Results. We first consider the characterization of small molecules. More precisely, the cross-validation of protein-ligand binding affinities prediction via solely ligand topological fingerprints is studied. We illustrate the excellent representability of our multi-component persistent homology by a comparison with a method using physics based descriptors. Additionally, we investigate the representational power of the proposed topological method on a few benchmark protein-ligand binding affinity data sets, namely, PDBBind v2007, PDBBind v2013, PDBBind v2015 and PDBBind v2016 [99]. These data sets contain thousands of protein-ligand complexes and have been extensively studied in the literature. Results indicate that multi-component persistent homology offers one of most powerful representations of protein-ligand binding systems. The aforementioned study of the characterization of small molecules and protein-ligand complexes leads to an optimal selection of features and models to be used for virtual screening. Finally, we consider the directory of useful decoys (DUD) database to examine the representability of our multi-component persistent homology for virtual screening to distinguish actives from non-actives. The DUD data set used in this work has a total of 128,374 ligand-target and decoy-target pairs containing 3961 active ligand-target pairs, and involves 40 protein targets from six families. A large number of state-of-the-art virtual screening methods have been applied to this data set. We demonstrate that the present multi-component persistent homology outperforms other methods with reported results on this benchmark. This paper ends with a conclusion.
Rational drug design and discovery have rapidly evolved into some of the most important and exciting research fields in medicine and biology. These approaches potentially have a profound impact on human health. The ultimate goal is to determine and predict whether a given drug candidate will bind to a target so as to activate or inhibit its function, which results in a therapeutic benefit to the patient. Virtual screening is an important process in rational drug design and discovery which aims to identify actives of a given target from a library of small molecules. There are mainly two types of screening techniques, ligand-based and structure-based. Ligand-based approaches depend on the similarity among small molecule candidates. Structure-based approaches try to dock a candidate molecule to the target protein and judge the candidate with the modeled binding affinity based on docking poses. Various molecular docking software packages have been developed for these purposes. Molecular docking involves both pose generation and binding affinity scoring. Currently, pose generation is quite robust while scoring power is still limited. Therefore, knowledge-based rescoring methods using machine learning or deep learning approaches can improve scoring accuracy [97, 98, 100]. We also apply our topological learning method as a rescoring machine to rerank the candidates based on docking poses generated by docking software.
This section explores the representational power of the proposed persistent homology methods for the prediction of protein-ligand binding affinities and the discrimination of actives and non-actives for protein targets. To this end, we use the present method to investigate three types of problems. First, we develop topological learning models for ligand based protein-ligand binding affinity predictions. This problem is designed to examine the representability of the proposed topological methods for small molecules. Then, we develop topological learning models for protein-ligand complex based binding affinity prediction. This problem enables us to understand the capability of the proposed topological learning methods for dealing with protein-ligand complexes. Finally, we examine the structure-based classification of active ligands and decoys which are highly possible to be non-actives, i.e., structure-based virtual screening (VS). The optimal selection of features and methods are determined by studying the first two applications and this finding leads to the main application studied in this work, the topological structure-based virtual screening. Computational algorithms used in this study are illustrated in Fig 1.
In this section, we address the representation of small molecules by element specific persistent homology, especially the proposed multi-level persistent homology designed for small molecules.
In this section, we develop topological representations of protein-ligand complexes.
In this section, we examine the performance of the proposed method for the main application in this paper, which is structure-based virtual screening which involves protein-compound complexes obtained by attempting to dock the candidates to the target proteins. The dataset is much larger than the two applications on protein-ligand binding affinity prediction which makes parameter tuning very time consuming. Therefore, the best performing procedures in ligand-based binding affinity prediction and protein-ligand-complex-based binding affinity prediction are applied in this virtual screening application.
We conduct several experiments on ligand based protein-ligand binding affinity prediction in this section which leads to the final models. To examine the strength and weakness of different sets of features and models, we first show a statistics fact of the S1322 data set of 7 protein clusters in Fig 2. The details of the S1322 data set is given in Section Results/Ligand based protein-ligand binding affinity prediction. All the gradient boosting trees models take the setup described in Section Methods/Machine learning algorithms/Gradient boosting trees.
Having demonstrated the representational power of the present topological learning method for characterizing small molecules, we further examine the method on the task of characterizing protein-ligand complex. Biologically, we consider the same task, i.e., the prediction of protein-ligand binding affinity, with a different approach that is based on the structural information of the protein-ligand complexes. Only gradient boosting trees and deep convolutional neural network algorithms are used in this section. All the gradient boosting trees models take the setup described in Section Methods/Machine learning algorithms/Gradient boosting trees.
In the present topological learning study, we use four versions of PDBBind core sets as our test sets. For each test set, the corresponding refined set, excluding the core set, is used as the training set.
In our final model TopVS reported in Table 6, we use topological descriptors of both protein-compound interactions and only the compounds (i.e., ligands and decoys) and take a consensus model on top of several ensemble of trees models and a deep learning model. We have also tested the behavior of our topological learning model TopVS-ML using either one of the aforementioned descriptions. The tests are done with TopVS-ML because that TopVS-DL is much more time consuming. When only topological descriptor of small molecules are used, which falls into the category of ligand-based virtual screening, an AUC of 0.81 is achieved. For the topological learning model using only the descriptions of protein-ligand interactions, an AUC of 0.77 is achieved. An AUC of 0.83 is obtained with a model combining both sets of descriptors which is better than each individual performance, suggesting that the two groups of descriptors are complementary to each other and are both important for achieving satisfactory results. The marginal improvement made by protein-compound complexes maybe due to the various docking quality. Similar situation was encountered by a deep learning method [98]. For the targets with high quality results by Autodock Vina (AUC of ADV > 0.8), the ligand-based features achieve an AUC of 0.81 and the complex-based features achieve an AUC of 0.86. On the other hand, for the targets with low quality results by Autodock Vina (AUC of ADV < 0.5), the ligand-based features achieve an AUC of 0.82 and the complex-based features achieve an AUC of 0.74. The results of these cases are listed in S1 Text, Tables H and I. This observation suggests that the performance of features describing the interactions and the geometry of protein-compounds complexes highly depends on the quality of docking results.
Our model with small molecular descriptors delivers an AUC of 0.81, which is comparably well to the other top performing methods. The performance of this model is also competitive in the regime of protein-ligand binding affinity prediction based on experimentally solved complex structures as is shown in Section Discussion/Ligand based protein-ligand binding affinity prediction. These results suggest that topology based small molecule characterization proposed in this work is potentially useful in other applications involving small molecules, such as predictions of toxicity, solubility and partition coefficient of small molecules.
Persistent homology is a relatively new branch of algebraic topology and is one of the main tools in topological data analysis. The topological simplification of biomolecular systems was a major motivation of the earlier persistent homology development [29, 36]. Persistent homology has been applied to computational biology [76, 77, 77–79], including our efforts [26, 87–91, 93]. However, the predictive power of primitive persistent homology was limited in early topological learning applications [92]. To address this challenge, we have recently introduced element specific persistent homology to retain chemical and biological information during the topological abstraction of biomolecules [14, 27, 94]. The resulting topological learning approach offers competitive predictions of protein-ligand binding affinity and mutation induced protein stability changes. However, persistent homology based approaches for small molecules have not been developed and its representability and predictive powers for the interaction of small molecules with macromolecules have not been extensively studied.
The present work further introduces multi-component persistent homology, multi-level persistent homology and electrostatic persistence for chemical and biological characterization, analysis and modeling. Multi-component persistent homology takes a combinatorial approach to create possible element specific topological representations. Multi-level persistent homology allows tailored topological descriptions of any desirable interaction in biomolecules which is especially useful for small molecules. Electrostatic persistence incorporates partial charges that are essential to biomolecules into topological invariants. These approaches are implemented via the appropriate construction of the distance matrix for filtration. The representation power and reduction power of multi-component persistent homology, multi-level persistent homology and electrostatic persistence are validated by two databases, namely PDBBind [99] and DUD [107, 108]. PDBBind involves more than 4,000 high quality protein-ligand complexes and DUD contains 128,374 compound-target pairs. Two classes of problems are used to test the proposed topological methods, including the prediction of protein-ligand binding affinities and the discrimination of active ligands from decoys (virtual screening). In both problems, we examine the representability of proposed topological learning methods on small molecules, which are somewhat more difficult to describe by persistent homology due to their chemical diversity, variability and sensitivity. Additionally, these methods are tested on their ability to handle the full protein-ligand complexes. Advanced machine learning methods, including Wasserstein metric based k-nearest neighbors (KNNs), gradient boosting trees (GBT), random forest (RF), extra trees (ET) and deep convolutional neural networks (CNN) are utilized in the present work to facilitate the proposed topological methods, rendering advanced topological learning algorithms for quantitative and qualitative biomolecular predictions. The thorough examination of the method on the prediction of binding affinity for experimentally solved protein-ligand complexes leads to a structure-based virtual screening method, TopVS, which outperforms other methods. The feature sets introduced in this work for small molecules and protein-ligand complexes can be extended to other applications such as 3D-structure based prediction of toxicity, solubility, and partition coefficient for small molecules and complex structure based prediction of protein-nucleic acid binding and protein-protein binding affinities.
The concept of persistent homology is built on the mathematical concept of homology, which associates a sequence of algebraic objects, such as abelian groups, to topological spaces. For discrete data such as atomic coordinates in biomolecules, algebraic groups can be defined via simplicial complexes, which are constructed from simplices, generalizations of the geometric notion of nodes, edges, triangles and tetrahedrons to arbitrarily high dimensions. Homology characterizes the topological connectivity of geometric objects in terms of topological invariants, i.e., Betti numbers, which are used to distinguish topological spaces by counting k-dimensional holes. Betti-0, Betti-1 and Betti-2, respectively, represent independent components, rings and cavities in a physical sense. In persistent homology, the generators in the homology groups are tracked along with a filtration parameter, such as the radius of a ball or the level set of a hypersurface function, that continuously varies over a range of values. Therefore, persistent homology is induced by the filtration. For a given biomolecule, the change and the persistence of topological invariants over the filtration offer a unique characterization. These concepts are very briefly discussed below. For more detailed theory and algorithms, the interested readers are referred to a book on computational topology [117].
The development of persistent homology was motivated by its potential in the dimensionality reduction, abstraction and simplification of biomolcular complexity [36]. In the early applications of persistent homology to biomolecules, emphasis was given on major or global features (long-persistent features) to derive descriptive tools. For example, persistent homology was used to identify the tunnel in a Gramicidin A channel [36] and to study membrane fusion [118]. For the predictive modeling of biomolecules, features of a wide range of scales might all be important to the target quantity [26]. At the global scale, the biomolecular conformation should be captured. At the intermediate scale, the smaller intra-domain cavities need to be identified. At the most local scale, the important substructures should be addressed, such as the pyrrolidine in the side chain of proline. These features of different scales can be reflected by barcodes with different centers and persistences. Therefore, applications in biomolecules can make a more exhaustive use of persistent homology [26, 87], compared to some other applications where only global features matter while most local features are mapped to noise. Earlier use of persistent homology was focused on qualitative analysis. Only recently had persistent homology been devised as a quantitative tool [26, 87]. While the aforementioned applications are descriptive and regression based analysis, we have also applied persistent homology to predictive modeling of biomolecules [92]. However, biomolecules are both structurally and biologically complex. Their geometric and biological complexities include covalent bonds, non-covalent interactions, effects of chirality, cis and trans distinctions, multi-leveled protein structures, and protein-ligand and protein-nucleic acid complexes. Covering a large range of spatial scales is not enough for a powerful model. The biological details should also be explored. We address the underlying biology and physics by modifying the distance function and selecting various sets of atoms according to element types, to describe different interactions. Some biological considerations are discussed in this section.
One important issue is how to protect chemical and biological information during the topological simplification. As mentioned earlier, one should not treat different types of atoms as homogeneous points in a point cloud data. To this end, element specific persistent homology or multi-component persistent homology has been proposed to retain biological information in topological analysis [14, 27, 94]. The element selection is similar to a predefined vertex color configuration for graphs.
When all atoms are passed to persistent homology algorithms, the information extracted mainly reflects the overall geometric arrangement of a biomoelcule at different spatial scales. By passing only atoms of certain element types or of certain roles to the persistent homology analysis, different types of interactions or geometric arrangements can be revealed. In protein-ligand binding modeling, the selection of all carbon atoms characterizes the hydrophobic interaction network whilst the selection of all nitrogen and/or oxygen atoms characterizes hydrophilic network and the network of potential hydrogen bonds. In the protein structural analysis, computation on all atoms can identify geometric voids inside the protein which may suggest structural instability and computation on only Cα atoms reveals the overall structure of amino acid backbones. In addition, combination of various selections of atoms based on element types provides very detailed description of the biomolecular system and the hidden relationships from the structure to function can then be learned by machine learning algorithms. This may lead to the discovery of important interactions not realized as a prior. This can be realized by passing the set of atoms of the selected element types to the persistent homology computation. This concept is used with the various definitions of distance matrix discussed as follows.
Biomolecular systems are not only complex in geometry, but also in chemistry and biology. To effectively describe complex biomolecular systems, it is necessary to modify the filtration process. There are three commonly used filtrations, namely, radius filtration, distance matrix filtration, and density filtration, for biomolecules [26, 90]. A distance matrix defined with smoothed cutoff functions was proposed in our earlier work to deal with interactions within a spatial scale of interest in biomolecules [26]. In the present work, we introduce more distance matrices to enhance the representational power of persistent homology and to cover some important interactions that were not covered in our earlier works. The distance matrices can be used with a more abstract construction of simplicial complexes, such as Vietoris-Rips complex.
Barcode representation of topological invariants offers a visualization of persistent homology analysis. In machine learning analysis, we convert the barcode representation of topological invariants into structured feature arrays for machine learning. To this end, we introduce two methods, i.e., counts in bins, barcode statistics, and persistence diagram slice and statistics, to generate feature vectors from sets of barcodes. These methods are discussed below. Python code is given in S1 Code for the generation of features used in the final models in the Results section.
Three machine learning algorithms, including k-nearest neighbors (KNN) regression, gradient boosting trees and deep convolutional neural networks, are integrated with our topological representations to construct topological learning algorithms.
|
10.1371/journal.pbio.1002217 | Sustained Pax6 Expression Generates Primate-like Basal Radial Glia in Developing Mouse Neocortex | The evolutionary expansion of the neocortex in mammals has been linked to enlargement of the subventricular zone (SVZ) and increased proliferative capacity of basal progenitors (BPs), notably basal radial glia (bRG). The transcription factor Pax6 is known to be highly expressed in primate, but not mouse, BPs. Here, we demonstrate that sustaining Pax6 expression selectively in BP-genic apical radial glia (aRG) and their BP progeny of embryonic mouse neocortex suffices to induce primate-like progenitor behaviour. Specifically, we conditionally expressed Pax6 by in utero electroporation using a novel, Tis21–CreERT2 mouse line. This expression altered aRG cleavage plane orientation to promote bRG generation, increased cell-cycle re-entry of BPs, and ultimately increased upper-layer neuron production. Upper-layer neuron production was also increased in double-transgenic mouse embryos with sustained Pax6 expression in the neurogenic lineage. Strikingly, increased BPs existed not only in the SVZ but also in the intermediate zone of the neocortex of these double-transgenic mouse embryos. In mutant mouse embryos lacking functional Pax6, the proportion of bRG among BPs was reduced. Our data identify specific Pax6 effects in BPs and imply that sustaining this Pax6 function in BPs could be a key aspect of SVZ enlargement and, consequently, the evolutionary expansion of the neocortex.
| During development, neural progenitors generate all cells that make up the mammalian brain. Differences in brain size among the various mammalian species are attributed to differences in the abundance and proliferative capacity of a specific class of neural progenitors called basal progenitors. Among these, a specific progenitor type called basal radial glia is thought to have played an important role during evolution in the expansion of the neocortex, the part of the brain associated with higher cognitive functions like conscious thought and language. In the neocortex, the expression of the transcription factor Pax6 in basal progenitors is low in rodents, but high in primates, including humans. In this study, we aimed to mimic the elevated expression pattern of Pax6 seen in humans in basal progenitors of the embryonic mouse neocortex. To this end, we generated a novel, transgenic mouse line that allows sustained expression of the Pax6 gene in basal progenitors. This elevated expression resulted in an increase in the generation of basal radial glia, in the proliferative capacity of basal progenitors, and, ultimately, in the number of neurons produced. Our findings demonstrate that altering the expression of a single transcription factor from a mouse to a human-like pattern suffices to induce a primate-like proliferative behaviour in neural progenitors, which is thought to underlie the evolutionary expansion of the neocortex.
| The evolutionary expansion of the mammalian neocortex is thought to be primarily the consequence of the increasing proliferative capacity of cortical stem and progenitor cells during development [1–9]. Recent studies have progressively focused on differences between species regarding the type, abundance, and modes of division of cortical stem and progenitor cells, which are thought to contribute to the variety of shapes and sizes of the neocortex present across mammals [1–8].
A hallmark of the developing cortical wall is its apical–basal polarity, with the apical side corresponding to the ventricular surface and the basal side contacting the basal lamina [4,10]. At the onset of neurogenesis, neuroepithelial cells, the primary cortical stem cells, transform into apical radial glia (aRG) [11,12]. aRG, together with apical intermediate progenitors, constitute apical progenitors (APs), as they repeatedly undergo mitosis at the apical surface of the cortical wall [8,10]. Apical intermediate progenitors (previously called short neural precursors) undergo self-consuming division generating two neurons [13–15]. In contrast, aRG undergo self-renewing divisions, generating neurons and, more frequently, basal progenitors (BPs) that delaminate from the apical surface, leave the ventricular zone (VZ) and move basally to the subventricular zone (SVZ) [16–24].
BPs comprise basal radial glia (bRG, also called outer radial glia) and basal intermediate progenitors (bIPs) [8,10]. BPs typically undergo mitosis in the SVZ and can undergo, in principle, neurogenic (i.e., neuron-producing) or proliferative (i.e., self-amplifying) divisions, albeit with profound differences in occurrence between species [8,16–18,20–22,25–31]. bRG can be distinguished from the process-lacking bIPs by their apically and/or basally directed processes at mitosis [8,17,18,21–28,31].
Comparison of BPs in various mammalian brains has revealed key differences in their abundance and mode of cell division [1–6,8,32–34]. Thus, such differences have been reported for bIPs, which can be classified into two principal types, neurogenic and proliferative, depending on the mode of cell division (generating two neurons and two bIPs, respectively) [8]. In the mouse and rat SVZ, neurogenic bIPs constitute the vast majority of BPs (>80%) [16–18,21,22], whereas proliferative bIPs and bRG exist in only small proportions [17,28–30,35]. Moreover, mouse bRG typically undergo asymmetric self-renewing neurogenic divisions but not symmetric proliferative divisions [28].
By contrast, in mammals exhibiting an increased abundance of BPs and an enlarged SVZ, as characterized in detail in species such as ferret, macaque, and human [1,4–6,8,23,32], bIPs are mostly of the proliferative type, and bRG constitute at least half of all BPs [23–27]. Moreover, in these species, both bRG and proliferative bIPs undergo mostly symmetric proliferative rather than neurogenic divisions [23,24,31]. These self-amplifying divisions significantly increase the number of BPs residing in the SVZ, consequently leading to the expansion of the SVZ. Moreover, the SVZ of these animals comprises not only a rodent SVZ-related layer called the inner SVZ (iSVZ) but in addition a novel layer called the outer SVZ (oSVZ) [32]. Importantly, these alterations in the mode of cell division and the resulting increase in BP abundance and formation of an oSVZ have been hypothesized to be major causes underlying the expansion of the neocortex [2–6,8,32].
A key question then is how these differences in BP abundance and mode of cell division between rodents and primates are brought about at the molecular level. A candidate regulatory mechanism is the differential expression of transcription factors. Of particular interest in this regard is Pax6 (accession number: AAH36957), a paired-box transcription factor [36–39]. Several mouse and rat mutant models have demonstrated that Pax6 is required for normal aRG abundance and mode of cell division [37,40–49]. Moreover, although Pax6 mRNA levels are generally lower in BPs than APs, this down-regulation is much greater for mouse than human [50]. Consistent with this, only a minority of mouse and rat BPs (<30%) show Pax6 immunoreactivity (which is of lower level than in APs) [3,51,52], whereas the opposite is the case for primate, notably human, BPs (>80% Pax6-positive), with essentially all bRG and the majority of bIPs containing this transcription factor [3,23–27,53,54]. Together, these findings raise the possibility that the differences in Pax6 expression between rodent and primate BPs may be responsible, at least in part, for the greater abundance and proliferative or self-renewal capacity of the latter.
We therefore sought to maintain Pax6 expression specifically in newly generated BPs in order to investigate if such expression would increase the abundance of BPs, notably of bRG, and their proliferative or self-renewal capacity. Using a novel approach of conditional Pax6 expression [16,21,55], we find that sustaining elevated Pax6 levels in BP-genic mouse aRG and the BP progeny derived therefrom increases both the proportion of bRG among the newly generated BPs and the self-renewing capacity of BPs.
In mouse, the aRG subpopulation that gives rise to BPs, in contrast to self-amplifying aRG, specifically expresses Tis21, a pan-neurogenic progenitor marker [16,21,55]. Thus, as a tool towards maintaining Pax6 expression in mouse BPs, we generated a Tis21–CreERT2 knock-in mouse line. In this mouse line, exon 1 of Tis21 is replaced by CreERT2 containing a herpes simplex virus (HSV) tag at its C-terminus via homologous recombination (Fig 1A; for details, see S1 Fig), in order to limit Cre expression to Tis21-positive cells. To assess the cellular specificity of Cre expression, Tis21–CreERT2 knock-in mice were crossed with Tis21–GFP knock-in mice [16]. Immunofluorescence of the dorsolateral telencephalon of double-transgenic mice at embryonic day (E) 10.5, corresponding to the onset of Tis21 expression, and at E13.5, corresponding to the time point at which the in utero electroporations described below were conducted, showed that Cre was expressed in essentially the same cells as GFP (Fig 1B and 1C), indicating its expression selectively in the neurogenic subpopulations of cortical progenitors. Specifically, quantitation at E10.5 revealed that 97% of the cells containing nuclear Tis21–GFP were also positive for cytoplasmic Cre (Fig 1D), and no Cre was detected in Tis21–GFP-negative cells.
We next ascertained that the Tis21–CreERT2 mouse exhibits tamoxifen-dependent recombination by crossing this mouse line with a conditionally activateable GFP reporter mouse line, RCE:loxP [56] (Fig 1E). In these double-transgenic mice, GFP should be expressed only when CreERT2 has been translocated from the cytoplasm into the nucleus and excised a stop cassette that prevents the transcription of the GFP mRNA; the estrogen analog tamoxifen induces such CreERT2 translocation [57]. Indeed, no GFP-positive cells were observed in the absence of tamoxifen (Fig 1G). In contrast, when treated with tamoxifen (Fig 1F), GFP fluorescence was observed throughout the double-transgenic mouse brain (Fig 1I), and GFP-positive cells were found in all layers of the embryonic neocortex (Fig 1I’). This reflected Cre recombinase activity, because no GFP expression was observed when tamoxifen was administered to RCE:loxP offspring lacking the Tis21–CreERT2 allele (Fig 1H). We conclude that Tis21–CreERT2 mouse embryos can be used to obtain tamoxifen-dependent recombination specifically in the neurogenic subpopulations of cortical progenitors.
To conditionally express Pax6 in BP-genic aRG of developing neocortex, we introduced a floxed Pax6 plasmid at midneurogenesis into APs of tamoxifen-treated Tis21–CreERT2 mouse embryos. Specifically, we generated a plasmid (referred to as Pax6-expressing plasmid) containing a constitutive promoter (CAG) followed by a membrane (GAP43)–GFP cassette flanked by two loxP sites, mouse Pax6, an internal ribosome entry site (IRES) sequence, and nuclear RFP (nRFP) (Fig 2A). Upon Cre-mediated recombination, the membrane–GFP cassette would be excised, leading to the simultaneous expression of Pax6 and nRFP. Introduction of this plasmid into APs of tamoxifen-treated Tis21–CreERT2 mouse embryos should ensure maintenance of Pax6 expression as mouse BPs arise from aRG divisions, as well as during their subsequent migration to, and function in, the SVZ. An identical plasmid but lacking the Pax6 and IRES sequences served as control (Fig 2A).
We first validated the Pax6-expressing plasmid by transfection of HEK 293T cells, a cell line in which the endogenous PAX6 gene is not expressed. Transfection with the Pax6-expressing plasmid alone resulted in GFP, but not nRFP, expression. Cotransfection of the Pax6-expressing plasmid and a Cre-expressing plasmid yielded both Pax6 and nRFP expression, whereas only nRFP expression was observed upon cotransfection of the control plasmid and the Cre-expressing plasmid (S2 Fig).
We then explored whether the Pax6-expressing plasmid could be used in Tis21–CreERT2 mouse embryos to obtain conditional Pax6 expression specifically in the neurogenic subpopulation of APs and their progeny. To this end, we used the in utero electroporation technique where an electric field is generated across the cortical wall in order to allow for the unidirectional delivery of the negatively charged plasmid DNA, injected into the ventricular lumen, into APs. Dorsolateral telencephalon of tamoxifen-pretreated (E12.5) Tis21–CreERT2 mice was electroporated with Pax6-expressing plasmid at E13.5 and analyzed at E14.5, the peak of BP generation from neurogenic aRG [22] (Fig 2B). For the ease of presentation, we shall refer to this approach from here onwards simply as conditional Pax6 expression. Analysis of the Pax6 expression pattern yielded the following observations.
First, analysis of the level of Pax6 immunoreactivity revealed that a subpopulation of cells had higher Pax6 immunoreactivity upon conditional Pax6 expression than in the control (Fig 2C and 2D). Upon closer inspection, all these highly Pax6-immunoreactive cells were RFP-positive, indicating that these cells constituted Pax6-expressing-plasmid–electroporated neurogenic APs and their progeny (Fig 2C,2D and 2F). The level of Pax6 immunoreactivity in these cells in the VZ was approximately 3-fold higher than that of the nonelectroporated APs or control-plasmid–electroporated neurogenic APs and their VZ progeny (Fig 2G), essentially all of which are known to express endogenous Pax6 [37,51,52]. In the SVZ, where mouse BPs normally down-regulate Pax6 expression [3,50–52], this difference was even greater (≈6-fold higher) (Fig 2H).
Second, the appearance of these highly Pax6-immunoreactive and RFP-positive cells upon Pax6-expressing plasmid electroporation was strictly dependent on tamoxifen pretreatment (S3 Fig). Together, these observations allow us to equate the RFP-positive cells with the cells containing Pax6 due to the electroporation. To distinguish these conditionally Pax6-expressing cells from the cells expressing Pax6 endogenously, we shall refer to them from here onwards as exogenous Pax6- (exoPax6-) expressing cells. In addition, considering the results shown in Fig 1, we conclude that these cells constitute specifically the neurogenic subpopulation of APs and their progeny, notably the aRG-derived BPs.
Third, we found that electroporation with Pax6-expressing plasmid did not affect, after 24 h, the distribution of the progeny (RFP+ cells) of the electroporated neurogenic APs between (Fig 2C–2E) and within (S4 Fig) the germinal layers (i.e., VZ and SVZ). This implies that conditional Pax6 expression in neurogenic APs and their progeny, even if this expression exceeds the normal endogenous level, does not cause any overt effects on cell migration within the first 24 h after electroporation. The finding that RFP-positive cells are similarly distributed in control and upon conditional Pax6 expression allows for a valid comparison between germinal layers of the effect of conditional Pax6 expression in subsequent experiments.
Conditional Pax6 expression in aRG has previously been found to induce apoptosis when pan-aRG Cre drivers based on Emx1 and hGFAP promoter and regulatory sequences were used. However, this phenomenon was not observed with a Cre driver based on Ngn2 expression [58], which, similar (but not identical) to Tis21 expression, is characteristic of neurogenic progenitors [59]. It was therefore important to ascertain that conditional expression of Pax6 in Tis21–CreERT2 mice would not induce apoptosis. Indeed, immunofluorescence for the apoptosis marker activated caspase-3 did not reveal any significant difference in the number of caspase-3–positive cells between the progeny of control-plasmid–and Pax6-expressing-plasmid–electroporated neurogenic APs (S5 Fig). We therefore conclude that the present approach of conditional Pax6 expression is suitable to maintain high levels of Pax6 expression specifically in neurogenic APs and their progeny, notably the aRG-derived BPs, thus recapitulating the Pax6 expression pattern observed in BPs of developing primate neocortex.
In assessing the functional consequences of sustained Pax6 expression in BPs, we sought to obtain initial clues as to the identity of the progeny of the Pax6-electroporated neurogenic APs. Using the cycling cell marker Ki67, we first investigated whether the exoPax6-expressing cells exhibited the same proportion of progenitors versus neurons as control cells (Fig 2I–2K). Whereas conditional Pax6 expression did not alter the percentage of Ki67-positive cells in the VZ, it did result in a significant increase in Ki67-positive cells in the SVZ (Fig 2K). This suggested that the conditional Pax6 expression increased the population of cycling BPs derived from electroporated aRG.
We noticed in some experiments that in both control and conditional Pax6 expression, more Ki67-positive cells were observed in the basal region of the SVZ, and in particular in the intermediate zone of the electroporated area, but not in the contralateral area nor in nonelectroporated dorsolateral telencephalon. This reflected a previously described side effect of in utero electroporation, that is, the displacement of some Pax6-positive cells towards the cortical plate [60]. Importantly, this side effect does not affect the findings described in the present study for three reasons. First, all our data are comparisons between control and conditional Pax6 expression, both of which involve identical conditions of in utero electroporation. Second, all our quantifications are confined to electroporated, RFP-positive cells, and the electroporation side effect has been reported to affect mainly nonelectroporated cells [60]. Third, our quantifications of cells in the SVZ exclude cells in the intermediate zone.
To gain further insight into a possible regulation of the cell cycle of cortical progenitors by conditional Pax6 expression, we examined specific cell cycle parameters. We first examined the effect of conditional Pax6 expression on the total cell cycle length (Tc) of neurogenic aRG by performing live imaging on E14.5 organotypic slices prepared from control or Pax6-expressing plasmid–electroporated brains. The time period between two successive aRG mitoses was taken to indicate the length of the cell cycle, Tc. In both control and conditional Pax6 expression, we observed no major difference in Tc, although there was a trend for a shorter Tc upon conditional Pax6 expression (control, 21.0 ± 3.3 h, n = 8 cells versus Pax6, 18.5 ± 1.2 h, n = 9 cells, S1 Table top).
To estimate the proportion of the progeny of control-plasmid–and Pax6-expressing-plasmid–electroporated neurogenic APs that were in S-phase, we performed pulse-labeling with the thymidine analog EdU one hour before analyzing the embryos at E14.5. This revealed that a significantly greater proportion of the exoPax6-expressing progeny than of the control progeny was in S-phase, in both the VZ and SVZ (Fig 3A–3C). Given that conditional Pax6 expression did not increase the population size of cycling APs (Fig 2K), nor alter much their Tc (S1 Table top), the increase in the proportion of cells in S-phase in the VZ (Fig 3C) likely reflected a greater share of S-phase in the AP cell cycle, rather than an increase in cycling APs as such.
To address this directly, we performed a dual pulse chase experiment as previously described [61] (see S6A Fig and Materials and Methods) in order to determine the length of S-phase. We observed a significant increase in the length of S-phase for the sum of the electroporated aRG and their progeny upon conditional Pax6 expression (S6 Fig).
We further corroborated this by analyzing the pattern of immunofluorescence of the cycling cell marker proliferating cell nuclear antigen (PCNA). Like other cycling cells, cortical progenitors in S-phase show a punctate nuclear PCNA pattern, whereas progenitors in G1 and G2 show diffuse nuclear PCNA immunoreactivity [23,52,62]. Based on punctate PCNA staining, we observed a proportion of neurogenic APs in S-phase upon control electroporation that was similar to previously published data on E14.5 Tis21-positive APs [52] (S1 Table middle). Conditional Pax6 expression, however, was found to significantly increase the percentage of PCNA-positive nuclei in the VZ that showed a punctate pattern (Fig 3D–3F), i.e., increased the proportion of neurogenic APs that were in S-phase. These findings, together with the Ki67 (Fig 2K) and EdU (Fig 3C) data, imply that conditional Pax6 expression increases the relative proportion of S-phase within the AP cell cycle.
As there was no significant difference in Tc but an increase in the proportion of cells in S-phase upon conditional Pax6 expression in Tis21-positive APs, we hypothesized that the G1-phase must have been shortened to compensate for the longer S-phase. Consistent with this hypothesis, a significantly smaller proportion of the exoPax6-expressing progeny in the VZ than of the control progeny of electroporated neurogenic APs was positive for cyclin D1, a cyclin that is expressed from mid- to late-G1 (Fig 3G–3I). To estimate the length of the G1-phase, we combined the data obtained from live imaging with the punctate PCNA staining data (S1 Table bottom). As none of the apical mitoses observed lasted for >1 h and no difference in G2 length was reported between neural progenitors [52], we assumed that the proportion of neurogenic aRG in G2- and M-phase remained unchanged upon conditional Pax6 expression. Similar to the data obtained for cyclin D1 (Fig 3I), we estimated a shorter G1-phase upon conditional Pax6 expression (control 15.6 h versus Pax6 12.8 h, S1 Table bottom).
As to BPs, the increase in the proportion of EdU-positive cells in the SVZ upon conditional Pax6 expression (Fig 3C) was consistent with that of Ki67-positive cells (Fig 2K), corroborating our conclusion that the population size of cycling BPs derived from electroporated aRG was increased under this condition. Further support for this population size increase was provided by immunofluorescence for phosphohistone H3, a marker of cells in late G2- and M-phase, which revealed a significant increase in mitotic BPs derived from electroporated aRG (Fig 3J–3L). Also in the case of BPs, conditional Pax6 expression significantly increased the relative proportion of S-phase within the cell cycle as revealed by the pattern of nuclear PCNA immunoreactivity (Fig 3D–3F), albeit not at the expense of decreasing the relative proportion of G1 (Fig 3I).
Our group previously reported a difference in S-phase length between Tis21- positive and Tis21-negative APs [52]. As Tis21-negative and Tis21-positive APs differ in the type of division (symmetric versus asymmetric) and progeny produced (APs versus BPs) [16,55], we wondered whether the increase in the relative proportion of S-phase within the cell cycle of the exoPax6-expressing APs (Fig 3F) may be indicative of an alteration in their mode of division.
To explore this possibility, we investigated the nature of the cycling BPs that were increasingly observed upon conditional Pax6 expression (Fig 2I–2K) by examining the expression of two characteristic transcription factors, Tbr2 (Fig 4A–4C) and Sox2 (Fig 4D–4F). Tbr2 is typically expressed by the differentiating progeny of Tis21-expressing aRG fated to become bIPs [22,51,52,63], whereas Sox2 expression is characteristic of aRG and bRG [23,24,26,28,29,31,64]. Upon conditional Pax6 expression, analysis for the abundance of Tbr2-positive cells revealed a significant reduction in the exoPax6-expressing progeny as compared to control (Fig 4A–4C). This reduction was largely accounted for by the decrease in Tbr2-positive cells in the SVZ, most of which presumably were bIPs (Fig 4C). Conversely, the abundance of Sox2-positive cells was higher in the exoPax6-expressing progeny as compared to the control (Fig 4D–4F). Remarkably, this increase occurred in the SVZ rather than the VZ (Fig 4F). This suggested that conditional Pax6 expression, which increased the population of BPs (Fig 2K), induced Tis21-expressing aRG to increasingly generate BPs with a radial glia-characteristic transcription factor expression (i.e., bRG), at the expense (at least relatively) of bIP production.
To directly investigate a possible effect of conditional Pax6 expression on the mode of cell division of neurogenic APs, we performed a daughter cell pair assay [65] by analyzing areas of dorsolateral telencephalon that contained only a few RFP-positive cells in the VZ 24 h after electroporation. Tbr2 immunofluorescence allowed us to distinguish three types of RFP+ daughter cell pairs: (1) Tbr2–/Tbr2– (no bIP daughter cells), (2) Tbr2+/Tbr2– (1 bIP daughter cell) and (3) Tbr2+/Tbr2+ (2 bIP daughter cells) (Fig 4G). Importantly, virtually all Tbr2– daughter cells in the VZ are likely to be radial glia, either aRG or newborn bRG, based on the following considerations. Essentially all daughter cell nuclei in the VZ were PCNA-positive (S7 Fig). This was in line with the findings that >80% and almost 90% of the progeny in the VZ that was derived from electroporated neurogenic APs were Ki67+ (Fig 2K) and Sox2+ (Fig 4F), respectively. Hence, the Tbr2– daughter cells were radial glial progenitors rather than neurons. Consistent with this, almost all cells in the mouse E14.5 VZ are cycling [52], and very few of them are newborn neurons [52].
Quantification of daughter cell pairs in the VZ showed that in the control, the majority (77%) of these pairs derived from AP divisions that had generated bIPs. Specifically, 56% of divisions were asymmetric (and presumably self-renewing) (Tbr2+/Tbr2–, Fig 4H, red), and 21% symmetric self-consuming (Tbr2+/Tbr2+, Fig 4H, green). These findings were in line with the fact that the progeny specifically of neurogenic APs was analyzed. Of note, only 23% of divisions did not generate any bIPs and hence were presumably symmetric proliferative with regard to the radial glia nature of the daughter cells (Tbr2–/Tbr2–, Fig 4H, blue). In contrast, upon conditional Pax6 expression, the majority (59%) of the daughter cell pairs were derived from neurogenic AP divisions that did not generate bIPs but radial glia (Tbr2–/Tbr2–, Fig 4H, blue). This occurred at the expense of bIP-generating divisions, that is, asymmetric self-renewing divisions (Tbr2+/Tbr2–, reduced to 32%, Fig 4H, red), and symmetric self-consuming divisions (Tbr2+/Tbr2+, reduced to 8%, Fig 4H, green).
The observations that conditional Pax6 expression increased (i) the non-bIP generating divisions (Tbr2–/Tbr2–, Fig 4H, blue) and (ii) the Sox2-positive progeny in the SVZ (Fig 4F) suggested that the former progeny increasingly consisted of newborn bRG. As bRG are known to delaminate from the ventricular surface [24–27,31,35], we explored whether the radial glia progeny in the VZ observed upon conditional Pax6 expression increasingly showed signs of delamination. To this end, we measured the distance of the ventricular-most nucleus of each Tbr2–/Tbr2– daughter cell pair from the ventricular surface (Fig 4I). In light of the observation that the mean distance of the ventricular-most nucleus of the control and exoPax6-expressing Tbr2+/Tbr2– and Tbr2+/Tbr2+ daughter cell pairs was always >40 μm (S8 Fig), whereas that of the Tbr2–/Tbr2– pairs was <26.5 μm (Fig 4I, S8 Fig), we focused our attention on the abundance of the ventricular-most nuclei of Tbr2–/Tbr2– daughter cell pairs with a distance from the ventricular surface of ≥27 μm (corresponding to >3 nuclear diameters and referred to as abventricular location [16]). Whereas only 1 of the 7 ventricular-most nuclei (14%) of the Tbr2–/Tbr2– daughter cell pairs in the control was found in an abventricular location, 7 of the 15 nuclei (47%) analyzed upon conditional Pax6 expression were abventricular (Fig 4I). This suggested that conditional Pax6 expression promoted a substantial proportion of the radial glia progeny derived from neurogenic AP divisions to delaminate from the ventricular surface, as would be expected for newborn bRG.
In species with a high abundance of bRG in the SVZ, the radial thickness of the VZ decreases concomitant with bRG generation [8,23,25–27,64]. In light of the findings described above, we investigated a possible reduction in VZ thickness upon conditional Pax6 expression by quantifying the total number of nuclei (both RFP–and RFP+) in the VZ within a 200-μm wide, electroporated region of the dorsolateral telencephalon. Indeed, we observed a significant, approximately 10%, reduction in the number of nuclei in the VZ upon conditional Pax6 expression (Fig 4J). The magnitude of this reduction was consistent with the efficiency of electroporation and the estimated increase in the proportion of the progeny of electroporated neurogenic APs that delaminated upon conditional Pax6 expression as compared to control (Fig 4H). Taken together, the findings presented so far strongly suggest that mouse neurogenic APs and their progeny that constitutively express Pax6 increasingly generate bRG at the expense of generating bIPs.
To corroborate and complement these findings, we next investigated the effect of conditional Pax6 expression on the proportion of bRG in the BP progeny of electroporated aRG. To this end, we analyzed the morphology of mitotic BPs using phosphovimentin immunofluorescence (Fig 5A–5C), which stains both the cell bodies and processes of mitotic cortical progenitors [66]. bRG characteristically extend basally and/or apically directed processes [23–29,31,35], whereas bIPs do not [17,18,21–26,28,35]. As the apically directed processes have been reported to be thinner than basal processes and may not be easily detected via phosphovimentin staining [23], we focused our analysis on basal process-bearing mitotic BPs. In the control, the vast majority (91%) of mitotic BPs were nonpolar and only a small minority (9%) extended a basal process (Fig 5C), consistent with the high abundance of bIPs and low abundance of bRG in the embryonic mouse SVZ [28,29,35]. In contrast, upon conditional Pax6 expression, we observed a more than 2-fold increase in the proportion of mitotic BPs with a basal process, i.e. of bRG within the BP population (23%, Fig 5C). These data show that, concomitant with the increase in the proportion of BPs among the aRG progeny (Fig 2K), conditional Pax6 expression more than doubled the proportion of bRG within these BPs.
As the apically-directed process of bRGs may be harder to detect via phosphovimentin immunofluorescence at mitosis [23], we next investigated the diversity of bRG morphology during interphase. To do this, we made use of the residual membrane-GFP (Fig 2A) expressed presumably due to incomplete Cre recombination (see Materials and Methods, live imaging) (Fig 5D and 5E). To distinguish bRG from migrating neurons, we stained for Sox2, which is expressed in radial glia but not in neurons. In the control, all of the bRG progeny of the electroporated neurogenic APs exhibited a basal process, and 40% of them an apically-directed process as well (Fig 5F). Upon conditional Pax6 expression, we found an increase in the proportion of bRG exhibiting both basally and apically directed processes (Fig 5E and 5F, 53%) and also observed bRG with an apically directed process only (Fig 5D and 5F, 7%). Interestingly, in the macaque, bRG with both processes and bRG with an apically directed process only have been reported to have a higher self-renewing capacity as compared to bRG with a basal process only [23]. Of note, the basal process of the bRG generated upon conditional Pax6 expression sometimes extended all the way to the pia (S9A Fig).
The bRG generated upon conditional Pax6 expression were nestin-positive (S9B Fig), could be Tbr2-negative (S9C Fig), and typically exhibited a perinuclear centrosome (S9D Fig). Furthermore, these cells underwent mitotic somal translocation, in which the cell soma moves rapidly in the basal or apical direction prior to mitosis [23,26,28,31], as revealed by live time-lapse imaging (S9E Fig).
The data presented so far show increased bRG generation upon elevating Pax6 levels in neurogenic aRG and sustaining it in the BPs derived therefrom. We sought to complement these findings by a converse, loss-of-function, approach. To this end, we investigated the proportion of mitotic (phosphovimentin-positive) bRG among BPs in the dorsolateral telencephalon of E14.5 homozygous small eye (Sey) mutant mice, which lack functional Pax6 because of a mutation that generates a premature translational stop codon (Fig 5G–5I). We found a significant reduction in the percentage of bRG as compared to littermates that have at least one copy of the Pax6 gene (Fig 5I). These data indicate that although Pax6 function is not absolutely required for bRG generation, its level of expression is crucial for determining the abundance of these cells in the developing mouse neocortex.
Ferret and primate bRG are known to undergo multiple rounds of self-renewing division [23–26,31], whereas bIPs in mouse and rat embryonic neocortex typically undergo one round of self-consuming division [16–18,20–22]. In light of the increase in cycling BPs (Fig 2K) and bRG (Fig 5C) upon conditional Pax6 expression, it was therefore of interest to investigate whether conditional Pax6 expression would subsequently lead to increased cell cycle re-entry of the BP progeny derived from electroporated aRG. To this end, a single pulse of EdU was administered at 24 h after electroporation and analyzed after an additional 24 h for the proportion of cycling, Ki67-positive cells among the EdU-labeled progeny of electroporated APs, in order to identify cells that had re-entered the cell cycle (Fig 6A).
In the control, 75% of such daughter cells present in the VZ, but only 23% of such daughter cells in the SVZ, had re-entered the cell cycle (Fig 6B and 6D). In contrast, upon conditional Pax6 expression, whereas daughter cell cycle re-entry was the same in the VZ, it nearly doubled in the SVZ (41%, Fig 6C and 6D). Again, we used the residual membrane-GFP fluorescence to determine the morphology of daughter cells that had re-entered the cell cycle. Two types of such daughter cells were observed, monopolar cells with a distinct basal process (Fig 6E), i.e., bRG, and multipolar cells with short extensions during interphase (Fig 6F), presumably bIPs.
We extended these data by analyzing the fate of the progeny derived from divisions of bRG, using live time-lapse imaging for at least 48 h of organotypic slices prepared from control and Pax6-expressing plasmid-electroporated E14.5 neocortex (Fig 7A). Despite the rare occurrence of bRG in mouse neocortex, we were able to identify several RFP-positive bRG, to image their divisions, and to track their progeny for at least an additional 20 h (i.e., for a time period longer than the average Tc of self-renewing bRG (see S2 Table)). In the control, in two out of the seven cases analyzed, the mitotic bRG underwent an asymmetric self-renewing division, as one of the daughter cells was observed to re-enter the cell cycle (S10 Fig). In the other five cases, similar to what has been previously reported for the embryonic mouse brain [28], both daughters did not enter mitosis during the time of our observations (S10 Fig).
Upon conditional Pax6 expression, half of the mitotic exoPax6-expressing bRG (three out of six) gave rise to progeny that subsequently underwent another round of cell division (S10 Fig). In two cases, these bRG divisions were asymmetric self-renewing (S10 and S11 Figs, S1 Movie). Remarkably, we also observed a bRG undergoing a symmetric proliferative division (Fig 7, S10 Fig, S2 Movie), with both daughters undergoing another round of cell division. These live imaging data are consistent with the notion that bRG generated upon conditional Pax6 expression and their progeny are endowed with greater proliferative potential as compared to control. Moreover, together with the cell cycle re-entry analysis (Fig 6), these data suggest that BPs show an increased proliferative capacity upon conditional Pax6 expression.
It has been reported that a nonvertical (i.e., oblique or horizontal) cleavage plane orientation in relation to the ventricular surface of dividing APs (for examples, see Fig 8A) increases the probability that daughter cells become bRG [24,31,35,64]. We investigated whether the increased generation of bRG upon conditional Pax6 expression involved such alterations in cleavage plane orientation. In the control, the vast majority (91%) of mitotic neurogenic APs showed a vertical, and only a small minority (9%) an oblique, cleavage plane orientation (Fig 8B), consistent with previous data on Tis21-expressing APs [21,67,68]. Strikingly, conditional Pax6 expression resulted in a significant increase in nonvertical cleavage planes in mitotic neurogenic APs (19%, Fig 8B). As this doubling matched the doubling in the proportion of bRG among BPs (Fig 5C), our observations suggest that the increase in nonvertical cleavage plane orientations of neurogenic APs upon conditional Pax6 expression (Fig 8B) causally contributed to the increased generation of bRG (Fig 4F and 4H, Fig 5C).
The doubling in cell cycle re-entry of BPs upon conditional Pax6 expression (Fig 6D) matched the doubling of bRG (Fig 5C), which in primates are endowed with constitutive cell cycle re-entry capacity [23,24,26]. However, the morphology of the BPs that had re-entered the cell cycle (Fig 6E and 6F, Fig 7) raised the possibility that the increased cell cycle re-entry of BPs upon conditional Pax6 expression (Fig 6D) may not only be due to the increase in the proportion of bRG (Fig 5C) but may in addition reflect an increased cell cycle re-entry of bIPs (Fig 6F, Fig 7). Moreover, inducing mouse Tbr2-positive BPs to re-enter the cell cycle by forced premature expression of the transcription factor Insm1 has been shown to be associated with an alteration in their cleavage planes from the normal near-random [21,24,63] to mostly horizontal orientations [63]. Finally, not only human bRG are thought to divide preferentially with a near-horizontal cleavage plane [24] but also Tbr2-positive progenitors in the human SVZ, which, like their macaque counterpart [23] and in contrast to mouse bIPs, are endowed with proliferative capacity [26] and show a near-horizontal cleavage plane orientation in the majority of cases [25]. These considerations prompted us to investigate whether conditional Pax6 expression, concomitant with increasing the cell cycle re-entry of the mouse BPs derived from electroporated aRG, would perhaps increase the proportion of horizontal cleavages of these BPs (for examples of vertical, oblique, and horizontal BP cleavage planes in relation to the ventricular surface, see Fig 8C).
In the control, the BP progeny derived from neurogenic aRG showed a random cleavage plane orientation at mitosis (Fig 8D), consistent with previously published data [21,63]. Interestingly, conditional Pax6 expression caused an increase (albeit not statistically significant) in the proportion of the BP progeny that divided with a horizontal cleavage plane, decreasing the proportion of oblique cleavage planes (Fig 8D). Given that conditional Pax6 expression increased not only the proportion of bRG among the BP progeny derived from electroporated neurogenic aRG (Fig 5C) but also the proliferative capacity of this progeny in general (Fig 2K, Fig 6D), our cleavage plane data are consistent with the notion that a horizontal cleavage plane may be a hallmark of BPs endowed with self-renewal capacity, that is, bRG and proliferative bIPs [24,31,63].
As conditional Pax6 expression increased the proliferative capacity of the BP progeny of neurogenic aRG and the proportion of bRG among these BPs, we finally investigated the consequences for cortical neurogenesis. To this end, we administered EdU 10 h after electroporation, at ≈E14.0, i.e., around the start of exo-Pax6 expression, in order to label the neuronal progeny born at this midneurogenesis stage, followed by their analysis in the cortical wall at E17.5 (Fig 9A).
We first quantified the population size of the total progeny at E17.5. Compared to E14.5, this progeny population size was increased in both the control (1.6-fold) and upon conditional Pax6 expression (2.1-fold), with the latter increase being significantly greater than the former (Fig 9D and 9E). This indicated that conditional Pax6 expression increased the total cell output observed at E17.5.
Of note, in the control, the majority (68%) of the progeny had migrated to the cortical plate (S12A and S12D Fig). In contrast, in the case of conditional Pax6 expression, this was observed for only approximately one third (31%) of the progeny, the majority of which exhibited heterotopia in the intermediate zone (S12B and S12D Fig). Strikingly, most of the heterotopia cells had a much higher level of Pax6 immunoreactivity than those that had reached the cortical plate (S12C and S12E Fig). These observations are consistent with previous findings in Pax6-overexpressing mouse models, in which aggregates of Pax6-overexpressing cells in the developing cortical wall have been described [58]. Further characterization of the progeny exhibiting heterotopia showed that these cells were immature neurons (S13 Fig).
Next, we analyzed the neuronal fate of the progeny that had migrated to the cortical plate. To distinguish between deep-layer and upper-layer neurons, we made use of established markers, the transcription factor Tbr1, which labels layer V and VI neurons, and the transcriptional regulators Satb2 and Brn2, which label layer II–IV neurons [69]. Conditional Pax6 expression reduced the proportion of EdU-labeled Tbr1-positive neurons originating from the electroporated neurogenic aRG (Fig 9B, 9C and 9F). Conversely, the proportion of EdU-labeled neurons that expressed Satb2 was increased (Fig 9G–9I). Similarly, the percentage of Brn2-positive cells among the neuronal progeny was significantly increased upon conditional Pax6 expression (Fig 9J). Together with the overall increase in progeny observed at E17.5 (Fig 9D and 9E), these data show that conditional Pax6 expression at midneurogenesis increases the generation of upper-layer neurons. This likely reflected the increase in BP proliferative capacity (Fig 2K, Fig 6D, Fig 7, S10 Fig) and relative bRG abundance (Fig 4F, Fig 5C) upon conditional Pax6 expression.
We complemented and extended the data obtained by the conditional Pax6 expression using in utero electroporation by taking a double-transgenic approach. Specifically, we crossed the Tis21–CreERT2 mice with JoP6 mice [58]. Like the Pax6-expressing plasmid, JoP6 mice contain, under a constitutive promoter, a floxed GFP-stop cassette followed by Pax6, an IRES sequence and a reporter [58]. Upon Cre recombination induced by tamoxifen administration at E13.5 (Fig 10A), Pax6 will be expressed at elevated levels in neurogenic aRG, and this expression sustained in their progeny throughout the embryonic neocortex.
With this approach, similar to the results obtained upon the conditional Pax6 expression via in utero electroporation of Tis21–CreERT2 mouse embryos (Fig 9), upper-layer neurons as identified by Satb2 and Brn2 expression were significantly increased when compared to control littermates (Fig 10D and 10E). By contrast, deep-layer neurons as identified by Tbr1 expression were not affected (Fig 10F). Strikingly, with this double-transgenic approach (Fig 10A), we did not observe the heterotopia seen upon conditional Pax6 expression using in utero electroporation (S12B Fig). This may reflect the more standardized way in which elevated and sustained Pax6 expression is achieved in the double-transgenic embryos.
In ferret and primate neocortex, the increase in proliferating BPs is accompanied by an expansion of the SVZ in the basal direction, that is, an increase in BPs residing in the oSVZ, the key basal-most germinal layer characterized by lesser cell density [4–6,8,32]. To explore whether the Tis21–CreERT2: JoP6 double-transgenic approach resulted in an increase in BPs residing in cortical low-cell-density layers basal to the mouse SVZ proper, that is, the intermediate zone and subplate, we examined the distribution of the cell proliferation marker Ki67. Whereas there was no significant difference in the abundance of Ki67-positive cells between control (Tis21–CreERT2+/–: JoP6–/–) and Pax6-overexpressing (Tis21–CreERT2+/–: JoP6+/–) neocortex in the VZ, we observed a significant increase in Ki67-positive cells not only in the SVZ, but also in the intermediate zone and subplate of Pax6-overexpressing mouse neocortex (Fig 10B, 10C and 10G). It is therefore interesting to note that the increase in progenitors residing in cortical low-cell-density layers basal to the SVZ of embryonic mouse neocortex observed with the Tis21–CreERT2: JoP6 double-transgenic approach of Pax6 overexpression is reminiscent of one of the features of the ferret and primate oSVZ.
BPs endowed with proliferative capacity, notably bRG, are a hallmark of the developing primate neocortex [1–6,8,23,34,70]. Here we show that a single transcription factor, Pax6, when specifically sustained in the aRG-to-BP lineage, is sufficient to generate such BPs (Fig 11). Our study differs in key aspects from previous studies in which Pax6 expression was increased in APs of dorsolateral telencephalon, as the latter either did not observe or address effects on BPs [48,58,71], or obtained opposite results [46]. Specifically, increased Pax6 expression was previously found to increase the mRNA levels for the bIP marker Tbr2 in the VZ and SVZ [46]. In contrast, in the present study, we found a decrease of Tbr2-positive BPs in the SVZ and nascent BPs in the VZ upon conditional Pax6 expression. Our observations are consistent with the increased generation of proliferative BPs, notably bRG. These differences in results presumably reflect the fact that in the previous study, Pax6 expression was increased in all APs, whereas in the present study, conditional Pax6 expression was confined to Tis21-positive, that is, neurogenic and BP-genic, APs and their progeny.
Our findings have three significant implications for elucidating the evolutionary expansion of the neocortex. First, they reveal a key role of Pax6 in the generation of a primate-like SVZ, that is, of proliferative BPs from aRG. We find that the effects elicited by increased Pax6 levels on aRG mitosis and daughter cell fate in embryonic mouse neocortex reproduce the normal situation in fetal human neocortex, which is characterized by higher Pax6 levels in human than mouse aRG (S14 Fig). Specifically, increasing Pax6 levels in mouse BP-genic aRG increased their oblique cleavage plane orientation at the expense of vertical cleavage plane orientation (Fig 11), consistent with previous studies reporting a greater proportion of oblique and horizontal cleavages in human [24] than mouse [19,35,67,68,72,73] aRG. This alteration in cleavage plane orientation may well have been promoted by the fact that conditional Pax6 expression was selective for BP-genic aRG, which are more susceptible to spindle orientation variability due to the reduction of apical and basal astral microtubules as compared to proliferative aRG [67]. The increased oblique cleavage plane orientation of BP-genic aRG likely caused, in line with previous findings [35], the observed increase in (i) self-consuming bRG-genic divisions of mouse aRG at the expense of self-renewing bIP-genic divisions and (ii) aRG daughter cell delamination, and consequently (iii) the decrease in mouse VZ thickness. Taken together, our findings provide a mechanistic explanation for the reduction in VZ thickness that occurs concomitant with the growth of the oSVZ during the progression of cortical neurogenesis in species with an enlarged neocortex [8,23,25–27,32,64].
As to the mechanism how increased Pax6 levels in BP-genic aRG promote oblique cleavage plane orientation, previous work has identified an intriguing Pax6 target gene—the mitotic spindle—and kinetochore-associated protein Spag5 [47]. An increase in nonvertical cleavage plane orientation of mouse aRG has been observed both upon knock-down of Spag5 and when Spag5 mRNA and protein levels in Pax6 mutant mice at midneurogenesis are elevated [47], suggesting that either too low or too high Spag5 levels perturb the normal, horizontal spindle orientation that is required for aRG vertical cleavage plane orientation. This is in line with the concept that for aRG, a horizontal spindle orientation is thought to reflect the active state of the mitotic spindle orientation machinery, and nonhorizontal spindle orientations can occur upon perturbation of this machinery [19,35,67,72,73].
By contrast, for mouse BPs, a default state of the mitotic spindle orientation machinery, with random cleavage plane orientations, is thought to be the normal situation [21,24,63], and activation of this machinery is thought to promote horizontal cleavage plane orientation. Such orientation prevails in primate BPs, notably bRG, which are endowed with much greater proliferative capacity than mouse BPs [23–26] and is increasingly observed when mouse BPs are induced to proliferate [63]. In this context, it is interesting to note that (i) the relative Spag5 mRNA levels are much higher in the human iSVZ and oSVZ than the mouse SVZ [50], and (ii) increasing the Pax6 level in mouse BPs, which likely results in increased Spag5 levels, was found here to increase their horizontal cleavage plane orientation. Taken together, the concept emerges that Pax6, via its downstream targets including Spag5, increases oblique, self-consuming aRG divisions generating proliferative BPs, notably bRG, and horizontal BP divisions promoting their proliferation or self-renewal.
Second, we observed that sustaining high Pax6 expression in BPs increases their cell cycle re-entry (Fig 6) and their abundance not only in the SVZ but even in the layers basal to the SVZ, the intermediate zone and subplate (Fig 10). This finding implies that the maintenance of expression of Pax6 in primate, but not mouse and rat, BPs, notably bRG, is a key feature of the machinery underlying their greater proliferative or self-renewal capacity [23–26]. It thus appears that Pax6 has the potential to promote proliferation and self-renewal of cortical progenitors in general, that is, for both APs [38,39,74] and BPs (this study). Conversely, as we observed a marked decrease in bRG in the dorsolateral telencephalon of Sey mouse embryos (Fig 5G–5I), it would be interesting to explore whether a similar decrease in bRG is observed in human embryonic stem cell-derived organoids [75] upon PAX6 knockdown after establishment of the SVZ. As a corollary, the molecular mechanisms underlying the sustained Pax6 expression in BPs, at the level of mRNA and protein generation and stability [76–83], then become the crucial issue for SVZ enlargement and neocortex expansion.
The increased cell cycle re-entry of BPs observed here upon sustained Pax6 expression is in contrast to the previously reported increase in cell cycle exit of cortical progenitors in Pax6 overexpressing (PAX77) mice [48]. Again, this discrepancy presumably reflects the difference between conditional Pax6 overexpression selectively in BP-genic APs (present study) and constitutive Pax6 overexpression in all APs [48].
As to the downstream targets of Pax6 that promote BP proliferation or self-renewal, at least two candidates exist. One is the transcription factor Sox2, a well-known stimulator of stem and progenitor cell proliferation and self-renewal [38,59,84]. Pax6 has been shown to induce Sox2 expression [85], and consistent with this, we observed that sustaining Pax6 expression in the aRG–BP lineage indeed increases the proportion of Sox2-positive BPs. The other class of candidates are extracellular matrix (ECM) constituents and their receptors, the integrins, which have been implicated in BP proliferation and self-renewal [25,50,52,62,86]. Interestingly, Pax6 induces the expression of ECM constituents such as tenascin-C [87] and integrin α5β1 [88]. Hence, the increased cell cycle re-entry of BPs upon sustained Pax6 expression may well reflect, at least in part, an altered, more human-like, microenvironment in the mouse SVZ that is now more conducive to BP proliferation and self-renewal.
Third, the increased proliferative capacity of mouse BPs achieved by sustained Pax6 expression resulted in a phenotypic change in the cortical plate that is characteristic of primates—an increase in upper-layer neurons (Figs 9 and 10). Also, this aspect of the present phenotype is in contrast to previous findings which showed, concomitant with increased progenitor cell cycle exit, an increase in deep-layer neurons at the expense of upper-layer neurons in the constitutively Pax6 overexpressing PAX77 mice [48]. It should be noted that conditional Pax6 expression in neurogenic aRG resulted in an increase in Pax6 levels that was substantially greater than that in human as compared to mouse APs (compare Fig 2G and S14 Fig). Moreover, upon the present approach of conditional expression, which used a constitutive promoter, Pax6 was found to be present even in neurons (S12 Fig). It is therefore comprehensible that the present approach of conditional Pax6 expression via in utero electroporation in embryonic mouse neocortex has phenotypic consequences, some of which are not observed in fetal human neocortex and upon more controlled Pax6 expression in the double-transgenic mouse (Fig 10), such as the heterotopia which consisted mostly of highly Pax6-positive immature neurons (S12 and S13 Figs).
Hence, considering all aspects of the present phenotype together, sustaining Pax6 expression in BP-genic aRG and the BPs derived therefrom, as is characteristically the case in fetal primate neocortex [23–27,50], is sufficient to induce primate-like progenitor behaviour in embryonic mouse neocortex, that is, (i) translocation of progenitors from the VZ to the SVZ, (ii) an increased proportion of bRG among the BPs generated, (iii) sustained proliferation or self-renewal of BPs in the SVZ, and (iv) an increased upper-layer neuron production. The differential regulation of Pax6 expression in cortical progenitors during development across mammals therefore emerges as a key issue of future studies aiming to understand the evolutionary expansion of the SVZ, and consequently the neocortex.
Although sustained Pax6 expression sufficed to generate primate-like bRG in developing mouse neocortex, it was insufficient to induce cortical folding (Figs 9 and 10). This is in contrast to previous studies in which the expression of specific genes implicated in neocortex expansion led not only to the expansion of BPs but also to folding of the mouse neocortex [89,90]. In these studies, the expansion of BPs comprised an increase in both bRG and bIPs. Expansion of bIPs alone has been reported to be insufficient to induce cortical folding in the mouse neocortex [91]. Moreover, the presence of bRG is essential for tangential dispersion of neurons [27] in order for the basal surface to expand more than the apical surface, and ultimately for cortical folding [4,5,8,27]. Hence, to increase the ratio of basal to apical surface, it appears to be critical to increase the proportion of bRG among the BPs in the SVZ above a certain level. This would increase the divergence of radial fibers emanating from the SVZ, allowing for a broader dispersion of migrating neurons. Our data suggest that a mere doubling of bRG abundance in the embryonic mouse neocortex (from 10% to 20% of all BPs), as was achieved by sustaining Pax6 expression, is still insufficient to result in cortical folding.
On a more general note, human-specific aspects of neocortex expansion can be considered to be caused by (i) the presence of a relevant gene in the human as well as nonhuman genomes, but with differential regulation of expression between human and nonhuman species [50,89,92]; (ii) the presence of a relevant gene in the human as well as nonhuman genomes, but with human-specific alterations in the coding sequence [93,94]; and (iii) the presence of a relevant gene in the human, but not nonhuman, genomes [90]. The present study demonstrates that Pax6, a central player in corticogenesis, can be regarded as a key example of the first scenario.
Human fetal brain tissue was obtained from the Klinik und Poliklinik für Frauenheilkunde und Geburtshilfe, Universitätsklinikum Carl Gustav Carus of the Technische Universität Dresden, following elective pregnancy termination and informed written maternal consents, and with approval of the local University Hospital Ethical Review Committees. All human fetal brain samples were anonymized. All animal experiments were performed in accordance with the German Animal Welfare legislation (“Tierschutzgesetz”). All procedures pertaining to animal experiments were approved by the Governmental IACUC ("Landesdirektion Sachsen”) and overseen by the Institutional Animal Welfare Officer(s). Mice were anaesthetised using isofluorane during the in utero electroporation procedure. Mice were killed via cervical dislocation. The license numbers concerned by the present experiments with mice are: 24–9168.11-9/2009-2 (in utero work, tamoxifen, BrdU) and 24–9168.24-9/2012-1 (tissue collection without prior in vivo experimentation).
Mice were maintained in strict pathogen-free conditions in the animal facility of the Max Planck Institute of Molecular Cell Biology and Genetics, Dresden, Germany. To characterize the Tis21–CreERT2 mouse line described below, females were crossed with either the Tis21–GFP knock-in homozygous males [16] or the RCE:LoxP line [56]. To perform the in utero electroporation experiments with heterozygous Tis21–CreERT2 embryos as described below, homozogyous Tis21–CreERT2 males were crossed with wildtype C57BL/6JOlaHSD females. For the double-transgenic mice, homozogyous Tis21–CreERT2 mice were crossed with heterozygous JoP6 mice. To study the loss of function of Pax6 on bRG generation, heterozygous Sey mice were crossed with one another. The day of the vaginal plug was defined as E0.5. Tis21–GFP [16], hACTB–Flpe [95], RCE:LoxP [56], and JoP6 [58] mouse lines were genotyped as previously described. Offspring from the above crossings were genotyped for the Tis21–CreERT2 allele by PCR using standard procedure as described below.
To obtain pCAGGS–LoxP-GAP43-GFP-LoxP-nRFP, we first generated the intermediate plasmid pCAGGS–nRFP. RFP containing 3 C-terminal tandem SV40 nuclear localization signals (nRFP) was PCR-amplified using pDSV-mRFPnls [98] as template and the primer pair nRFP-forward & nRFP-reverse. The nRFP PCR product was cloned into the pCAGGS eukaryotic expression vector [99] opened with AgeI and EcoRI, yielding pCAGGS–nRFP. Subsequently, the LoxP-GAP43-GFP-LoxP cassette was PCR-amplified using a DFRS plasmid harboring GAP43-GFP [100] as template and the primer pair LoxP-GAP43-GFP-forward & LoxP-GAP43-GFP-reverse. The LoxP-GAP43-GFP-LoxP PCR product was cloned into the pCAGGS–nRFP vector opened with AgeI and XhoI, yielding pCAGGS–LoxP-GAP43-GFP-LoxP-nRFP (Fig 2A top).
To obtain pCAGGS–LoxP-GAP43-GFP-LoxP-Pax6-IRES-nRFP (referred to as Pax6-expressing plasmid), the Pax6 and IRES sequences were amplified from DNA constructs kindly provided by Magdalena Götz [44], using the primer pair Pax6-forward & IRES-reverse. The PCR product was cloned into the control plasmid opened with XhoI, yielding the Pax6-expressing plasmid pCAGGS–LoxP-GAP43-GFP-LoxP-Pax6-IRES-nRFP (Fig 2A bottom).
HEK293T cells were plated at 5x104 cells on 24-well plates and kept in culture in DMEM supplemented with 10% fetal calf serum. At 24 h after plating, cells were transfected, using 1 μl of Lipofectamine2000 (Invitrogen), with 250 ng of pCAGGs-Cre [101] and/or 250 ng of either control or Pax6-expressing plasmids diluted with serum-free DMEM. Cells were incubated for 48 h, followed by fixation with 4% paraformaldehyde in 120 mM phosphate buffer pH 7.4 for 10 mins. The paraformaldehyde was then removed and cells were kept in PBS until further processing.
Tamoxifen (Sigma) was dissolved in corn oil at 20 mg ml-1. Unless specified otherwise, tamoxifen was administered orally via gavage (0.1 ml) to pregnant dams carrying E12.5 embryos. This single dose was administered when animals were killed at E13.5. When animals were killed at E14.5, tamoxifen was administered at E12.5 and at E13.5 (see Fig 2B). When animals were killed at E15.5 or later, tamoxifen was administered at E12.5, at E13.5 and at E14.5 (see Fig 5D and Fig 7A). For the Tis21–CreERT2: JoP6 experiments, tamoxifen was administered orally (0.2 ml) to pregnant dams carrying E13.5 embryos.
In utero electroporation was carried out essentially as previously described [100,102]. Briefly, tamoxifen-treated pregnant dams carrying E13.5 embryos were anesthesized using isofluorane. Embryos were injected intraventricularly either with 0.5–3 mg ml-1 control or Pax6-expressing plasmid in PBS containing 0.25% Fast Green (Sigma) using a glass micropipette followed by electroporation (30 V, six 50-msec pulses with 1 sec intervals). Electroporated brains were dissected at the indicated developmental stages and fixed for 20–70 h at 4°C in 4% paraformaldehyde in phosphate buffer for further analysis.
Single EdU pulses were administered by injecting 0.1 ml of 1 mg ml-1 EdU intraperitoneally into pregnant dams carrying embryos of the indicated developmental stages.
For the cell cycle re-entry experiments, we injected such a single pulse of EdU at E14.5 and sacrificed the animals 24 h later (Fig 6A). At this developmental stage, the length of S+G2+M-phase of cortical progenitors is ≤11 h [52], and a single EdU pulse is unlikely to be effective for >5 h [103]. Hence, the 24 h period between the EdU administration and analysis should be more than sufficient for essentially all cortical progenitors that incorporated EdU and that had been derived from electroporated aRG (i.e., that were RFP+) to go through M-phase, and thus for determining by Ki67 immunofluorescence whether or not the resulting daughter cells had re-entered the cell cycle.
For the dual pulse chase experiments, 0.1 ml of 1 mg ml-1 of IdU and BrdU were sequentially injected intraperitoneally into pregnant dams carrying embryos of the indicated developmental stages (S6A Fig). The length of S-phase was calculated as described previously [61].
It has previously been shown that electroporation does not randomly target APs irrespective of the phase of the cell cycle, but preferentially targets APs in late S-, G2- and M-phase [104]. Conditional Pax6 expression upon electroporation would thus be confined to a synchronized cohort of progenitors, which precludes the use of cumulative labeling with a thymidine analog to determine the length of the cell cycle and its various phases. We therefore used live imaging to measure the cell cycle length of electroporated Tis21-positive aRG. In these analyses, we have exploited the fact that the RFP+ cells still contain residual membrane-GFP fluorescence (either by inheritance, or because not all plasmid copies electroporated into a given aRG underwent Cre recombination, or both).
Live time-lapse imaging of dorsolateral telencephalon tissue in organotypic slice culture was prepared and carried out as previously described [67]. Stacks of 1024 x 1024 pixels x 18–21 optical sections (xyzt sampling: 0.346 × 0.346 × 2.5 μm × 22 or 24 min) were acquired for at least 48 h, using a confocal laser-scanning microscope LSM 780 equipped with a 40× C-Apochromat 1.2 N.A. W objective (Carl Zeiss, Germany).
AP divisions were defined as those occurring at the ventricular surface. The time period between two successive mitoses of the neurogenic aRG is taken to be the length of the cell cycle, Tc.
In addition, we used live imaging to track the fate of the bRG progeny and for the reconstruction of the RFP-positive bRG lineage tree. We defined bRG divisions as those occurring away from the ventricular surface (with no apical contact) and as BPs exhibiting a basally and/or apically directed process just prior to, and often persisting through, mitosis. We included only RFP-positive bRG that had undergone division and tracked their progeny for at least an additional 20 h (i.e., for a time period longer than the average Tc of self-renewing bRG).
For immunofluorescence of transfected cells [65], fixed cells were permeabilised with 0.3% Triton X-100 in PBS for 30 min and then quenched with 0.1 M glycine in PBS for 30 min. Cells were sequentially incubated with primary antibodies for 3 h followed by secondary antibodies for 1 h at room temperature. Coverslips were mounted onto glass slides using Mowiol.
For vibratome sectioning [105], fixed brains were embedded in 3% low-melting agarose. Sections (50–70 μm) were cut using a vibratome (Leica 1000) and were stored in PBS (maximally for 2 wk) until further processing. For cryosectioning [105], fixed brains were equilibrated in 30% (wt/vol) sucrose in PBS overnight at 4°C. Brains were embedded with Tissue-TEK (O.C.T, Sakura Finetek) and stored at −20°C. Brains were cryosectioned at 10–12 μm. Cryosections were rehydrated with PBS before further processing. Both vibratome and cryosections were subjected to an antigen retrieval protocol as follows. Unless indicated otherwise, sections were heated in 0.01 M citrate buffer pH 6.0 at 70°C for 1 h. For comparative quantification of Pax6 and phosphohistone H3 immunofluorescence levels in mouse and human mitotic APs, cryosections of paraformaldehyde-fixed embryonic mouse and fetal human neocortex were heated in the citrate buffer using a microwave oven at 800 W for 1 min followed by 140 W for 10 min. Sections were permeabilized using 0.3% Triton X-100 in PBS for 30 min and quenched with 0.1 M glycine for 30 min. Sections were then incubated with primary antibody overnight at 4°C, followed by secondary antibody for 1 h at room temperature in a solution of 0.2% gelatin, 300 mM NaCl, and 0.3% Triton X-100 in PBS. Floating sections were mounted to Superfrost Plus microscope slides (Thermo Scientific) using Mowiol (Merck Biosciences). For BrdU and IdU detection, slices were processed after RFP immunofluorescence as follows. An additional antigen retrieval step was performed by using HCl (2 N HCL, 30 min incubation at 37°C). Slices were then blocked with 10% goat serum and incubated for 3 h at room temperature followed by 1 h of secondary antibody incubation.
The following primary antibodies were used; ßIII-tubulin (Sigma, T8578 1:500), BrdU and IdU (Becton Dickinson, 347580, 1:100), BrdU only (Abcam, ab6326, 1:100), Brn2 (Santa Cruz, SC-6029, 1:200), caspase-3 (Abcam, ab2302, 1:500), cyclinD1 (Thermo, MA1-39546, 1:200), γ-tubulin (Sigma, T5326, 1:200), GFAP (Millipore, MAB 360, 1:500), HSV tag (Abcam, ab19354, 1:200), Ki67 (Abcam, ab16667, 1:300), nestin (Abcam, AB5968, 1:200), Olig2 (Thermo, MA5-15810, 1:200), Pax6 (Covance, PRB-278P, 1:200), PCNA (Millipore, MAB424, 1:100), PH3 (Millipore, 06–570, 1:500), phosphovimentin (Abcam, ab22651, 1:500), RFP (Chemotek, 5F8, 1:500), SATB2 (Abcam, ab51502, 1:200), Sox2 (Santa Cruz, SC17320, 1:500), Tbr1 (Abcam, ab31940, 1:200), and Tbr2 (Abcam, ab23345, 1:200). Alexa Fluor 488, 594, 647 labeled secondary antibodies (Molecular Probes) were used (1:500). Nuclei were counterstained with DAPI (Sigma, 1:1,000). In case of thymidine analog-labeled samples (i.e., BrdU, EdU, and IdU), sections were postfixed with 4% paraformaldehyde for 20 min after the secondary antibody incubation. Incorporated EdU was detected using the Click-iT EdU kit with Alexa Fluor 647 (Invitrogen) as described previously [52].
Vibratome sections (50-μm) and 12-μm cryosections were used for Tbr2 and PCNA analysis, respectively. We used similar criteria as used previously [65], with some modification. In summary, we examined only sparsely electroporated areas and defined two closely located RFP-positive cells as a pair of daughter cells derived from a single electroporated AP if (i) no other RFP-positive cells were observed within the distance of one cell body around the two cells in the z-stack; (ii) both cells exhibited the same RFP fluorescence intensity; and (iii) the two RFP-positive cells were aligned in the same radial axis and were located above one another. In the case of the Tbr2–/Tbr2– daughter cell pairs, we measured the distance of the center of the nucleus of the ventricular-most daughter cell from the ventricular surface in Fiji.
Fluorescence images were acquired using a Zeiss 700 confocal microscope using 25x and 63x objectives. Images were taken as either 2.1 μm (25x) or 0.9 μm (63x) single optical sections. All images used for scoring of parameters in control versus conditional Pax6 expression had comparable RFP fluorescence intensities. All images showing these parameters for control versus conditional Pax6 expression were acquired with the same settings during each microscope session. Images taken as tile scans were stitched together using the ZEN software (Zeiss). Quantifications were performed using Fiji. Whole-brain images were acquired with an Olympus SZX12 stereomicroscope.
Cleavage plane orientation of electroporated mitotic APs and BPs was measured in 2-D based on the position of the DAPI-stained sister chromatids during late anaphase and was expressed relative to the ventricular surface. A cleavage plane parallel to the ventricular surface (i.e., horizontal cleavage plane) is defined as 0°.
Germinal zones were identified based on their different histological characteristics. The VZ was identified as the ventricular-most layer of densely packed, radially aligned, elongated nuclei. The SVZ was identified as the layer basal to the VZ containing less densely packed, randomly orientated, rounded nuclei.
Unless specified otherwise, cells were counted in a rectangular area, 200-μm wide at the ventricular surface, within the electroporated region of the dorsolateral telencephalon. For quantifications using double-transgenic animals (Tis21–CreERT2: JoP6), cells were counted in a rectangular area, 100-μm wide at the ventricular surface. Cells were counted without using pseudocolour. All quantifications were confined to RFP-positive cells only, with the exception of (i) the determination of the total nuclei present in the VZ (Fig 4J), and (ii) the analyses of the neocortex of the double-transgenic animals, (Tis21–CreERT2: JoP6); in both cases, all DAPI-stained nuclei were quantified.
For quantification of immunofluorescence intensity levels, the area of the nucleus of interphase cells in VZ and SVZ was selected using the DAPI staining as a guide, and the area of the cell body of mitotic APs was selected using the phosphohistone H3 immunofluorescence as a guide. Selected areas were quantified using Fiji [106].
Data was further processed using the Prism software (GraphPad software). Student's t test was used to determine statistical significance.
|
10.1371/journal.pgen.1003228 | Hepatocyte Growth Factor, a Determinant of Airspace Homeostasis in the Murine Lung | The alveolar compartment, the fundamental gas exchange unit in the lung, is critical for tissue oxygenation and viability. We explored hepatocyte growth factor (HGF), a pleiotrophic cytokine that promotes epithelial proliferation, morphogenesis, migration, and resistance to apoptosis, as a candidate mediator of alveolar formation and regeneration. Mice deficient in the expression of the HGF receptor Met in lung epithelial cells demonstrated impaired airspace formation marked by a reduction in alveolar epithelial cell abundance and survival, truncation of the pulmonary vascular bed, and enhanced oxidative stress. Administration of recombinant HGF to tight-skin mice, an established genetic emphysema model, attenuated airspace enlargement and reduced oxidative stress. Repair in the TSK/+ mouse was punctuated by enhanced akt and stat3 activation. HGF treatment of an alveolar epithelial cell line not only induced proliferation and scattering of the cells but also conferred protection against staurosporine-induced apoptosis, properties critical for alveolar septation. HGF promoted cell survival was attenuated by akt inhibition. Primary alveolar epithelial cells treated with HGF showed improved survival and enhanced antioxidant production. In conclusion, using both loss-of-function and gain-of-function maneuvers, we show that HGF signaling is necessary for alveolar homeostasis in the developing lung and that augmentation of HGF signaling can improve airspace morphology in murine emphysema. Our studies converge on prosurvival signaling and antioxidant protection as critical pathways in HGF–mediated airspace maintenance or repair. These findings support the exploration of HGF signaling enhancement for diseases of the airspace.
| The airspace compartment of the mammalian lung, comprised of spherical sacs termed alveoli, harbors the architecture, cellular composition, and molecular armamentarium to perform the critical function of gas exchange or oxygen uptake. Despite the necessity of this alveolar compartment for organismal viability, the mechanism by which alveoli are formed and maintained is obscure. Furthermore, no treatments are currently available that can regenerate the airspace once damaged. In this manuscript, we sought to determine whether hepatocyte growth factor, a cytokine with a functional armamentarium that subserves the critical events of alveolar formation (epithelial proliferation, migration, resistance from apoptosis and angiogenesis), could be an important mediator of alveolar formation and airspace maintenance. Our simple paradigm was that critical homeostatic pathways for the lung should operate both in lung formation and in lung maintenance/regeneration. Using an informative battery of mouse models and cell lines, we show that hepatocyte growth factor is a determinant of alveolar formation and that the enhancement of hepatocyte growth factor signaling can both protect and repair the airspace from pathologic airspace enlargement or emphysema.
| One approach to identifying mediators of alveolar formation and regeneration in the mammalian lung is to delineate the elemental events that attend airspace formation and then systematically investigate candidate proteins that harbor a compatible signaling repertoire in animal or cellular model systems. From a developmental perspective, the eruption of alveolar septae from a primordial saccule in early postnatal murine life requires localized epithelial proliferation, migration and resistance to apoptosis [1], [2]. Furthermore, epithelial morphogenesis must be accompanied by a microvasculature that permits efficient diffusion of gases from the airspace lumen to the systemic vascular bed. A candidate mediator of such events is hepatocyte growth factor (HGF). HGF is a pleiotrophic cytokine that promotes epithelial proliferation, morphogenesis, migration and survival [3], [4]. HGF is also known to induce angiogenesis and inhibit epithelial apoptosis. We sought to determine whether HGF is critical for alveolar formation and might have a therapeutic role in alveolar regeneration.
The HGF/c-Met signaling pathway incorporates all of the features of a key alveolar survival factor [5]. HGF is expressed with its receptor (c-Met or Met) in the vertebrate lung parenchyma. Upon HGF binding, c-Met undergoes autophosphorylation which initiates the recruitment of a variety of downstream signal transduction molecules (reviewed in [6]). In selective models, HGF signaling supports postpneumonectomy lung growth, induces branching morphogenesis and ameliorates inflammatory lung injury [7]–[9]. Unfortunately, these artificial models imperfectly approximate the physiologic events required for alveolar formation and maintenance. Moreover, since elevated local and systemic HGF levels in patients with lung injury correlate with disease severity and poor outcomes, whether HGF is a participant in airspace repair or a marker of ongoing injury is a subject of controversy [3]. When HGF is administered during neonatal hyperoxia or early in the course of bleomycin lung injury, some measure of protection against lung damage is observed. However, no studies have addressed whether HGF/c-Met signaling is 1) required for alveolar formation or 2) has therapeutic value in animals with established pathologic airspace enlargement. Our studies utilize both loss-of-function and gain-of-function strategies in the murine lung to investigate the function of this pathway in murine airspace formation and airspace repair. To delineate a developmental role, we present a postnatal interrogation of mice deficient in HGF signaling in the airspace epithelia. To investigate mechanisms of airspace repair, we use the TSK/+ mouse model of genetic emphysema, a convenient platform to evaluate strategies for airspace regeneration. In studies here, we both assess whether HGF infusion can improve airspace caliber in this model and evaluate downstream pathways engaged in the therapeutic response. The goal is to invoke candidate mechanisms for HGF-mediated airspace repair with possible broad therapeutic utility.
In the present study, we find that mice deficient in Met expression in alveolar epithelial cells exhibit impaired airspace morphology accompanied by a reduced abundance and survival of alveolar type II cells. We also show a reduction of vascularization in the lung parenchyma of Met-deficient mice, suggesting an intimate morphogenic connection between the epithelial and endothelial compartments. Pharmacologic augmentation of HGF signaling in a murine model of emphysema reverses airspace enlargement in the adult lung and is marked by akt and stat3 activation. Finally, in whole cell assays using primary and immortalized alveolar epithelial cells, we establish that hepatocyte growth factor signaling promotes cell survival, induces proliferation and scattering of alveolar epithelial cells, confers protection against cell death via akt activation and mediates antioxidant production. These data 1) support a critical role for HGF signaling in alveolar development and regeneration and 2) implicate downstream prosurvival signaling as a contributor to the airspace maintenance and reparative effects of HGF. Importantly, the studies suggest that developmental strategies for airspace formation are recapitulated in reparative contexts.
In order to invoke a role of the HGF/c-Met pathway in alveolar morphogenesis, we first established that the ligand and receptor are expressed in alveoli. Using immunohistochemistry, we show that c-Met is expressed in alveolar type II cells in the lungs of two- week old C57Bl/6 mice (Figure 1A). We also show that the HGF ligand is expressed diffusely in the interstitium of the alveolar septum. HGF is notably excluded from alveolar epithelial cells, consistent with its known localization and deposition in other tissues (Figure 1A and Figure S1A). Coimmunostaining for an alveolar epithelial marker (SPC) and c-Met in 2 week old lungs shows that the sites of c-Met expression are alveolar epithelial cells, airway epithelial cells and a subset of alveolar macrophages (Figure S1B).
Having established that HGF and c-Met are expressed in the developing murine lung, we investigated whether this pathway was critical for alveolarization. Yamamoto recently showed that mice with an alveolar epithelial cell specific deletion of Met had impaired late embryonic lung development, implicating the HGF/c-Met pathway in late lung development [10]. However, a dedicated analysis of the postnatal phenotype was not pursued. We generated mice deficient in Met expression in alveolar epithelial type II cells (AECII). Conditional alveolar epithelial cell specific deletion of Met was achieved by crossing the well characterized SPC-rtta;otet-Cre cassette provided by Dr. Jeffrey Whitsett into Metf/f mice that harbor an inactivating conditional deletion in exon 16 provided by Dr. Snorri Thorgeirsson [11], [12]. These tritransgenic mice (SPC-rtta/+;otet-Cre;Metf/f, termed SPCMetf/f) were treated with doxycycline from conception and harvested at two and three weeks of age. The doxycycline-treated tritransgenic mice were normal in birthweight and showed no gross extrapulmonary phenotypic or histologic abnormalities observed at 6 months. We examined activated c-Met (phosphorylated c-Met) by immunohistochemical staining in the airspace of doxycycline treated SPCMetf/f mice compared with bitransgenic or single transgenic controls. Alveolar epithelial staining for p-met was largely ablated in the SPCMetf/f mice, consistent with inducible compartmental deletion of Met (Figure 1B, 1C). Modest and marked increases in airspace caliber were seen in two and three week old doxy-treated SPCMetf/f mice, respectively, (Figure 1D, 1E). This finding suggested that c-Met expression in alveolar epithelial cells contributes to normal alveolarization.
Since HGF is a known epithelial mitogen and survival factor, we investigated whether AECII abundance was altered in the SPCMetf/f lung. By immunohistochemistry, we found reduced number of AECII cells in the 1 month old SPCMetf/f lung (Figure 1F and Figure S2). As alveolar epithelial and endothelial morphogenesis are often interdependent, we examined microvascular abundance in the mutant lung. By thrombomodulin staining and quantitative immunohistochemistry, we found a marked truncation in the pulmonary vascular bed of the mutant mice (Figure 1G, 1H). These data suggest that the architectural defects observed in the mutant mice are likely secondary to both the primary impairment in alveolar epithelial cell survival and secondary cell-nonautonomous effects on the microvascular bed.
To further parse the alveolar epithelial cell phenotype of Met-deficient mice, we assessed measures of cell survival and stress in the alveolar compartment. The distribution of airspace proliferation, as assessed by Ki67 immunostaining, in the SPCMetf/f mice compared to wild type controls was different (Figure 2A). Whereas Ki67 staining was predominantly localized to airspace epithelial cells in the wild-type lung, staining was most prominent in the alveolar macrophages of the mutant lung (Figure 2A, 2B). By contrast, TUNEL staining, reflecting parenchymal cell death, was not enhanced in the mutant lung (data not shown). Of note, this combination of reduced target cell proliferation without enhanced cell death was also observed in hepatocyte-specific Met deletion [13]. Because oxidative stress in the airspace compartment can reduce proliferation and increase airspace dimension, we examined nitrotyrosine staining in the lungs of SPCMetf/f mice compared to wild type controls. We found a more than 50% increase in nitrotyrosine staining in the mutant lungs (Figure 2C, 2D). We assessed the expression of a panel of antioxidants in the lungs of mutant mice and found no significant change in the levels of NAD(P)H: quinone oxidoreductase-1 (Nqo1), heme oxygenase 1 (Hmox1) and glutamate-cysteine ligase catalytic subunit (Gclc) but a trend towards reduction in mutant mice (Revised Figure S3). Oxidant injury often associates with inflammation, especially macrophage influx, in various models of parenchymal lung disease. We indeed found increased macrophage abundance and proliferation in the lungs of SPCMetf/f mice (Figure 2E, 2F). Thus, the loss of c-Met expression in lung epithelial cells culminates in enhanced oxidative stress, reduced epithelial cell proliferation, mononuclear inflammation in the airspace compartment and a truncated microvascular bed. These insults likely confer the reduced AECII abundance and increased airspace dimension which define the airspace phenotype.
Loss of alveolar c-Met expression does not affect extracellular matrix expression or abundance. Since airspace homeostasis requires extracellular matrix integrity and HGF is known to attenuate fibrosis in animal models, we assessed the deposition of elastin and collagen in the lungs of wild-type and Met-deficient mice. Trichrome and modified Hart's staining showed no altered deposition of collagen or elastin respectively in the lungs of mutant mice compared with age-matched controls (Figure S4A, S4B).
To establish whether HGF exerts a direct effect on epithelial activities involved in alveolar septation, we employed pharmacologic and genetic enhancement of HGF signaling in MLE12 cells, an established murine alveolar epithelial cell line. Treatment of cells with recombinant HGF induced proliferation which was more robust at higher concentrations than that seen with EGF treatment, a known epithelial mitogen (Figure 3A). Similarly, transient transfection of human MET into MLE12 cells induced a significant increase in cell proliferation at 24 h when compared with vector control (Figure 3B). Because epithelial cell scattering approximates the cellular migratory activity that is required for alveolar formation and can be mediated selectively by HGF [14], [15], we determined whether HGF could promote such behavior in MLE12 cells. Treatment of serum-starved MLE12 with recombinant HGF resulted in marked dispersion compared with media or EGF controls (Figure 3C). Of note, the total cell counts were comparable in the EGF and HGF treated cells. Hepatocyte growth factor treatment also attenuated staurosporine-induced apoptosis in MLE12 cells (Figure 3D, 3E). In order to determine whether HGF enhances cell survival of primary alveolar epithelial cells, we performed a survival analysis of isolated murine alveolar type 2 (ATII) cells from wild-type and SPCMetf/f mice. Others have shown that HGF induces proliferation of primary rat alveolar type II cells [16], [17]. We found a significant increase in the survival of wild-type AEC cells at 48 h, consistent with a prosurvival effect of HGF signaling (Figure 3F). Using primary cells, we examined whether antioxidant and antiapoptotic signaling contributed to the short-term prosurvival effect of HGF. We found a significant induction of the antioxidants Nqo1 and Gclc but no change in the expression of the antiapoptotic genes Bcl2 and Bax with HGF treatment (Figure 3G and data not shown). This battery of in vivo and whole cell studies suggested that HGF/c-Met signaling utilizes both antioxidant and antiapoptotic effects to mediate selected components of the complex series of cellular events that are needed for alveolar formation and alveolar epithelial cell survival.
Given the multiple alveolar epithelial responses conferred by HGF signaling, we queried whether augmented HGF signaling might induce a protective or reparative response in murine models of emphysema. TSK/+ mice, a spontaneously mutant strain heterozygous for a mutant allele of the matrix protein fibrillin-1 which compromises fibrillin-1 activity, are a well-accepted model of genetic emphysema. They display alveolar septation defects that evolve into overt emphysema [18]. We recently showed that the TSK/+ airspace phenotype is partially attributable to matrix-associated susceptibility to oxidative stress resulting in alveolar cell death [19]. Before proceeding with an HGF augmentation strategy, we determined whether there are alterations in HGF expression and signaling in the TSK/+ mouse model of impaired septation. Real-time PCR, ELISA analysis and immunoblotting of whole lung specimens showed no alteration of HGF and c-Met expression in the PD14 TSK/+ lung compared with age-matched controls (Table S1, Figure S5A, S5B). However, at 2 months of age, activated HGF levels were reduced in the TSK/+ lung. By immunohistochemical analysis, although we saw some regions of reduced c-Met expression there was overall no consistent reduction in c-Met expression in the PD14 TSK/+ lung (Figure S5C, top). By contrast, we observed reduced and discontinuous expression of HGF within the interstitium of the airspace compartment of TSK/+ mice (Figure S5D, bottom). Given the antifibrotic effects of HGF in rodent models and the proposed role of alveolar myofibroblasts in alveolar homeostasis, we used alpha smooth muscle actin immunohistochemistry to gauge the abundance of alveolar myofibroblasts in the TSK/+ lung. We found few myofibroblasts in the alveolar compartment of both wild-type and TSK/+ mice as well as preserved abundance of SMCs in the airway submucosa (Figure S5E). These data suggest that airspace disorders characterized by abnormal extracellular matrix composition (e.g. TSK/+ mice) may exhibit altered HGF activation and deposition and that the TSK/+ mouse is an excellent model system to examine the therapeutic effects of HGF augmentation.
Since alveolar cell specific deletion of Met compromises alveolar formation, we examined whether enhanced HGF signaling might rescue the airspace phenotype in TSK/+ mice. A subcutaneous pump containing active, recombinant human HGF, kindly provided by Drs. Ralph Schwall and Mark Merchant at Genentech, or carrier protein was inserted into adult TSK/+ mice and wild type controls. The HGF pumps delivered 50 µg/day over a two week period. We measured human HGF levels in HGF pump mice by ELISA assay, comparing intratracheal and subcutaneous pump delivery of comparable doses (50 µg/d for 3 d). A marked elevation in serum HGF was observed after pump delivery but not intratracheal delivery (Figure 4A). Immunohistochemical staining for enhanced HGF signaling in the lung as evidenced by activated c-Met (p-Met) expression showed increased staining in the airspace compartment in the HGF-treated mice (Figure 4B). We assessed airspace morphology as an index of airspace protection and repair in these mice. The TSK/+ mice treated with short-term HGF (low dose and high dose) demonstrated >17% improvement in airspace caliber (Figure 4C, 4D). We also found reduced alveolar oxidative stress by both nitrotyrosine in the HGF-treated mice suggesting that HGF is able to antagonize oxidative stress (Figure 4E, 4F). Using immunoblotting, we analyzed downstream signaling patterns that corresponded to the morphologic and oxidative stress rescue in the TSK/+ mice after HGF augmentation (Figure 4G). We found increased stat3 and akt activation in the TSK/+ lung after HGF augmentation.
Since activation of akt and stat3 associated with reparative effects of HGF on pathologic airspace enlargement, we assessed phosphoprotein activation in MLE12 cells treated with HGF. Treatment of MLE12 cells with recombinant human HGF induced activation of ERK, JNK and akt (Figure 5A). Notably, we saw no stat3 activation with HGF treatment (Figure S6A). We assessed whether akt was involved in critical prosurvival events by using staurosporine treatment of MLE12 cells to induce apoptosis. Wortmannin, a known akt inhibitor, inhibited HGF induced akt activation but not ERK activation in MLE12 cells (Figure S6B). We found that staurosporine-induced apoptosis was inhibited by HGF treatment but pretreatment with an akt inhibitor wortmannin fully blocked the protective HGF effect (Figure 5B). Taken together, the in vivo and in vitro findings not only suggest an important role for HGF/c-Met signaling in alveolar epithelial cell survival and maintenance of postnatal airspace homeostasis but also a role for prosurvival signaling cascades as mediators of these events.
Compelling experimental evidence reveals HGF as a “master” driver of organ regeneration and repair processes [20]. HGF promotes functional tissue regeneration in the liver and kidney [21], [22]. Moreover, enhanced HGF secretion accompanies various organ injuries, involving the kidney, liver, lung and heart [23]. Lung injury from bleomycin [24], liver injury from CCl4 administration and renal injury from acid administration [25] in rodent models can be attenuated by exogenous HGF treatment, suggesting that the HGF pathway is involved in conserved organ repair mechanisms [26]. In this study, we showed that alveolar epithelial cell specific HGF signaling is required for airspace homeostasis. We further showed that augmentation of HGF signaling in an established murine model of emphysema can improve airspace enlargement. Finally, a battery of in vivo and in vitro studies utilizing both gain of function and loss of function maneuvers invoke prosurvival phosphoprotein activation and antioxidant signaling as important downstream mediators of the reparative effects of HGF augmentation.
Our investigation of the loss of function phenotype attached to the cell-specific deletion of Met identifies its activation in AECII cells as an important cell-autonomous event in alveolar formation. How does HGF promote alveolarization? The major consequence of Met deletion in airspace epithelial cells is the reduced abundance of these cells in the juvenile and adult lungs, increased oxidative stress and inflammation and truncation of the vascular bed. Since whole cell data by our lab and others suggests that c-Met activation induces cell survival and enhanced migration in primary and immortalized alveolar epithelial cells, the septation defect is likely attributable to these mechanisms that converge to compromise septation [8], [27]. Recent work by Factor et al examining mice with hepatocyte- and liver-specific deletion of Met similarly demonstrated a profound cell autonomous defect in cell cycle progression, invoking an Erk-1 dependent mechanism [12]. An additional aspect revealed by our study is the truncation of the pulmonary microvasculature which accompanies loss of epithelial expression of c-Met. Yamamoto showed that mice deficient in lung epithelial VEGF-A displayed impaired pulmonary capillary formation and reduced HGF production, implicating the c-Met pathway as a critical mediator of epithelial-endothelial crosstalk in lung homeostasis [10]. Similarly, our findings of combined therapeutic and developmental effects of HGF in the lung support a critical homeostatic role for HGF signaling in the airspace that likely incorporates proliferative, migratory and morphogenic agendas and distinct prosurvival pathways.
In the lung, hepatocyte growth factor, secreted by endothelial cells, epithelial cells and interstitial fibroblasts, is sequestered in a precursor state in the extracellular matrix. Although the precise activation events are not well understood in the lung, inflammatory insults trigger the liberation of active HGF and the initiation of HGF/c-Met signaling in many tissues [28], [29]. If HGF signaling is a homeostatic mechanism needed for the maintenance of airspace morphology and cell composition, then an effective deficit in HGF signaling may occur in airspace disorders that are not marked by inflammation. We show that active HGF levels are reduced but c-Met protein levels are preserved in the TSK/+ lung compared with wild-type controls. We suspect this is secondary to defective HGF deposition and activation on the abnormal TSK/+/+ extracellular matrix evident by HGF immunostaining (Figure S2C). Accordingly, downstream HGF signaling is impaired in TSK/+ mice. Since TSK/+ mice exhibit marked postnatal oxidative stress in the lung that promotes airspace enlargement, this lack of maintained or enhanced HGF signaling may be a cause of the enhanced oxidant stress and lead directly to the airspace phenotype [19]. The improved airspace caliber resulting from HGF administration in the TSK/+ mouse suggests suboptimal HGF signaling may be a hospitable context for for airspace protection/repair with HGF administration. As stated above, reduced p-Met activation in the TSK/+ lung combined with enhanced apoptosis and oxidative stress is consistent with impairment in both proximal HGF/c-Met signaling and reparative downstream pathways. Investigators report both reduced and maintained HGF levels in the lungs of patients with COPD/emphysema [30], [31]. Interestingly, patients with acute lung injury typically have increased HGF levels in the bronchoalveolar fluid reflecting a reparative response [32]. Whether those levels are maintained as theinjury evolves is unknown. Infants with bronchopulmonary dysplasia who have reduced HGF levels typically have worse outcomes [33]. Thus, the lack of an increase in active HGF signaling in the TSK/+ lung likely reflects an impaired response to epithelial injury.
Evidence that selective cytokines which contribute to alveolar morphogenesis, such as epidermal growth factor (EGF), fibroblast growth factor 10 (FGF10), platelet derived growth factor A (PDGFA), and vascular endothelial growth factor (VEGF), have protective or therapeutic efficacy for animal models of adult airspace disorders is limited [10], [34]–[36]. This limitation largely reflects the extraepithelial effects of these cytokines that may antagonize normal lung repair (reviewed in [37]), a therapeutic requirement for the neonatal rather than adult milieu or simply absence of well-constructed preclinical trials. For example, VEGF is required for airspace formation but overexpression or pharmacologic augmentation of VEGF can have injurious effects, nicely discussed in [38]. Although without a clear role in alveolar formation, keratinocyte growth factor (KGF) is the only growth factor which has a similar functional repertoire as HGF. However, a major limitation for the therapeutic potential of KGF is the probable requirement for pre- or concurrent injury administration (protective rather than therapeutic effects) [3], [39]. In fact, since overexpression of KGF in the murine lung results in severe malformations, the protective/therapeutic window must be carefully defined in neonatal or developing mice [40]. We show that HGF administration has reparative effects in adult TSK/+ mice suggesting a more flexible therapeutic repertoire than KGF. We plan to dissect this difference in future studies.
Growth factor induced repair of epithelium may incorporate proliferative, antiapoptotic, migratory and morphogenic agendas [5], [41]. These converge to alter cellular turnover and increase cellular survival, permitting the regeneration of functional structures. We found in the TSK/+ model that stat3 and akt activation, known prosurvival mediators, correlate with the maintenance or reestablishment of airspace integrity.
HGF/c-Met signaling induces stat3 activation frequently resulting in both cellular migration and morphogenesis in a variety of cell systems [42]. Further, a loss of stat3 activation in the murine lung increases susceptibility to hyperoxic injury and overexpression of an activated stat3 confers protection [43], [44]. Similarly, akt is involved in airspace maintenance in a neonatal model of lung injury [45], [46]. HGF-induced akt activation ameliorates cigarette smoke extract induced epithelial cell death [47]. We show here that HGF treatment of immortalized murine lung alveolar epithelial cells (MLE12), an established model of AECII cells, activates akt and appears to mediate prosurvival signaling. We also show that HGF treatment of primary alveolar type II cells promotes survival and expression of antioxidants. Future efforts will focus on dissecting the interface between prosurvival signaling and antioxidant protection in the airspace compartment.
In summary, although alveolar septal loss is the most intractable functional and anatomic lesion in COPD, the molecular basis of this process remains elusive. The mitogenic, motogenic and morphogenic features of HGF make it an attractive candidate mediator of airspace repair. We propose that reduced c-Met activation and expression underlie the inadequate reparative response in the emphysematous lung. Mice deficient in Met expression in alveolar epithelial cells display compromised epithelial cell abundance, pruning of the microvascular bed and airspace enlargement. Reduced HGF expression and c-Met activation are evident in inbred mice with genetic emphysema. We have also found that pharmacologic augmentation with recombinant HGF in a murine model of emphysema results in both reduced oxidative stress in the airspace and improved airspace dimension. We define here an important homeostatic role of HGF signaling in airspace formation, maintenance and regeneration suggesting that the HGF/c-Met pathway should be explored for airspace disorders such as bronchopulmonary dysplasia and emphysema.
Adult C57Bl6 and TSK/+ mice were housed in a controlled environment and provided with standard water and chow. Animal care was in compliance with IACUC recommendations. Mice conditionally deficient in Met expression, SPC-rtta/+;otet-Cre/+;metf/f, in alveolar epithelial cells were generated by crossing Metf/f mice harboring an inactivating conditional deletion of exon 16 of the mouse Met gene [12] with bitransgenic mice expressing SPC-rtta;otet-Cre [11]. Pups resulting from these matings were produced in comparable litter sizes and the genotypes represented in Mendelian ratios. Controls were bitransgenic or single transgenic mice. The mice were housed in a facility accredited by the American Association of Laboratory Animal Care, and the animal studies were reviewed and approved by the institutional animal care and use committee of Johns Hopkins School of Medicine. To induce Cre recombinase, mice were treated with doxycycline (Sigma) at 5 mg/ml in drinking water from conception to time of harvest. Tight-skin (TSK/+) mice backcrossed into a C57Bl/6 background without the pallid allele were generated and maintained as described [19]. Mice were genotyped using standard protocols [11], [12], [48].
Recombinant HGF provided by Genentech was administered through an intraperitoneal miniosmotic pump placed under isoflurane anesthesia. The pumps were loaded with HGF with carrier in PBS or carrier in PBS alone (control) producing a total daily amount of 25–50 µg for 2 weeks. For intratracheal delivery, HGF was administered per catheter inserted into the tracheal under isoflurane anesthesia and direct inspection.
Results are expressed as means ± SEM unless otherwise stated. Comparisons between 2 experimental groups were examined using the Student T test or Mann-Whitney rank sum test. Comparisons among 3 or more groups were performed by one-way ANOVA. All statistical analyses were performed with Sigmastat (version 3.5; systat Software, Chicago, IL). A p<0.05 was considered significant.
Additional and more detailed methods are provided in Text S1.
|
10.1371/journal.ppat.0040043 | Structure–Function Aspects of PstS in Multi-Drug–Resistant Pseudomonas aeruginosa | The increasing prevalence of multi-drug–resistant (MDR) strains of Pseudomonas aeruginosa among critically ill humans is of significant concern. In the current study, we show that MDR clinical isolates of P. aeruginosa representing three distinct genotypes that display high virulence against intestinal epithelial cells, form novel appendage-like structures on their cell surfaces. These appendages contain PstS, an extracellular phosphate binding protein. Using anti-PstS antibodies, we determined that the PstS-rich appendages in MDR strains are involved in adherence to and disruption of the integrity of cultured intestinal epithelial cell monolayers. The outer surface–expressed PstS protein was also identified to be present in P. aeruginosa MPAO1, although to a lesser degree, and its role in conferring an adhesive and barrier disruptive phenotype against intestinal epithelial cells was confirmed using an isogenic ΔPstS mutant. Formation of the PstS rich appendages was induced during phosphate limitation and completely suppressed in phosphate-rich media. Injection of MDR strains directly into the intestinal tract of surgically injured mice, a known model of phosphate limitation, caused high mortality rates (60%–100%). Repletion of intestinal phosphate in this model completely prevented mortality. Finally, significantly less outer surface PstS was observed in the MPAO1 mutant ΔHxcR thus establishing a role for the alternative type II secretion system Hxc in outer surface PstS expression. Gene expression analysis performed by RT-PCR confirmed this finding and further demonstrated abundant expression of pstS analogous to pa5369, pstS analogous to pa0688/pa14–55410, and hxcX in MDR strains. Taken together, these studies provide evidence that outer surface PstS expression confers a highly virulent phenotype of MDR isolates against the intestinal epithelium that alters their adhesive and barrier disrupting properties against the intestinal epithelium.
| The resistance of bacteria to multiple antibiotics is a major problem in critically ill patients who often become colonized by highly lethal pathogens such as Pseudomonas aeruginosa. During the course of critical illness, as many as 50% of patients' intestinal tracts become colonized with P. aeruginosa, with as many as 30% of strains being resistant to multiple antibiotics. Concomitantly, critical illness is characterized by acute depletion of phosphate, which itself has been shown to be an independent predictor of infection-related mortality. In the present study we determined that during low phosphate conditions, highly virulent multi-antibiotic–resistant strains of P. aeruginosa isolated from critically ill patients produce an abundance of the phosphate-binding protein, PstS, located on extracellular finger-like structures. These PstS rich appendages participate in the binding of P. aeruginosa to intestinal lining cells and may allow P. aeruginosa to acquire phosphate from its host while remaining at arm's length from the host immune system. This clever tactic may be one example by which successful opportunistic pathogens such as P. aeruginosa survive within complex ecological niches such as the intestinal tract and harm their hosts during the course of critical illness.
| Infection due to P. aeruginosa continues to be a major cause of mortality among critically ill and immuno-compromised patients despite the development of newer and more powerful antibiotics. Both the immunoevasive nature of P. aeruginosa as well as its acquisition of multi-drug resistance makes elimination of this organism a particular challenge. Multi-drug–resistant (MDR) strains of P. aeruginosa, defined as resistant to at least three of the following antibiotics: ceftazidime, imipenem, gentamicin or ciprofloxacin, are often isolated from patients exposed to prolonged intensive care-type therapies [1]. Yet antibiotic resistance itself does not confer enhanced virulence [2], and therefore the ability to discriminate between virulent versus non-virulent phenotypes among colonizing multi-drug resistant isolates would be a major step in predicting the particular threat of a colonizing strain of P. aeruginosa. The primary site of colonization and a frequent source of subsequent infection of P. aeruginosa is the gastrointestinal tract reservoir, where as many as 50% of critically ill patients are colonized within 3 days of admission with as many as 30% of strains displaying antibiotic resistance [3]. Yet little is known about the behavior of these pathogens in this site, especially those that are multi-drug resistant. We recently screened several strains of MDR isolates from hospitalized patients and characterized their virulence against the intestinal epithelium using an in vitro model of cultured intestinal epithelial monolayers [2]. The majority of strains (60%) were found to be either attenuated or have no effect in their ability to adhere to or disrupt the integrity of the intestinal epithelium. However several strains representing three distinct genotypes, showed extremely high adherence capacity and a profound ability to disrupt the barrier function of cultured intestinal epithelial cells. These strains harbored the exoU gene, known to encode the most toxic effector protein of the type III secretion apparatus, thus possibly explaining their extreme toxicity against cultured intestinal epithelial cells. However, exoU expression is dependent on contact to host epithelial cells, and as recently shown with the exoU positive strain P. aeruginosa PA103 [4], lack of adherence leads to a loss of cytotoxicity against cultured epithelial cells [5] despite an intact exoU gene. Therefore we studied selected exoU positive MDR strains of P. aeruginosa displaying unusually high adherence and disrupting properties against the intestinal epithelium and determined whether surface structures might exist to explain their enhanced adhesiveness to cultured intestinal epithelial cells. In this report we show that these strains express previously un-described appendages that contain significant quantities of PstS, a high affinity phosphate binding protein. We characterized the structural and functional aspects of these PstS rich appendages and determined that they play a significant role in the adherence to and disruption of intestinal epithelial cells. Outer surface expression of PstS rich appendages was induced under low phosphate conditions and suppressed in high phosphate media. Lethality assays in a mouse model of gut-derived sepsis in which low phosphate conditions are known to exist, demonstrated high lethality rates that were completely abrogated when mice were supplemented with intestinal phosphate. Taken together, these data provide evidence that low phosphate conditions increase the presence of PstS rich appendages on MDR P. aeruginosa whose presence facilitates binding to the intestinal epithelium and whose expression in vivo may play a significant role in the development of gut-derived sepsis in critically ill patients.
In our previous work we screened consecutive MDR P. aeruginosa clinical isolates and identified a subset of strains that displayed a highly destructive phenotype against cultured intestinal epithelial cells (Caco-2 bbe) [2]. Among this subset of isolates, high swimming motility, increased adhesiveness to Caco-2 monolayers, and the presence of the exoU gene predicted a cytotoxic phenotype against the intestinal epithelium [2]. In the present study, we screened these highly adhesive MDR clinical isolates by their cell surface morphology using electron microscopy, and identified appendage-like structures on the surfaces of the most cytotoxic isolates (1, 13, and those of genotype 20) (Figure 1A–1D and Table 1). The identified appendages were 20 nm in diameter, up to 500 nm in length, and were visually distinct from flagella (Figure 1A and 1D). The identified appendages were not detected on any of the remaining clinical isolates of the previously reported series of strains (see Figures S1 and S2).
To identify proteins involved in the formation of the visualized appendages, surface-associated proteins were obtained by extensive vortexing of bacterial cells grown on Pseudomonas isolation agar (PIA), denatured by boiling with sample buffer and then separated by 10% Tris-glycine SDS-PAGE. Figures 2A and 2B show the presence of abundant protein bands at an approximate MW of 32 kDa from surface sheared proteins in strain MDR25 (Figure 2A, lane 2); 32 and 40 kDa bands from strain MDR1 (Figure 2A, lane 3); and a 40 kDa band from strain MDR13 (Figure 2B, lane 2). Proteins were transferred to a PVDF membrane and N-terminal sequencing of the 32 and 40 kDa proteins in strain MDR1 and the 40 kDa protein in strain MDR13 were performed with ABI-Procise cLC Protein Sequencer (Mayo Proteomics Research Center). The N-terminal peptide sequence of the 32 and 40 kDa proteins in strain MDR1 were found to be AIDPALPEYQK and EINGGGATLPQQLXQEPGV, respectively. The N-terminal peptide sequence of 40 kDa protein in strain MDR13 was identified as DINGGGATLPQQLYQ. The peptide sequences were searched with BLAST, and the best hit of the 32 kDa band sequence was found to be the PstS protein PA5369 in P. aeruginosa PAO1. The sequence AIDPALPEYQK was located to aa 25–35 on ORF PA5369. PA5369 was demonstrated to contain a cleavable type I signal peptide of 24 aa [6], therefore the 32 kDa protein in strain MDR1 might correspond to periplasmic orthologous protein PA5369 in PAO1. The best hit of the 40 kDa band sequence in both strains 1 and 13 was found to be PA55410 from P. aeruginosa PA14 (http://v2.pseudomonas.com/getAnnotation.do?locusID=PA14_55410). The sequence DINGGGATLPQQLYQ was located to aa 24–38 on ORF PA55410. Orthologous to PA55410, PA0688 protein in PAO1 was also demonstrated to contain a cleavable type I signal peptide of 23 aa, MFKRSLIAASLSVAALVSAQAMA [6], which was 100% identical to N-terminus of PA55410. Therefore the 40 kDa proteins in strains 1 and 13 might be orthologous to PA55410 in P. aeruginosa PA14 [7], the strain known to be highly virulent, and PA0688 from P. aeruginosa PAO1.
We next amplified and sequenced genes analogous to pa5369 in MDR strains 1, 13, and 25 (GenBank Accession numbers EF601157, EF601158, and EF601159). We determined them to be very conserved with few differences in nucleotide sequences that did not affect amino acid sequences which were 100% identical to PA5369 in P. aeruginosa PAO1 (http://www.pseudomonas.com/). We therefore created anti-PA5369 antibodies against the specific peptide 192–212 KEEALCKGDFRPNVNEQPGS that was chosen based on hydrophobicity, surface probability, flexibility, and antigenic index, as well as the Advanced BLAST Search for the absence of significant homology to other P. aeruginosa proteins. Antibodies were subjected to affinitive purification using the native peptide 192–212, and then used to detect appendage-like structures in the clinical isolates. We first performed immunobloting of cell surface structures using anti-192–212 peptide antibodies, now referred to as anti-PA5369 antibodies, and found high antibody affinity to cell surface proteins of strain MDR25, moderate affinity to strains MDR1 and MPAO1, and minimal affinity to strain MDR13 (Figure 2C). These data corresponded to the results of the SDS-PAGE demonstrating an abundance of the 32 kDa protein band in strain MDR25 but not in strain MDR13. The specificity of the anti-PA5369 antibody was confirmed by examining both sheared appendages and bacterial pellets in wild-type MPAO1, PA5369 mutant, and clinical strain MDR25 (Figure 2D). Results demonstrated that anti-PA5369 recognized abundant amounts of protein in both the cell pellet and sheared appendages in strain 25. In strain MPAO1, antibodies recognized an abundant amount of protein in the cell pellet but a low amount in sheared appendages. No recognizable protein in either the cell pellet or sheared surface fractions in the MPAO1 mutant ΔPA5369 was observed. ELISA assays performed with sheared proteins from different clinical isolates (Figure 2E) demonstrated the presence of highly abundant PA5369-like protein in MDR clinical strains 25, 27, and 28, all of which share the same genotype 20. Interestingly, anti-PA5369 antibodies recognized significantly lower amount of proteins in sheared fractions isolated from other clinical isolates previously shown to be unable to alter the epithelial resistance of Caco-2 monolayers (see Table 1) [2]. Another interesting finding was the presence of PA5369 in sheared fractions of MPAO1 by both ELISA (Figure 2E) and immunobloting (Figure 2D), although much lower in abundance compared to the highly adherent strains MDR25 and MDR1. PA5369 has been predicted by COG (Clusters of Orthologous Groups, http://www.pseudomonas.com/) to be PstS, a phosphate transport system substrate-binding protein whose expression in P. aeruginosa is induced at phosphate concentrations < 1 mM [8–12]. In order to determine if the formation of appendages in clinical isolate 25 was phosphate dependent, we suspended a single colony in 10% glycerol, and plated equal amounts on either PIA that we measured to contain 300 μM of phosphate or PIA supplemented with 1 mM K-phosphate buffer, pH 7.0. Cells grown on these plates were analyzed for the presence of appendages by transmission electron microscopy (TEM) and immunobloting. TEM images clearly demonstrated the absence of appendages in strain MDR25 grown on phosphate-enriched PIA (Figure 2F) and an abundance of appendages in the same strain grown on PIA only (Figure 2G). We also noted the differences in colony phenotype when smooth surface colonies appeared on high Pi media (Figure 2Fa), and rough surface colonies appeared on low Pi media (Figure 2Ga). Immunoblot analysis (Figure 2H and 2I) demonstrated the absence of proteins recognized by anti-PA5369 antibody on phosphate rich media versus their abundance on phosphate poor media in both sheared cell surface fractions (Figure 2H) and cell pellets (Figure 2I). Finally, we performed immuno-gold electron microscopy of strain MDR25 to confirm the presence of PstS protein on the appendages. Whole cells of strain MDR25 were directly harvested from PIA plates and incubated with anti-PA5369 antibody followed by incubation with gold-labeled goat anti-rabbit antibody. Figure 2J demonstrates gold spots localized on the cell surface structures in strain MDR25. Gold spot localization was not observed in negative controls performed in the absence of primary anti-PA5369 antibodies (data not shown). We noted the fragility of these appendages on EM as detached and fragmented appendages were observed (see Figure S3).
In order to determine the contributory role of the PA5369-like protein on the ability of MDR P. aeruginosa to adhere to and disrupt barrier function of cultured intestinal epithelial cells, we examined the effect of anti-PA5369 antibodies on the adhesiveness and barrier disrupting capability. In order to avoid the non-specific interference of whole antibodies, we purified Fab fragments of anti-PA5369 antibodies to use in these experiments. Using stain MDR25, we performed adhesion assays to Caco-2 monolayers and determined the transepithelial resistance (TER) of Caco-2 cells, a measure of barrier function, in the presence or absence of Fab fragments of anti-PA5369 antibodies. Both the adhesiveness of strain MDR25 to Caco-2 monolayers (Figure 3A) and the ability of strain MDR25 to disrupt the TER of Caco-2 monolayers (Figure 3B) were significantly attenuated when pre-incubated with the Fab fragments of anti-PA5369 antibodies.
In order to determine if PstS could influence the ability of non-multi-drug resistant strains to alter intestinal barrier function, we performed complementary experiments using P. aeruginosa MPAO1 and its derivative mutant ΔPA5369 [13]. The mutant ΔPA5369 was complemented with the pa5369 gene on a multi-copy plasmid pUCP24 (Δ5369/pa5369). First strains were verified for the presence of surface-associated PstS by ELISA using specific anti-PA5369 antibodies (Figure 3C). In order to determine if PstS contributed to the adhesiveness of MPAO1, we apically inoculated Caco-2 monolayers with P. aeruginosa strains and assessed the degree of adhesiveness after one hour of co-incubation. We observed the adhesiveness of MPAO1 to cultured intestinal epithelial cells to be as low as 1% of the initial inoculum; an effect that was further decreased with the mutant strain Δ5369 (Figure 3D). Strain Δ5369/pa5369 demonstrated increased adhesiveness to Caco-2 cells compared to both the wild type and 5369 mutant (Figure 3D). Reiterative experiments were then performed to assess the ability of the strains to alter epithelial barrier function, as measured by TER of Caco-2 cells. We have previously reported strain MPAO1 to display low virulence against Caco-2 monolayers (∼5% decrease in TER at 3 hours) compared to clinical strain MDR25 (∼70% decrease in TER at 3 hours) (see Table 1). However at later time points (7 hours) strain MPAO1 decreased resistance of Caco-2 monolayers by 60%–70%. Therefore, we measured TER after 7 hours of co-incubation of Caco-2 cells using MPAO1 and its derivatives and found that strain Δ5369 was significantly attenuated in its ability to decrease the TER of Caco-2 monolayers (Figure 3E). Complementation of Δ5369 with pa5369 gene restored its effect to decrease the TER of Caco-2 cells similar to the wild type PAO1 (Figure 3E).
We incidentally noticed the spontaneous appearance of smooth colonies among rough-edged colonies of MDR25 when grown on PIA where the phosphate level (Pi) was determined to be ∼300 μM (Figure 4A, black arrows). When smooth colonies were isolated and re-plated on PIA, the rough-edged colonies re-appeared interspersed among smooth colonies (Figure 4B, shown by white arrow) suggesting the possibility of colony phase variation. We also noted that rough (MDR25R) and smooth (MDR25S) colonies were distinct in their production of biofilm, whereby MDR25S produced significantly greater amounts of biofilm compared to MDR25R (Figure 4C). Growth curves for MDR25S and MDR25R were similar in liquid Pseudomonas broth (see Figure S4). We next examined smooth and rough colonies for their PstS content on surface sheared fractions by ELISA using anti-PA5369 antibodies and determined that smooth edge colonies produced significantly less PstS compared to rough colonies (Figure 4D). Finally, we determined if MDR25R and MDR25S differentially induced mortality in mice using an established model of lethal gut-derived sepsis [14,15]. This model involves creating a surgical stress with a 30% hepatectomy and simultaneous intestinal exposure to P. aeruginosa via direct injection into the cecum [15]. This model is of particular clinical relevance as it is well established that surgical hepatectomy results in severe hypophosphatemia [16]. Rough and smooth colonies were suspended in 10% glycerol at OD 0.25 (600 nm) and injected into the cecum at the time of hepatectomy. Mice were followed for 48 hours for mortality. Results demonstrated that mice injected with the smooth edged, PstS poor strain MDR25S displayed 10% mortality at 48 hours whereas mice injected with the rough edged PstS rich strain MDR25R, displayed 60% mortality at 48 hours (Figure 4E). Data were analyzed using Kaplan-Meier survival curves with SPSS software, n = 10/group, p = 0.021.
We determined if phosphate supplementation in mice subjected to a 30% hepatectomy, could prevent lethality due to MDR25R or MDR1 by performing reiterative experiments in which mice were fed varying concentrations of phosphate ([Pi]). Group 1 (n = 8) were fed water only, Group 2 (n = 8) were fed 0.2x PBS (phosphate buffered saline, pH 7.4, [Pi] = 2 mM) as their drinking water, and Group 3 (n = 5) were fed 1x PBS ([Pi] =10 mM) as their drinking water for 36 hours before surgical hepatectomy and injection of bacteria. In addition, prior to injection of MDR25R into the cecum, bacteria were suspended in either water containing 10% glycerol (group 1) or 0.2x PBS (group 2) or 1x PBS (group 3). Results shown in Figure 5A demonstrate that MDR25R caused 100% mortality within 48 hours when mice drank water only (Group1) whereas mice drinking a water solution containing 2 mM phosphate (Group 2) had significantly decreased mortality (50%) while mice drinking a water solution containing 10 mM phosphate (Group 3) had no mortality (100% survival). Data were analyzed using Kaplan-Meier surviving curves in SPSS software, p = 0.004. Similar results were found in reiterative experiments with the clinical isolate MDR1 (Figure 5B) (n = 10, p = 0.001).
The outer surface expression of PstS PA5369 observed in the current study is at variance with its previously reported characterization as a periplasmic protein. In order to clarify this we hypothesized that knockout of adjacent low phosphate responsive elements might impair the surface expression of PstS. PstS PA5369 is clustered to the phosphate ABC transporter locus (Figure 6A). Based on the KEGG SSDB (Kyoto Encyclopedia of Genes and Genomes Sequence Similarity DataBase http://www.genome.jp/kegg/ssdb/) search, PstS PA5369 can be considered as a paralogous protein of PstS PA0688 in P. aeruginosa PAO1 that is clustered to the alternative type II secretion locus (Figure 6B). According to recent data [17], PA0688 is characterized as alkaline phosphatase that is secreted by the alternative type II secretion system Hxc induced under low phosphate conditions. We therefore hypothesized that hxc might also play a role in the outer surface expression of PstS PA5369 and performed immunoblot analysis of sheared appendages in strains P. aeruginosa PAO1 and its derivative mutant ΔHxcR (PA0686). As shown in Figure 6D, the ΔHxcR mutant was attenuated in the production of the outer surface but not intracellular PA5369 suggesting at least partial involvement of hxc system to present PstS on outer surface appendages. Moreover, complementation of the mutant with hxcR restored its ability to express outer surface PstS (Figure 6D), confirming involvement of the Hxc system. N-terminal sequence of proteins from sheared fractions of the MDR isolates 1 and 13 by Blast Search correspond to PA14–55410 which is orthologous to PA0688 and clusters to the hxc system (Figure 6C). We amplified and sequenced the corresponding gene in strain MDR1, and found that the protein encoded by this gene had 90% identity to PA14 55410 from P. aeruginosa PA14, 45.3% identity to PA0688 from P. aeruginosa PAO1, and 64.3% identity to the human plasma phosphate-binding protein HPBP which is classified as a DING protein [18–20]. We therefore named the protein from MDR1 as DING and its respective gene dinG (GenBank Accession number EF616488). We next performed experiments to determine the expression level of pstS and related genes from the phosphate ABC transporter system, pstS pa5369, phoB and phoU, as well as pstS and related genes from hxc system, pa0688, dinG, and hxcX, a gene of hxc operon (Figure 6A–6C). We examined four strains: MPAO1, MDR1, MDR25R, and MDR25S. Strains were grown overnight on PIA and PIA complemented with 10 mM K-Ph buffer, pH 7.0, and RNA was directly isolated from cell colonies. Results are presented in Figure 6E. While the expression of the housekeeping enzyme citrate synthase demonstrated no significant change in response to phosphate limitation, the expression of all genes tested was increased in response to low phosphate media with each expressing a distinct pattern. While a similar increase in phoU expression was observed between all strains, only a modest increase in phoB expression was detected in MDR25R compared to other strains. Similarly only a modest increase in pstS pa5369 expression was observed in the MDR isolates compared to MPAO1. The most intriguing finding however was that, although pstS pa5369 expression was similar between the rough (high outer surface PstS) and smooth (low outer surface PstS) colony variants, a dramatic difference in expression was observed in the hxc operon. In fact, hxcX expression was ten times higher in MDR25R compared to MDR25S. The low response of hxcX to phosphate limitation was also observed in MPAO1 strain compared to MDR25R and MDR1. Although the expression of pa0688, the orthologous of DING protein in MPAO1, was 20-fold higher at low phosphate concentrations, this effect was small compared to that observed for MDR strains where a 150-fold increase was observed with MDR25S, a 1,400-fold in MDR25R, and 5,000-fold increase in expression in MDR1.
Numerous reports have documented that the rise in MDR nosocomial pathogens continues to threaten hospitalized patients despite the implementation of various countermeasures including isolation techniques and antibiotic de-escalation measures [21]. While the mere culture of a MDR resistant pathogen such as P. aeruginosa is perceived to be a real and present danger to patients primarily because it cannot be readily eliminated by antibiotics, the evidence that antibiotic resistance itself confers a more virulent phenotype is highly variable. Our previous work on screening consecutively isolated MDR strains of P. aeruginosa from critically ill hospitalized patients demonstrated that strains express extremely polar virulence phenotypes against the intestinal epithelium from those that are essentially inert, to those that are highly motile, adhesive, and destructive [2]. In fact among the consecutively collected strains in this series, only a minority of strains displayed a virulent phenotype against the intestinal epithelium (∼15%). A better understanding of the virulence determinants of MDR P. aeruginosa and their mechanism of action against the intestinal epithelium is important given the high prevalence of colonization of this organism in the intestinal tract of critically ill and immuno-compromised patients [22–24].
Human critical illness represents a unique ecological niche for P. aeruginosa because of its prolonged exposure to antibiotics and physiologic disturbances that have no historical precedent in terms of host survival. Extensive life sustaining measures employed during the care of the critically ill such as the use of gastric acid suppression therapy, vasoactive agents that result in profound luminal hypoxia, continued use of opioids that impair the ability of the intestinal tract to excrete non-commensal pathogens, and the delivery of highly processed artificial nutrition, all favor the exposure of pathogens like P. aeruginosa to a composite of environmental cues that can directly activate its virulence circuitry [25–29]. In this regard a major environmental cue within the intestinal tract that could shift the virulence of P. aeruginosa to that of a more virulent phenotype against the epithelium may be low extracellular phosphate which is often present during severe critical illness [16,30–32]. Hypophosphatemia is reported to be present following a variety of physiologic stress states such as myocardial infarction [33], ischemia-reperfusion injury [33], major liver resection [16], use of insulin to control hyperglycemia [34,35], use of intravenous nutrition [36,37], and during sepsis [38]. In such circumstances, phosphate depletion appears to be severe and an independent predictor of mortality due to infection and sepsis [38]. While the mechanisms for this observation are unknown, it is possible that colonizing strains of P. aeruginosa present in the intestinal lumen of critically ill patients become activated to express a more virulent phenotype against the intestinal epithelium in response to low phosphate concentrations resulting from surgical injury and catabolic stress.
In the present study, we determined that MDR P. aeruginosa clinical strains displaying a high degree of virulence against cultured intestinal epithelial cells express an extraordinary amount of surface-associated PstS proteins orthologous to PA5369 from P. aeruginosa PAO1 and PA14 55410 protein from P. aeruginosa PA14. The observation that PstS on appendages contributes to intestinal epithelial adherence coupled with its known role as a phosphate binding protein, raises the possibility that the PstS present on appendages might facilitate the ability of MDR P. aeruginosa to acquire phosphate from intracellular stores within the host. This latter effect appeared to be dependent on the presence and expression of the alternative type II secretion system Hxc, which itself is activated in the presence of low phosphate [17]. This finding establishes a link between the phosphate binding ABC transporter and the Hxc system in P. aeruginosa. Simultaneous expression of both systems is necessary for the formation of outer surface PstS-rich appendages as neither ΔPstS nor ΔHxcR mutants produce them in P. aeruginosa MPAO1. Both the PstS and the Hxc systems are highly inducible in MDR clinical isolates that express a particularly adhesive and barrier disrupting phenotype against intestinal epithelial cells. In this regard, certain MDR P. aeruginosa strains may have adapted unique genetic changes in response to unusually harsh selective pressures that typify critically ill humans including multiple antibiotic use, severe hypoxia, and the ability to sustain life with prolonged intravenous nutrition. Among such changes may be the ability to acquire phosphate and other nutrients from within host cells given that physiologic stress and tissue injury are know to shift phosphate into the intracellular compartment resulting in severe hypophosphatemia [39]. Under such circumstances outer surface expression of PstS on appendages may confer an evolutionary advantage to P. aeruginosa by expressing phosphate acquiring structures capable of scavenging intracellular phosphate at arm's length from the host immune system.
Data from the present study are not the first to demonstrate that PstS is secreted by bacteria during nutrient manipulation of the media. For example, it has been recently reported that Streptomyces lividans secretes PstS into liquid cultures containing high concentrations (>3 %) of certain sugars, such as fructose, galactose, and mannose [40]. Another example in which PstS has been shown to be secreted is PA0688 protein in P. aeruginosa PAO1. Inquiry of the PA0688 protein into the “Clusters of Orthologous Groups (COG) Program” predicted it to be PstS. However PA0688 clusters to the hxc loci and has been shown to be secreted by the Hxc system under phosphate depleted conditions and functions as alkaline phosphatase [17]. In the current study, the proteins identified in MDR P. aeruginosa orthologous to PA0688 were expressed several hundred fold higher than in MPAO1. That sequence analysis revealed this protein in MDR1 to belong to DING proteins is intriguing, given that the origin and function of DING proteins have remained a focus of speculation. DING proteins are characterized by their N-terminal DINGGGATL-sequence and are highly conserved in both animal and plants, although they are more variable as microbial proteins [19,41–43]. There are some functional similarities between DING proteins from pro- and eukaryotes including structural homology with phosphate-binding proteins [19,43]. It has been hypothesized that pathogenic or symbiotic bacteria might acquire the DING gene via horizontal gene transfer from eukaryotes in order to sense and respond to host signals or to modify intercellular signaling pathways in host cells [42]. Others have suggested that DING proteins do not exist in eukaryotes at all, and that their detection in human tissues has been a result of microbial contamination or infection [41]. Based on the codon usage analysis of DNA, it has been assumed that DING sequences found in eukaryotes are of Pseudomonas origin [41]. Further work is in progress to characterize the role of the DING-protein related appendages found in this series of MDR clinical isolates, the results of which may add to our understanding of the impact of bacterial DING proteins on the modulation of signal transduction in animals.
The gene pstS pa5369 is part of the pst operon encoding a specific phosphate transport system that is activated under low phosphate conditions. The high affinity phosphate transport system pst belongs to the Pho regulon that is controlled by the two-component regulatory system PhoB/PhoR, which responds to local phosphate concentration. PhoB/PhoR is highly conserved and widely present in Gram negative and Gram positive microorganisms. In addition, PhoB/PhoR controls the expression of multiple genes [44,45] many of which are involved in phosphate uptake and metabolism and various other metabolic pathways such as the de novo biosynthesis of NAD [46], the initiation of chromosome replication [47], the acid shock response [48], the RpoS-mediated stress response [49], and AMP hydrolysis [44,50]. The phosphate regulon might be also involved in the activation of quorum sensing in P. aeruginosa as evidenced by the recent observation that the transcriptional activation of rhlR and production of PQS and pyocyanin develop during phosphate limitation [12]. Furthermore, a link between the expression of the ABC phosphate transporter and penicillin resistance in Streptococcus pneumoniae has been reported thereby proposing a novel role for PstS [51]. These investigators reported that the pstS gene product was overproduced in resistant isolates, the inactivation of which resulted in penicillin sensitivity [51]. Further evidence linking PstS to antibiotic resistance has been demonstrated in fluoroquinolone resistant Mycobacterium smegmatis [52] where amplification of the phosphate specific transporter suggested that the efflux mediated fluoroquinolone resistance might be an intrinsic function of the Pst system [52–54]. Thus is it plausible that the development of multi-drug resistance in P. aeruginosa clinical isolates MDR1, MDR13, and MDR25 might be related to the overproduction of PstS proteins.
Mouse lethality experiments from the present study strongly suggest a significant role for PstS in the virulence of MDR25R P. aeruginosa virulence in vivo. The importance of PstS in in vivo virulence has been previously addressed in various models including a mouse infection model using Mycobacterium tuberculosis and pstS1 and pstS2 knockout strains [55], a fish infection model using Edwardsiella tarda, a facultative aerobic enterobacterium that causes hemorrhagic septicemia in fish and gastrointestinal infections in humans [56], and a chicken infection model using Escherichia coli O78, an organism associated with extraintestinal infections and septicemia in poultry, livestock, and humans [57].
In summary, we have identified PstS-rich appendage-like structures on the outer surfaces of selected strains of MDR P. aeruginosa that confer a highly adhesive and virulent phenotype against cultured intestinal epithelial cells. Further characterization of these appendages and better understanding of their molecular regulation are needed to fully define their role in the virulence of multi-drug resistant P. aeruginosa. The observation that critical virulence factors such as PstS in P. aeruginosa are highly responsive to environmental phosphate, in conjunction with the observation that intestinal phosphate repletion completely prevented mortality in surgically injured mice exposed to MDR strains, underscores the importance of recognizing intestinal phosphate depletion following catabolic stress and a possible strategy of intestinal phosphate loading as a countermeasure against colonizing strains of P. aeruginosa that are resistant to all conventional antibiotics.
The consecutively collected clinical strains of MDR P. aeruginosa used in the present study (Table 1) have been characterized and described previously [2]. P. aeruginosa strains MPAO1, MPAO1 mutant ΔPA5369 (PA5369:: ISphoA/hah, ID 29772, and MPAO1 mutant ΔPA0686 (PA0686:: ISphoA/hah, ID 957) were obtained from the P. aeruginosa mutant library [13]. The mutant ΔPA5369 was complemented with pa5369 gene to create strain ΔPA5369/ pa5369, and the mutant ΔPA0686 was complemented with DNA comprising pa0686 plus pa0687 genes to create the strain ΔPA0686/ pa0686-pa0687. The MDR clinical isolates were routinely subcultured from frozen stocks on Pseudomonas isolation agar (PIA) containing Gm, 50 μg/ml. Note that strain 25, herein referred to as MDR25, and all strains of genotype 20 (see Table 1) did not grow on rich media (LB- liquid or agarized) or TSB (liquid or agarized), and did not grow in the specially designed phosphate limited liquid media described by Hancock [58]. We observed strain MDR25 growth was best supported in Pseudomonas broth that we determined to contain 2 mM Pi, and PIA determined to contain ∼0.3 mM Pi.
Human intestinal epithelial cells Caco-2bbe were grown to confluence in 0.3 cm2 transwells (Costar), and their barrier function was assessed by measuring the transepithelial electrical resistance (TER) to a fixed current across cells as previously described [2]. All experiments were performed in triplicate.
Adhesiveness of P. aeruginosa to Caco-2 bbe cells was determined as previously described [2]. All experiments were performed in triplicate.
Biofilm formation was assayed as described with modifications [59]. Briefly, P. aeruginosa MDR clinical strains were grown overnight in 2 ml of PB, Gm 50 μg/ml in 15 ml culture tubes at 37°C, 200 rpm (C24 Incubator Shaker, New Brunswick Scientific, Edison, NJ). The wells were then rinsed thoroughly with water and the attached material was stained with 3 ml of 0.1% crystal violet, washed with water, and solubilized in 3 ml of ethanol. Solubilized fractions were collected and absorbance measured at 550 nm. All experiments were performed in triplicate.
P. aeruginosa strains were grown on PIA plate, than bacteria were harvested, suspended in PBS, and surface-associated structures were sheared by vigorous vortexing for 2 min. After centrifuging at 5,000g, for 5 min, proteins in the supernatant were separated using 10% Tris-glycine SDS-polyacrylamide gel and detected by Coomassie brilliant blue staining. For amino-terminal peptide sequence analysis, the proteins were electroblotted onto polyvinylidene fluoride (PVDF) membranes, and sequenced by Edman degradation chemistry using an Applied Biosystems Procise 492 HT Protein Sequencer (Applied Biosystems, Foster City, CA) at the Mayo Proteomics Research Center (Mayo Clinic College of Medicine, Rochester, MN).
Polyclonal rabbit antiserum against 192–212 peptide KEEALCKGDFRPNVNEQPGS of PA5369 (anti-PA5369) was produced in rabbits (SynPep Corporation, Dublin, CA). Anti- PA5369 antibodies were affinity purified by AminoLink Plus Immobilization Kit (Pierce) using 192–212 peptide to create an affinity column. For immunoblot analysis, proteins were electrotransferred from SDS-polyacrylamide gels to PVDF membrane (Immobilon-P, Millipore) and primed with affinity pure anti-PA5369 antibodies at 1:1,000 dilution. Affinity pure F(ab)2 fragments of anti-rabbit IgG conjugated with horseradish peroxidase (Jackson Immunological Res Lab) at 1:5,000 dilution was used as secondary antibody. Detection was performed using SuperSignal West Dura Extended Duration Substrate (Pierce).
P. aeruginosa strains were grown on PIA plates for 2 days, bacteria were harvested, suspended in PBS containing protease inhibitor cocktail (Roche) to create a cell density of 5.0 (OD 600 nm) in a total volume of 500 μl, and centrifuged at 6,000 rpm, 5 min. The pellet was resuspended in 500 μl PBS containing protease inhibitor cocktail, and vigorously vortexed for 2 min. Cell surface associated proteins were separated by centrifuging for 5 min at 5,000 g. After centrifugation, the supernatants were diluted (1:5) with carbonate-bicarbonate buffer (Sigma), and 200 μl/well was used for coating Maxisorp Loose Immuno-modules (Nunc). Plates were incubated at 4°C, overnight, washed with PBS, and unbound sites were blocked with 3% bovine serum albumin in PBS for 30 min at room temperature. Rabbit polyclonal affinity purified anti-5369 antibody (1:1,000) followed by HRP-labeled affinity purified F(ab)2 fragments of anti-Rb IgG (Jackson Immunological Research Laboratories) (1:5,000), and o-phenylaminediamine (Sigma) were used to detect PA5369-like protein at 450 nm optical density.
Polyclonal affinity purified rabbit anti-PA5369 antibody against 192–212 peptide of PA5369 were used to isolate Fab fragment by ImmunoPure Fab Preparation Kit (Pierce) accordingly to manufacturer protocol.
Transmission electron microscopic analysis was performed as previously described [60] with minor modifications. Briefly, bacteria were grown for 48 hours on PIA media with/without Gm, 20 μg/ml. A drop of water was deposited on the edge of colony, and a Formvar-coated copper grid was immediately floated on the drop for 30 s, then rinsed with TE buffer (10 mM Tris-HCl, 1 mM EDTA, pH 8.0) and stained with a 1% aqueous solution of uranyl acetate. Samples were examined under 300 KV with a FEI Tecnai F30 electron microscope.
For immuno-gold labeling, 200 mesh formvar-coated nickel grids were rinsed with TE buffer, rehydrated with PBS for 30 min and blocked with 1% BSA for 30 min followed by transferring to anti-PA5369 antibodies diluted as 1:100 in 1% BSA. Incubation was allowed at a humidified chamber for 3.5 hours, at room temperature, followed by extensive washing with PBS, blocking with 0.1% BSA for 25 min, and incubating in the humidified chamber for 1 hour with goat anti-rabbit IgG conjugated with 10 nm gold particles (TED PELLA) at 1:10 dilution in 0.1% BSA. Grids were washed with PBS, fixed with 1% glutaraldehyde in PBS for 10 min, washed with water, and stained briefly with uranyl acetate and lead citrate. Air dried grids were examined under 300KV with FEI Tecnai F30.
The pa5369 gene was amplified using PAO1 DNA and primers forward 5369F EcoRI 5' CCGGAATTCGATGAAACTCAAGCGTTTG 3'and reverse 5369R XbaI 5' GCTCTAGACAAGTCACTGGATTACAG 3' and cloned in E.coli-P. aeruginosa shuttle vector pUCP24 [61] using EcoRI and XbaI restriction sites to create pUCP24/5369 where pa5369 is expressed from Plac promoter. The plasmid pUCP24/5369 was electroporated in ΔPA5369 to create strain Δ5369/5369.
The DNA containing pa0686 and pa0687 was amplified using PAO1 DNA and primers forward PA0686-744301F-EcoRI 5' CCGGAATTCGCGCGGTACCGTTGG 3' and PA0687-746961R-XbaI 5' GCTCTAGACGGACTACTGGACCAGTTG 3'and cloned in E.coli-P. aeruginosa shuttle vector pUCP24 [61] using EcoRI and XbaI restriction sites to create pUCP24/0686–0687 under regulation of Plac promoter. The plasmid pUCP24/0686–0687 was electroporated in the MPAO1 mutant strain ΔHxcR (ΔPA0686) to create strain ΔHxcR/0686–0687.
The forward primer 5369F EcoRI 5' CCGGAATTCGATGAAACTCAAGCGTTTG 3'and reverse primer 5369R XbaI 5' GCTCTAGACAAGTCACTGGATTACAG 3' were designed based on the sequence of P. aeruginosa PAO1 genome and used to amplify genes using genome DNA isolated from strains MDR 1, MDR 13, and MDR 25. The gene analogous to pa14–55410 was amplified using genome DNA of strain MDR1, and primers 55410F EcoRI 5'CCGGAATTCGATGTACAAGCGCTCTCTGAT 3' and 55410R XbaI 5' GCTCTAGACAAG TTAGAGCGGACGGCCGAT 3' designed based on the sequence of P. aeruginosa PA14 genome. Amplified DNAs were cloned directly into pCR2.1 (Invitrogen), and the sequence was obtained using standard M13 Forward and M13 Reverse primers on an Applied Biosystems 3730XL genetic analyzer (University of Chicago, Cancer Research Center, DNA Sequencing & Genotyping Facility).
Strains MPAO1, MDR25R, MDR25S, and MDR1 were grown overnight on PIA and PIA supplemented with 10 mM K-Ph buffer, pH 7.0, and collected directly in the RNA protect buffer (Qiagen), and RNA isolation, DNA degradation, and cDNA synthesis were performed as previously described [27]. Real time PCR was performed on the ABI 7900HT System using SYBR Green qPCR SuperMix-UDG (Invitrogen), cDNA, and respective primers: for citrate synthase PA1580 gene gltA, PA1580–434 5' TCTACCACGACTCCCTGGAC 3' and PA1580–590 5' TTTTCCGCGTAGTTCAGGTC 3'; for PstS PA5369 gene pstS, PA5369–148 5' ACTCTGGCCAACCTGATGAC 3' and PA5369–335 5'CCGTACTTCTGCTCGAAAGC 3'; for phosphate uptake regulator PhoU PA5365 gene phoU, PA5365–523 5' CGCGAACTGGTCACCTACAT 3' and PA5365–711 5' CTCGACCTCTTCCTTCATGC 3'; for low phosphate response regulator PhoB PA5360 gene phoB, PA5360–7 5' GGCAAGACAATCCTCATCGT 3' and PA5360–164 5'CAGTCGAGCAGGATCAGGTC 3'; for PstS PA0688 gene pa0688, PA0688–693 5'GGTGAACATCAACAGCAACG 3' and PA0688–872 5'TAACCGACGATGGAGTAGCC 3'; for PstS analogous to PA14–55410) gene dinG, S1-DING-427 5' CTCTGCCGTTCAACAAGTCA 3' and S1-DING-604 5'CGGGTGAACAGTTCGGTAGT 3', for HxcX atypical pseudopilin PA0682 gene hxcX , PA0682–299 5' AAGACGAGCAGGGCAAGTT 3'and PA0682–454 5'GTGCATAGGAGGCGAGTACC 3'. 0.5 μg of RNA after DNAse treatment was converted to cDNA in 20 μl of reaction mixture (High Capacity cDNA Reverse Transcription kit, Applied Biosystems). The cDNA and RNA (-RT control) were diluted either as 1:50 (for cS, pstS 5369, phoU, phoB) or 1:10 (for hxcX) or 1:500 (for pa0688 and dinG), and 5 μl of diluted mixture was used as a template added to 7.5 μl of master mix containing as manufactured (Invitrogen) SYBR green, ROX, and respective primers. The amplification was run in 384 well plates. Expression levels were calculated based on differences in Ct levels.
All primers were confirmed for DNA amplification using genome DNAs from MPAO1 and MDR clinical isolates prior the Real Time experiments.
All experiments were approved by the Animal Care and Use Committee at the University of Chicago (Protocol IACUC 71744). Six-seven week old male C57BL6 mice were ordered from Harlan Spraque Dawley Animal facility and allowed at least four days for housing acclimation prior to experiments. The mouse model of gut-derived sepsis was performed as previously described [15] with following modifications. Mice drank either water (no phosphate supplementation), or 0.2x PBS (2 mM Pi), or 1x PBS (10 mM Pi ) for 36 hours prior to lethality experiments. Animals were anesthetized (ketamine 100 mg/kg, xylazine 10 mg/kg) intraperitoneally and a 30% hepatectomy was performed on the left lobe of the liver, and the bacterial suspension of P. aeruginosa clinical isolates MDR25 or MDR1 were injected directly into the distal ileum and cecum with a fine high gauze needle. The abdomen was closed in two layers with suture, and mice were allowed to drink either water or PBS but were given no food for 48 hours. Animals were followed for mortality and sacrificed when they appeared septic and moribund.
Eight-week-old male C57BL6 mice were given only water for 36 hours prior to hepatectomy. P. aeruginosa MDR25S and MDR25R were grown overnight in Pseudomonas broth (PB) containing Gm, 50 μg/ml. Overnight cultures were diluted as 1:500–1:250 in water containing 10% glycerol, and 20–50 μl were plated on PIA, Gm, 50 μg/ml. After 24–36 hours of growth, cells were collected directly from plates using an Olympus SZX16 stereo microscope to insure proper collection of rough and smooth colonies. The cells were diluted in water containing 10% glycerol to OD600 nm 0.25, and 200 μl of bacterial suspension was injected in the cecum of mice immediately after hepatectomy. The abdomen was closed, and mice were allowed drinking water. Mice were followed for mortality and sacrificed when they appeared septic and moribund. Data were analyzed using Kaplan-Meier surviving curves and SPSS software employing the Long-rank (Mantel-Cox) test for significance.
Statistical analysis of the data was performed with Student t-test using Sigma plot software and Kaplan-Meier survival curves using SPSS software.
PstS PA5369 (Pseudomonas aeruginosa PAO1), NP_254056; PA0688, probable binding protein component of ABC transporter (Pseudomonas aeruginosa PAO1), NP_249379; PA14_55410, Hypothetical, unclassified, unknown (Pseudomonas aeruginosa UCBPP-PA14), complete genome NC_008463; PhoB, two-component response regulator (Pseudomonas aeruginosa PAO1), NP_254047; HxcX, atypical pseudopilin (Pseudomonas aeruginosa PAO1), NP_249373; PA0686, probable type II secretion system protein (Pseudomonas aeruginosa PAO1), complete genome NC_002516; PhoU, phosphate uptake regulatory protein (Pseudomonas aeruginosa PAO1), NP_254052; citrate synthase (Pseudomonas aeruginosa PAO1), NP_250271; Pseudomonas aeruginosa PAO1, complete genome, NC_002516; ExoU (Pseudomonas aeruginosa), AAC16023; human plasma phosphate-binding protein (HPBP), P85173.
In the present study: PstS (Pseudomonas aeruginosa MDR1), EF601157; PstS (Pseudomonas aeruginosa MDR13), EF601158; PstS (Pseudomonas aeruginosa MDR25), EF601159; DING (Pseudomonas aeruginosa MDR1), EF616488. |
10.1371/journal.pgen.1002358 | A Novel Protein LZTFL1 Regulates Ciliary Trafficking of the BBSome and Smoothened | Many signaling proteins including G protein-coupled receptors localize to primary cilia, regulating cellular processes including differentiation, proliferation, organogenesis, and tumorigenesis. Bardet-Biedl Syndrome (BBS) proteins are involved in maintaining ciliary function by mediating protein trafficking to the cilia. However, the mechanisms governing ciliary trafficking by BBS proteins are not well understood. Here, we show that a novel protein, Leucine-zipper transcription factor-like 1 (LZTFL1), interacts with a BBS protein complex known as the BBSome and regulates ciliary trafficking of this complex. We also show that all BBSome subunits and BBS3 (also known as ARL6) are required for BBSome ciliary entry and that reduction of LZTFL1 restores BBSome trafficking to cilia in BBS3 and BBS5 depleted cells. Finally, we found that BBS proteins and LZTFL1 regulate ciliary trafficking of hedgehog signal transducer, Smoothened. Our findings suggest that LZTFL1 is an important regulator of BBSome ciliary trafficking and hedgehog signaling.
| Primary cilia are considered to be a signaling hub coordinating multiple signaling pathways. Impairment of ciliary function results in developmental defects in vertebrates and also underlies many human disorders including obesity, polycystic kidney disease, and retinopathy. BBS is a prototypical human genetic disorder associated with ciliary dysfunction. Among the known BBS proteins, seven form a complex, the BBSome, which was recently defined as a coat complex transporting membrane proteins between plasma and ciliary membranes. However, the molecular mechanisms controlling BBSome trafficking and the cargos transported by the BBSome are not well understood. In this work, we performed tandem affinity purification using transgenic mice expressing one of the BBSome subunits and identified a novel protein, LZTFL1, as a BBSome interacting protein. We determined that LZTLF1 negatively regulates BBSome ciliary trafficking and that reduction of LZTFL1 activity can compensate for loss of certain BBS proteins and restores BBSome ciliary trafficking. Furthermore, we discovered that BBSome and LZTFL1 regulate ciliary trafficking of Smoothened, a 7-transmembrane Hedgehog signal transducer. Our findings identify an important player in cilia biology and provide novel insights into the regulation of Hedgehog signaling, a crucial signaling pathway for organizing the body plan, organogenesis, and tumorigenesis.
| Primary cilia are microtubule-based subcellular organelles projecting from the surface of cells. Studies during the last decade have shown that primary cilia play essential roles in regulating cell cycle, embryonic development, and tissue homeostasis by acting as a cellular antenna transducing extracellular signals into the cells [1], [2]. Loss of cilia or ciliary dysfunction has been linked to a series of related genetic disorders in humans [3], [4]. These disorders, collectively termed ciliopathies, share common features such as cystic kidney disease, retinal degeneration, and polydactyly.
Bardet-Biedl Syndrome (BBS) is one of the human genetic disorders associated with ciliary dysfunction. Patients with BBS display obesity, polydactyly, retinal degeneration, renal abnormalities, diabetes, hypertension, hypogenitalism, and cognitive impairment. To date, as many as 16 genes have been reported to be involved in BBS [5], [6], [7], [8] (and references therein) and molecular functions of BBS proteins have begun to emerge. Among the known BBS proteins, seven proteins (BBS1, BBS2, BBS4, BBS5, BBS7, BBS8, BBS9) and BBIP10 form a stable complex, the BBSome, which mediates protein trafficking to the ciliary membrane [9], [10], [11]. BBS3 is a member of the Ras superfamily of small GTPases and controls BBSome recruitment to the membrane and BBSome ciliary entry [11]. Of the remaining, three BBS proteins (BBS6, BBS10, BBS12) form another complex with the CCT/TRiC family of group II chaperonins and mediate BBSome assembly [8].
Many receptor proteins and signaling molecules localize to cilia, and the BBSome is involved in transporting at least some of these proteins. For example, several G-protein coupled receptors such as MCHR1, SSTR3, and Dopamine receptor 1 (D1) fail to localize to or abnormally accumulate within the neuronal cilia in Bbs2 and Bbs4 null brains [12], [13]. In Chlamydomonas bbs4 mutants, several proteins aberrantly accumulate within flagella [14]. However, most of the BBSome cargos are currently unknown in mammalian cells. Also unknown is how the trafficking activity of the BBSome is regulated.
In an effort to understand how BBSome function is regulated, we initiated studies to identify BBSome interacting proteins in vivo. In this work, we show that LZTFL1 interacts with the BBSome and negatively regulates its trafficking activity to the cilia. We also provide evidence that the BBSome and LZTFL1 are part of the transport mechanism of Sonic Hedgehog (SHH) signal transducer, Smoothened (SMO), that localizes to cilia [15].
To isolate BBSome interacting proteins in vivo, we generated a transgenic mouse line expressing LAP-BBS4, which allows localization studies and tandem affinity purification using GFP and S tags, under the control of cytomegalovirus (CMV) immediate early promoter. We used mouse testis, where BBS genes are the most abundantly expressed. Expression of LAP-BBS4 in the testis is approximately 2–3 fold higher than that of endogenous Bbs4 (Figure S1A). We first tested whether the recombinant LAP-BBS4 protein is functionally equivalent to endogenous Bbs4. We tested this by three criteria: 1) whether LAP-BBS4 physically associates with other BBSome subunits and is incorporated into the BBSome, 2) whether LAP-BBS4 properly localizes and reproduces the endogenous Bbs4 localization pattern, and 3) whether LAP-BBS4 can functionally rescue the BBS phenotype caused by loss of Bbs4. To confirm the incorporation of LAP-BBS4 into the BBSome, extracts from wild-type, Bbs4−/−, and LAP-BBS4 transgenic mouse testes were subjected to co-immunoprecipitation (co-IP) with anti-GFP antibody. As shown in Figure S1B, pull-down of LAP-BBS4 efficiently co-precipitated all BBSome subunits tested (BBS1, BBS2 and BBS7). In spermatozoa, endogenous Bbs4 is found in the middle piece and the principle piece of the flagella (Figure S1C). Within the principle piece, Bbs4 staining is more intense at the proximal end and gradually decreases toward the distal end of the flagellum. LAP-BBS4 detected by GFP antibody shows a similar localization pattern. Finally, introduction of the LAP-BBS4 transgene into Bbs4−/− mice restores sperm flagella, which are lost in Bbs4−/− mice (Figure S1D), and fertility to Bbs4−/− mice; all four transgenic Bbs4−/− males mated with wild-type females produced pups, while Bbs4−/− males without the transgene did not produce any pups. Based on these criteria, we concluded that LAP-BBS4 is functionally equivalent to endogenous Bbs4.
Protein extracts from wild-type and LAP-BBS4 transgenic testes were subjected to tandem affinity purification. In LAP-BBS4 transgenic testis, all BBSome subunits (Bbs1, Bbs2, Bbs5, Bbs7, Bbs8, and Bbs9) were co-purified with LAP-BBS4, while no BBS proteins were purified when wild-type testes were used (Figure 1A and Table S1). In the LAP-BBS4 transgenic sample, one prominent additional protein was co-purified with BBSome subunits. Mass spectrometry analysis revealed that this protein is Leucine zipper transcription factor-like 1 (Lztfl1). To confirm the interaction and to identify other LZTFL1 interacting proteins, we generated a stable cell line expressing LZTFL1 with FLAG and S tags (FS-LZTFL1) and conducted tandem affinity purification. In this experiment, at least three BBSome subunits (BBS2, BBS7, and BBS9) were co-purified with FS-LZTFL1 (Figure 1B and Table S2). In addition, we found endogenous LZTFL1 proteins were co-purified, indicating that LZTFL1 forms homo-oligomers. Homo-oligomerization of LZTFL1 was confirmed by co-IP and in vitro crosslinking experiments (Figure S2D and S2E). We also found two additional, smaller isoforms of LZTFL1. Although it is unclear whether these smaller forms are bona fide LZTFL1 isoforms or cleavage products derived from the over-expressed FS-LZTFL1, EST database searches revealed the presence of smaller isoforms of LZTFL1 (AK093705 and AK303416) with predicted molecular weights approximately the same as the isoforms observed by us.
To determine the LZTFL1-interacting subunit of the BBSome, each individual subunit of the BBSome was co-transfected with LZTFL1 and analyzed for co-IP (Figure 1C). In this experiment, we found that BBS9 is the LZTFL1-interacting subunit of the BBSome. Based on interaction domain mapping studies, the C-terminal half of LZTFL1 (amino acid (aa) 145–299) was found to interact with BBS9 (Figure 1D). Within BBS9, the fragment containing aa 685–765 interacted with LZTFL1 (Figure S2F), which is a part of the α-helix domain at the C-terminus after the α/β platform domain [11]. In size exclusion chromatography, the peak of LZTFL1 was separated from that of the BBSome, suggesting that LZTFL1 is not a constitutive component of the BBSome and only a subset of LZTFL1 is associated with the BBSome (Figure 1E).
LZTFL1 maps to human chromosome 3p21.3, which is often deleted in several types of cancer [16]. Recently, tumor suppressor function of LZTFL1 has been proposed [17]. However, very little is known about the molecular functions of LZTFL1. To gain insight into the structure and functions of LZTFL1, we performed homology searches. Reciprocal BLAST searches yielded LZTFL1 orthologs in all vertebrates and the flagellate Chlamydomonas reinhardtii, but not in plants, amoebae, or fungi (Figure S2). LZTFL1 homologs are also not found in Caenorhabditis elegans, Drosophila melanogaster, and planaria (Schmidtea mediterranea). Sequence and secondary structure analyses indicated that LZTFL1 is mostly alpha-helical and has a coiled-coil domain in its C-terminal half (Figure S2). A leucine-zipper domain is present as part of the coiled-coil domain. InterPro Domain Scan analysis indicates that aa 212–295 of LZTFL1 has sequence homology to the t-SNARE domain (IPR010989). The structure of this domain in rat Syntaxin-1A (Stx1A) was previously determined (Figure S2C) [18]. The t-SNARE domain of Stx1A forms a rod-like α-helix and is involved in hetero-tetramer formation with other SNARE proteins. Purified Stx1A t-SNARE domain also forms a homo-tetramer [19], suggesting that LZTFL1 may form a similar structure.
We examined expression and localization of LZTFL1. Antibodies against LZTFL1 selectively recognized a protein band at ∼36 kDa in SDS-PAGE (Figure 2A and Figure S3B). This protein was diminished in cells transfected with siRNAs against LZTFL1, verifying the specificity of the antibody. Lztfl1 expression was detected in almost every tissue tested except for skeletal muscle and white adipose tissue (Figure S3A). In immunolocalization studies using hTERT-RPE1 cells, LZTFL1 was detected throughout the cytoplasm (Figure 2B). In contrast to BBS proteins, which show ciliary and centriolar satellite localization (Figure 2D and Figure S4), we did not find enrichment of LZTFL1 around the centrosome or within cilia. GFP-tagged LZTFL1 also showed cytoplasmic localization with no enrichment in the cilia or centrosomes (data not shown). Similar results were obtained from IMCD3 cells and HEK293T cells (Figure 2B and Figure S3C). Consistent with this, we found most of the Lztfl1 immuno-reactivity in the inner segment of the photoreceptor cells and posterior side of the cell body of spermatozoa (Figure S3D, S3E). Since the localization pattern of LZTFL1 is significantly different from that of BBS proteins, we examined where the BBSome-LZTFL1 interaction occurs. Using the in situ proximity-mediated ligation assay [20], LZTFL1 bound BBSomes were found scattered throughout the cytoplasm (Figure 2C). These data indicate that LZTFL1 and BBSome interaction occurs within the cytoplasm.
Next, we sought to determine the biological functions of LZTFL1. Since the BBSome is involved in protein trafficking to the ciliary membrane and LZTFL1 interacts with the BBSome, we investigated whether LZTFL1 has any cilia related functions. To this end, we examined whether LZTFL1 is involved in cilia formation, cilia stability, or cilia length. We found no differences in cilia formation after serum withdrawal, cilia stability upon serum treatment, or the length of cilia in LZTFL1 depleted hTERT-RPE1 cells (data not shown). We next examined whether LZTFL1 is involved in BBSome assembly, BBS protein stability, or BBSome trafficking. We found no defects in BBSome assembly or BBS protein stability in LZTFL1 depleted cells (Figure 3B and data not shown). However, we noticed a dramatic alteration in BBSome localization when we ablated LZTFL1. Normally, BBS proteins localize either within cilia or around centrosomes (Figure 2D and Figure S4). In our experimental conditions, some 42% of control siRNA transfected cells show ciliary localization of BBS9 (Figure 2D, 2F). However, when LZTFL1 was depleted by RNA interference (RNAi), we observed a consistent and striking increase of BBS9 within the cilia with a concomitant decrease in the centriolar satellite pool of BBS9. In contrast, over-expression of wild-type LZTFL1 inhibited ciliary localization of BBS9 (Figure 2E, 2F). Interestingly, over-expression of an LZTFL1 deletion mutant, which lacks the N-terminal 70 amino acids, behaved as a dominant negative form and increased ciliary localization of BBS9. Substitution of two highly conserved basic amino acids (KR) within that region (Figure S2A) was sufficient to cause the same increase in BBS9 ciliary localization. Similar results were observed for BBS4 and BBS8 (Figure S4), indicating that LZTFL1 regulates ciliary localization of the entire BBSome rather than merely BBS9. Combined with the LZTFL1 localization and in situ PLA results, our data indicate that LZTFL1 binds to the BBSome in the cytoplasm and inhibits BBSome ciliary entry. Alternatively, LZTFL1 may promote ciliary exit of the BBSome. Our data also suggest that LZTFL1 has bipartite functional domains; the C-terminal half of LZTFL1 is responsible for BBSome binding and the N-terminal half is for regulating BBSome trafficking activity.
To test whether LZTFL1 regulates general intraflagellar transport (IFT), we examined localization of IFT proteins in LZTFL1 over-expressing and depleted cells. In contrast to BBS proteins, IFT57 and IFT88 were found in almost every cilium (Figures S4 and 5). Depletion of LZTFL1 did not further increase the frequency of ciliary localization of IFT proteins or the fluorescence intensity. More importantly, over-expression of wild-type or mutant variants of LZTFL1 had no impact on ciliary localization of IFT proteins. These data suggest that LZTFL1 is a specific regulator of BBSome ciliary trafficking but not general IFT.
Previously, several BBSome subunits have been shown to be essential for BBSome ciliary localization in C. elegans and C. reinhardtii [14], [21]. However, it has not been systematically investigated which BBSome subunits are essential for BBSome ciliary localization and which are dispensable. To address this, we transfected siRNAs against each BBSome subunit and BBS3, which is involved in BBSome ciliary trafficking [11], into hTERT-RPE1 cells and probed localization of BBS8 and BBS9. Quantitative real-time PCR results confirmed efficient knock-down of each BBS gene expression after siRNA transfection (Figure S6). By using immunofluorescence microscopy, we found that all BBSome subunits and BBS3 are required for BBSome ciliary localization and loss of any single BBSome subunit precluded ciliary entry of the BBSome (Figure 3A and Figure S7). Interestingly, however, the localization pattern of BBS9 (and BBS8) was distinct depending on which BBSome subunit was depleted, suggesting differences in their roles within the BBSome. For example, in BBS1 depleted cells, both BBS8 and BBS9 showed concentric enrichment around the centrosomes with a great increase in the staining intensity (Figure 3A and Figure S7A). This suggests that BBSome components may form aggregates near the centrosomes in the absence of BBS1. Alternatively, BBS1 may be required to return the BBSome back to the cytoplasm. Sucrose gradient ultracentrifugation results are more consistent with the second possibility (Figure 3B). In BBS2 depleted cells, overall staining intensity of BBS9 was greatly decreased. When we measured BBS9 protein level in BBS2 depleted cells, the amount of BBS9 was significantly reduced compared to control siRNA or other BBS gene siRNA transfected cells (Figure S6B). In addition, we found BBS2 protein level was also significantly decreased in BBS9 depleted cells, suggesting that BBS2 and BBS9 are dependent on each other for stable expression. Together, these data indicate that only the intact BBSome can enter the cilia.
To gain insight into the molecular basis of this requirement of each BBSome subunit for BBSome ciliary entry, we investigated the status of BBSome assembly after individual BBSome subunit depletion (Figure 3B). To this end, hTERT-PRE1 cells were transfected with siRNAs against control and each BBS gene, and cell lysates were analyzed by 10–40% sucrose gradient ultracentrifugation. Consistent with the previous result and formation of the BBSome [9], BBS4 and BBS9 were found in the same fraction in control siRNA transfected cells. However, in BBS1 depleted cells, BBS4 and BBS9 were found in separate and lower molecular weight fractions, indicating that BBS4 and BBS9 were not associated in the absence of BBS1. Similarly, depletion of BBS2 and BBS9 also caused disintegration of the BBSome with a significant reduction in BBS9 levels. In BBS4 and BBS8 depleted cells, although there was some disassembly of the BBSome, a significant proportion of the BBSome was still found to be intact. In BBS7 depleted cells, BBS4 and BBS9 were found in the same fractions but the peak was significantly shifted from that of control cells, suggesting that at least one additional subunit is missing from the BBSome in the absence of BBS7. Of note are BBS3 and BBS5 depleted cells. In these cells, the vast majority of the BBSome remained intact, suggesting that BBS3 and BBS5 are not required for BBSome assembly. LZTFL1 depletion or over-expression did not cause any change in BBSome assembly, suggesting that LZTFL1 does not function by modulating BBSome assembly. IFT88, a component of the IFT-B subcomplex, was found in separate fractions from the BBSome and none of the BBS or LZTFL1 perturbations caused any change in IFT88 migration. Together, these data suggest that BBS1, BBS2, BBS7, and BBS9, all of which have β-propeller domains, are required for BBSome assembly (e.g. forming a core scaffold and required for recruitment of at least one additional BBSome subunit), while BBS4, BBS5, and BBS8 have relatively minor or no impact on BBSome assembly and are likely to be in the periphery of the BBSome.
Since LZTFL1 is a negative regulator of BBSome ciliary entry, we tested whether LZTFL1 depletion can rescue the BBSome mislocalization phenotype caused by loss of BBSome subunits. We ablated LZTFL1 expression together with each of the BBSome subunits by RNAi and probed localization of BBS8 and BBS9. Indeed, LZTFL1 knock-down significantly increased ciliary localization of BBS8 in most cells, particularly in BBS3 and BBS5 depleted cells (Figure S7). Ciliary localization of BBS9 was also rescued by LZTFL1 knock-down in all cases except for BBS1 and BBS4 depleted cells (Figure 3C and Figure S7). It is unclear whether the more efficient rescue of BBS9 ciliary localization compared to BBS8 is due to the higher sensitivity of anti-BBS9 antibody or whether it represents features of partial BBSome complexes lacking some BBSome subunits (such as BBS8). Whichever is the case, it is clear that LZTFL1 knock-down can restore ciliary localization of the BBSome at least in BBS3 and BBS5 depleted cells. It is noteworthy that BBS3 and BBS5 are the BBS proteins that have minimal impact on BBSome assembly. Therefore, as long as the BBSome forms, reducing LZTFL1 activity can restore BBSome trafficking to cilia.
Since polydactyly is one of the cardinal features of BBS and a hallmark phenotype of Sonic Hedgehog (SHH) signaling defect, we examined roles of BBSome and LZTFL1 in the SHH pathway. We first examined the requirement of LZTFL1 for SMO ciliary localization in hTERT-RPE1 cells, which express SMO endogenously. In control siRNA transfected cells, SMO localizes to the cilia in response to SMO agonist (SAG) but not in SAG-untreated cells (Figure 4). However, in LZTFL1 depleted cells, ciliary localization of SMO was found in a significant number of cells even without SAG treatment and further increased by SAG treatment. For BBS genes, we tested BBS1, BBS3, and BBS5: two genes (BBS3 and BSB5) that LZTFL1 depletion can restore ciliary trafficking of the BBSome, and one gene (BBS1) that cannot be restored with LZTFL1 depletion. In contrast to LZTFL1 depleted cells, ciliary localization of SMO was significantly decreased in BBS depleted, SAG-treated cells. Consistent with the restoration of BBSome trafficking to cilia in LZTFL1 depleted cells, ablation of LZTFL1 expression in BBS3 and BBS5, but not in BBS1, depleted cells restored ciliary localization of SMO. These data indicate that BBSome function facilitates ciliary localization of SMO and that LZTFL1 suppresses SMO localization to cilia.
Although we were not able to detect physical interactions between the endogenous BBSome and SMO in hTERT-RPE1 cells, presumably due to the transient nature of the interaction and the difficulty of extracting membrane proteins without disrupting protein-protein interactions, we found that several BBS proteins can associate with the C-terminal cytoplasmic tail domain of SMO (aa 542–793) in transiently transfected cells (Figure 4C). Previously, two amino acids (WR; aa 549–550) immediately downstream of the 7th transmembrane domain of SMO were shown to be essential for SMO ciliary localization [15]. Deletion of this WR motif from the cytoplasmic tail of SMO abolished interaction between BBS proteins and SMO (Figure 4D). These data suggest that the BBSome may directly interact with SMO and mediates SMO ciliary localization.
Finally, we investigated whether downstream HH target gene expression is affected by loss of BBSome and LZTFL1 function. In hTERT-RPE1 cells, HH target gene GLI1 expression was relatively mildly induced by SAG treatment (Figure S8A). Depletion of BBS1, BBS3, and BBS5 modestly but consistently reduced GLI1 expression. We also used mouse embryonic fibroblast (MEF) cells, which show robust responsiveness (Figure S8B). Consistent with the results from hTERT-PRE1 cells, knock-down of BBS gene expression significantly reduced Gli1 expression in MEF cells. Although statistically not significant compared with BBS3 and BBS5 single knockdowns, reduction of Lztfl1 activity tends to restore Gli1 expression in Bbs3 and Bbs5 depleted cells. Interestingly, while LZTFL1 depletion resulted in ciliary translocation of SMO even in SAG-untreated cells, GLI1 expression did not increase in LZTFL1 depleted cells, suggesting that SMO accumulated within the cilia in LZTFL1 depleted cells is not activated. This is consistent with the idea that Smo is constantly transported in and out of the cilia even in the inactive state [22], [23] and the recent finding of the 2-step model of SMO activation [24]. Finally, we used MEF cells derived from our Bbs2 and Bbs4 null embryos [25], [26] and found similar reductions in Gli1 expression upon SAG treatment (Figure S8C). Together, our data suggest that the BBSome and LZTFL1 are a part of the Smo ciliary trafficking mechanisms and contribute to the cellular responsiveness to the SHH signaling agonist.
BBSome functions as a coat complex to transport membrane proteins between plasma and ciliary membranes [11]. In this work, we identify LZTFL1 as a negative regulator of BBSome trafficking to the ciliary membrane. Our data indicate that LZTFL1 associates with the BBSome within the cytoplasm and inhibits ciliary entry of the BBSome. Alternatively, LZTFL1 may promote the exit of BBSomes from the cilia. Although we cannot completely rule out the second possibility, currently available data favor the entry inhibition model over the exit promotion model. First, LZTFL1 does not show any enrichment around the basal body or within the cilia either by immunological methods or by GFP-fused recombinant protein. Many proteins involved in vesicle trafficking commonly show some degree of enrichment around the donor compartment. BBSomes also show enrichment around the basal body and within the cilia with a gradual increase toward the ciliary base, which is consistent with the model that the BBSome is recruited to the plasma membrane near the basal body for ciliary entry and to the ciliary membrane at the ciliary base for exit. However, LZTFL1 does not show this localization pattern. Second, although BBSomes show enrichment around the basal body and within the cilia, LZTFL1-associated BBSomes are found scattered throughout the cytoplasm, which is consistent with the idea that LZTFL1 associates with the BBSome within the cytoplasm and limits ciliary access of the BBSome. These observations strongly favor the BBSome ciliary entry inhibition model. Localization of LZTFL1 to the cytoplasm with cilia-related function is similar to the recently characterized seahorse/Lrrc6, which also localizes to the cytoplasm and regulates cilia-mediated processes [27].
Our findings indicate that every BBS gene tested so far is required for ciliary entry of the BBSome and loss of any single BBS protein commonly results in a failure of BBSome ciliary trafficking. This is not limited to BBSome subunits but also found with depletion of other BBS proteins that are not part of the BBSome including BBS3, BBS6, BBS10, and BBS12 (this study and S.S. and V.C.S. unpublished results). These findings suggest that a failure of BBSome ciliary trafficking is a common cellular feature of BBS and support the idea that BBS results from a trafficking defect to the cilia membrane. However, the precise mechanism leading to BBSome mis-localization is different depending on the missing BBS proteins. For example, BBS1, BBS2, BBS7, and BBS9, which commonly contain β-propeller domains, are likely to form the scaffold/core of the BBSome and required to recruit other BBSome subunits. BBS6, BBS10, and BBS12 were previously shown to be required for BBSome assembly by interacting with these β-propeller domain containing BBS proteins [8]. BBS4, BBS5, and BBS8 are likely to be at the periphery of the BBSome and have limited impact on BBSome assembly. However, BBS4 interacts with p150glued subunit of the cytoplasmic dynein machinery and may link the BBSome to the cytoplasmic dynein motor protein [28]. BBS5 was shown to interact with PIPs and may be involved in the association of the BBSome to the membrane [9]. The function of BBS8 is currently unknown and requires further characterization. Despite these functional differences, all BBS proteins are essential for BBSome assembly or ciliary trafficking and only the holo-BBSome enters the cilia. BBSome mis-localization may be used as a cell-based assay to evaluate BBS candidate genes. Remarkably, reducing LZTFL1 activity restores BBSome trafficking to the cilia at least in BBS3 and BBS5 depleted cells. It appears that decreased LZTFL1 activity can compensate for the loss of certain BBS proteins, which are required for BBSome ciliary entry, as long as the BBSome is formed. This implies that certain BBS subtypes may be treated by modulating LZTFL1 activities.
Polydactyly is one of the cardinal features of BBS found in the vast majority of human BBS patients [3] and also a hallmark phenotype of disrupted SHH signaling or IFT function [2], [29], [30]. Several additional features including mid-facial defects and neural crest cell migration defects found in BBS mutant animals are also linked to SHH signaling defects [31]. Therefore, it is speculated that BBS proteins may be involved in SHH signaling. In this work, we show that BBS proteins are involved in ciliary trafficking of SMO. We found that the SMO cytoplasmic tail domain physical interacts with the BBSome at least in overexpressed conditions, while ciliary localization defective mutant SMO does not. Furthermore, ablation of LZTFL1 increases SMO ciliary localization. These data indicate that BBS proteins and LZTFL1 are at least part of the SMO ciliary trafficking mechanism. It remains to be shown that ciliary transport of Smo occurs together with the BBSome (e.g. by live cell imaging) and in a directly BBSome-dependent manner.
Our data also suggest the presence of alternative ciliary trafficking mechanisms for SMO. For example, although ciliary localization of SMO is significantly reduced by BBSome depletion, it is not completely abolished, while ciliary localization of BBS8 and BBS9 is severely reduced. HH target gene GLI1 expression is also relatively mildly affected by BBS protein depletion. This suggests that the BBSome is only part of the transport mechanism and that there is likely to be another mechanism by which SMO is transported to the cilia. Currently, it is unknown whether IFT proteins can directly mediate SMO ciliary trafficking or are involved indirectly. These findings suggest that although BBS proteins are involved in Smo ciliary trafficking, the presence of alternative mechanisms is sufficient to support normal development of neural tubes and survival in BBS mutants. In addition, while polydactyly is very common in human BBS patients, none of the BBS mouse models generated so far display polydactyly as is seen in Smo or Shh mutants [2], implying a potential species-specific requirement of BBS proteins or differential threshold in SHH signaling in the limb bud.
While BBS and IFT proteins are part of the transport machinery (like a train) to deliver cargos to and from the cilia, NPHP, MKS, and Tectonic proteins localize to the transition zone and function like a station (or immigration official) to control the entry and exit of the ciliary proteins. For example, Garcia-Gonzalo, et al. recently showed that the Tectonic complex, which consists of Tctn1, Tctn2, Tctn3, Mks1, Mks2, Mks3, Cc2d2a, B9d1, and Cep290, localizes to the transition zone of cilia and controls ciliary membrane composition [32]. NPHP1, NPHP4, and NPHP8 form another complex at the transition zone and functions as a ‘gate keeper’ together with other NPHP and MKS proteins [33], [34], [35]. These NPHP, MKS, and Tctn protein functions also appear to be partly redundant. In this model, one can envisage certain ciliary proteins transported by the BBSome and allowed to enter the cilia by the Tectonic complex. Some other proteins may be transported by a non-BBSomal mechanism and granted access to cilia by the Tectonic complex or by NPHP1–4–8 complex. LZTFL1 is a specific regulator of BBSome ciliary trafficking activity. Some other proteins may regulate activities of IFT complex or specific transition zone complexes. Obtaining a complete picture of ciliary protein transport will be the focus of future studies to understand the precise mechanisms of cilia-related disorders.
All animal work in this study was approved by the University Animal Care and Use Committee at the University of Iowa.
Expression vectors for BBS genes were published [8]. Human and mouse LZTFL1 cDNA clones (NM_020347 and NM_033322) were purchased from OpenBiosystems and subcloned into CS2 plasmids with Myc, FLAG, HA, or FS (FLAG and S) tag after PCR amplification. Site-directed mutagenesis was performed by using QuikChange protocol (Agilent) and PfuUltra II Fusion HS DNA polymerase (Agilent). Small interfering RNAs (siRNAs) were purchased from Dharmacon (ON-TARGETplus SMARTpool) and transfected at 100 nM concentration for Gli1 gene expression analysis and at 50 nM concentration for all other experiments with RNAiMAX (Invitrogen) following manufacturer's protocol.
Antibodies against BBS1, BBS2, BBS4 and BBS7 were described previously [9]. To produce rabbit polyclonal antibody for mouse Lztfl1, recombinant NusA-Lztfl1 protein (full-length) was purified using HIS-SELECT Nickel Affinity Gel (Sigma) and used as antigen to immunize rabbits (Proteintech Group). Smo-N antibody was a generous gift from Dr. R. Rohatgi (Stanford University). Other antibodies used were purchased from the following sources: mouse monoclonal antibodies against acetylated tubulin (6–11B-1; Sigma), γ-tubulin (GTU-88; Sigma), LZTFL1 (7F6; Abnova), Myc (9E10; SantaCruz), FLAG (M2; Sigma), HA (F-7; SantaCruz), β-actin (AC-15; Sigma), rabbit polyclonal antibodies against ARL13B (Proteintech Group), BBS7 (Proteintech Group), BBS8 (Sigma), BBS9 (Sigma), γ-tubulin (Sigma), IFT57 (Proteintech Group), IFT88 (Proteintech Group), Smo (Abcam), and rabbit monoclonal antibody against GFP (Invitrogen).
The LAP-BBS4 transgenic mouse line was generated by injecting the LAP-BBS4 expression cassette of the pLAP-BBS4 construct [9] into 1-cell pronuclear stage mouse embryo from B6SJL (C57BL/6J X SJL/J; Jackson Laboratory) strain in the University of Iowa Transgenic Animal Facility. Transgenic animals were maintained in the mixed background of C57BL/6J and 129/SvJ. Genotype was determined by PCR using the following primers (5′-GTCCTGCTGGAGTTCGTGAC-3′ and 5′-GGCGAAATATCAATGCTTGG-3′). Progenies from three founders expressed the LAP-BBS4 transgene. In two lines, LAP-BBS4 levels were less than 50% of endogenous Bbs4 in the testis and only one line expressed LAP-BBS4 higher than endogenous Bbs4 (Figure S1). This line was used for the entire study. Bbs4 knock-out mouse model and hematoxylin/eosin staining was previously described [26].
Testes from six wild-type and LAP-BBS4 transgenic animals were used for TAP. Proteins were extracted with lysis buffer (50 mM HEPES pH 7.0, 200 mM KCl, 1% Triton X-100, 1 mM EGTA, 1 mM MgCl2, 0.5 mM DTT, 10% glycerol) supplemented with Complete Protease Inhibitor cocktail (Roche Applied Science). The remaining TAP procedure was described previously [9]. TAP of FS-LZTFL1 was conducted with HEK293T cells stably expressing FS-LZTFL1 or parental cells. Cell lysates from twenty 15-cm dishes were loaded onto anti-FLAG affinity gel (M2; Sigma), and bound proteins were eluted with 3xFLAG peptide (100 µg/mL; Sigma). Eluate was loaded onto S-protein affinity gel (Novagen), and bound proteins were eluted in 2x SDS-PAGE sample loading buffer. Purified proteins were separated in 4–12% NuPAGE gels (Invitrogen) and visualized with SilverQuest Silver Staining Kit (Invitrogen). Excised gel slices were submitted to the University of Iowa Proteomics Facility and protein identities were determined by mass spectrometry using LTQ XL linear ion trap mass spectrometer (Thermo Scientific). Co-IP was performed as previously described [8].
Protein extract from one 10-cm dish of hTERT-RPE1 cell was concentrated with Microcon Centrifugal Filter Devices (50,000 MWCO; Millipore), loaded on a 4 ml 10–40% sucrose gradient in PBST (138 mM NaCl, 2.7 mM KCl, 8 mM Na2HPO4, 1.5 mM KH2PO4, 0.04% Triton X-100), and spun at 166,400 x Gavg for 13 hrs. Fractions (∼210 µl) were collected from the bottom using a 26 G needle and concentrated by TCA/acetone precipitation. Proteins were re-suspended in equal volume of 2x SDS-PAGE sample loading buffer and analyzed by SDS-PAGE and immunoblotting. Size exclusion chromatography was previously described [8]. Briefly, protein extracts from testis and eye were concentrated by Amicon Ultra-15 (30 kDa; Millipore) and loaded on a Superose-6 10/300 GL column (GE Healthcare). Eluted fractions were TCA/acetone precipitated and re-suspended in 2x SDS loading buffer. The column was calibrated with Gel Filtration Standard (Bio-Rad).
hTERT-RPE1 cells were maintained in DMEM/F12 media (Invitrogen) supplemented with 10% FBS. Immortalized MEF cells expressing Smo-YFP was kindly provided by Dr. M. Scott (Stanford University) and cultured in DMEM with 10% FBS. Primary MEF cells from wild-type and Bbs2 and Bbs4 null embryos [25], [26] were prepared following a standard protocol at embryonic day 13.5. Cells were transfected with siRNAs using RNAiMAX for 48 hrs and further incubated in serum-free medium for 24 hrs for ciliation. For RNA extraction, cells were treated with 100 nM SAG (EMD Chemicals) for additional 18 hrs in a fresh serum-free medium. Total RNA was extracted using TRIzol Reagent (Invitrogen) following manufacturer's instruction. Quantitative PCR was performed as previously described [8]. RPL19 mRNA levels were used for normalization and ΔΔCt method [36] was used to calculate fold inductions. The PCR products were confirmed by melt-curve analysis and sequencing. Knock-down efficiencies of BBS genes were measured by qPCR and samples with more than 90% reduction in BBS gene expression levels were used for GLI1 gene expression analysis. PCR primer sequences are in Table S3. For immunofluorescence microscopy, cells were seeded on glass coverslips in 24-well plates and transfected with siRNAs using RNAiMAX or with plasmid DNAs using FuGENE HD (Roche Applied Science). Cells were cultured for 72 hrs before fixation with the last 30 hrs in serum-free medium to induce ciliogenesis. For SAG treatment, 100 nM SAG was added to fresh serum-free medium and incubated for 4 hrs at 37°C. Cells were fixed with cold methanol, blocked with 5% BSA and 2% normal goat serum in PBST, and incubated with primary antibodies in the blocking buffer. Primary antibodies were visualized by Alexa Fluor 488 goat anti-mouse IgG (Invitrogen) and Alexa Fluor 568 goat anti-rabbit IgG (Invitrogen). Coverslips were mounted on VectaShield mounting medium with DAPI (Vector Lab), and images were taken with Olympus IX71 microscope. For in situ PLA, Duolink in situ PLA kit with anti-mouse PLUS probe and anti-rabbit MINUS probe (OLINK Bioscience) was used with mouse monoclonal antibody for LZTFL1 (Abnova) and rabbit polyclonal antibody for BBS9 (Sigma) following manufacturer's instruction. Briefly, RPE1 cells were seeded onto an 8-well Lab-Tek II chamber slide (Nunc) and treated as for immunofluorescence microscopy until the primary antibody binding step. After washing, cells were decorated with PLA PLUS and MINUS probes (1∶20 dilution) for 2 hrs in a 37°C humidified chamber. Hybridization and ligation of probes, amplification, and final SSC washing were performed per manufacturer's instruction in a humidified chamber. Complex formation was detected by Duolink Detection kit 563 (OLINK Bioscience) and Olympus IX71 microscope.
Homology search, secondary structure prediction, and modeling were performed by using human LZTFL1 protein sequence (NP_065080), SWISS-MODEL Workspace, and InterProScan [37], [38]. Rat Stx1A H3 domain (PDB: 1hvv) was used as a template and aligned with human LZTFL1 aa 212–284.
For paraformaldehyde (PFA) cross-linking, HEK293T cells were incubated with 1% PFA in DMEM at 37°C for 10 min. For cross-linking with photoreactive amino acids, cells were cultured in Dulbecco's Modified Eagle's Limiting Medium (DMEM-LM; Pierce) with L-Photo-Leucine and L-Photo-Methionine (Pierce) supplemented with 10% dialyzed FBS (Pierce). Cross-linking was performed following manufacturer's instruction. Cells were irradiated in the UV Stratalinker 2400 (Stratagene) 5-cm below the UV lamp for 6 minutes. After cross-linking, cells were washed with PBS three times and lysed in the lysis buffer (50 mM Tris pH 7.0, 150 mM NaCl, 0.5% Triton X-100, 0.5% CHAPS, 2 mM EDTA, 2 mM NaF, 2 mM NaVO4) supplemented with Complete Protease Inhibitor cocktail (Roche Applied Science). Protein extracts were loaded onto a 4–12% SDS-PAGE gel and subjected to immunoblotting.
Wild-type mice at 2–3 months of age were used for retinal sections. Animals were perfused with 4% PFA in PBS (2.5 ml/min, 50 ml), and excised eyes were post-fixed for 30 min in the same solution. After washing with PBS, eyes were frozen in OCT and sectioned using cryostat with CryoJane system (myNeuroLab) with 7 µm thickness. For mouse spermatozoa, testis and epididymis from wild-type animals were minced by forceps in Ham's F10 solution (Invitrogen) and incubated for 20 minutes at 37°C in 5% CO2 incubator. The turbid upper fraction (“swim-up” fraction) was collected using wide-opening tips and spread onto positively charged slide glasses. Excess liquid was slowly aspirated and samples were air-dried. After fixation with 4% PFA in PBS for 5 min, samples were processed for immunofluorescence following standard protocol.
|
10.1371/journal.pcbi.1005152 | Interrogating Emergent Transport Properties for Molecular Motor Ensembles: A Semi-analytical Approach | Intracellular transport is an essential function in eucaryotic cells, facilitated by motor proteins—proteins converting chemical energy into kinetic energy. It is understood that motor proteins work in teams enabling unidirectional and bidirectional transport of intracellular cargo over long distances. Disruptions of the underlying transport mechanisms, often caused by mutations that alter single motor characteristics, are known to cause neurodegenerative diseases. For example, phosphorylation of kinesin motor domain at the serine residue is implicated in Huntington’s disease, with a recent study of phosphorylated and phosphomimetic serine residues indicating lowered single motor stalling forces. In this article we report the effects of mutations of this nature on transport properties of cargo carried by multiple wild-type and mutant motors. Results indicate that mutants with altered stall forces might determine the average velocity and run-length even when they are outnumbered by wild type motors in the ensemble. It is shown that mutants gain a competitive advantage and lead to an increase in the expected run-length when the load on the cargo is in the vicinity of the mutant’s stalling force or a multiple of its stalling force. A separate contribution of this article is the development of a semi-analytic method to analyze transport of cargo by multiple motors of multiple types. The technique determines transition rates between various relative configurations of motors carrying the cargo using the transition rates between various absolute configurations. This enables a computation of biologically relevant quantities like average velocity and run-length without resorting to Monte Carlo simulations. It can also be used to introduce alterations of various single motor parameters to model a mutation and to deduce effects of such alterations on the transport of a common cargo by multiple motors. Our method is easily implementable and we provide a software package for general use.
| Molecular motors such as kinesin and dynein facilitate directed transport of intracellular cargo over tracks called microtubules. Inside cells, multiple motor proteins are known to bind and move cargoes. These teams of motors enable the transport of cargoes over longer distances, extending beyond the processive runlengths of a single motor. Impaired transport, possibly due to mutations that affect single motor parameters, is known to cause neurodegenerative diseases. A recent study reported that phosphorylation of a kinesin motor implicated in Huntington’s disease, leads to a reduction in the single motor stalling force. In this work, we investigate how heterogeneity in motor stall forces can affect the coordinated transport properties of multi-motor ensembles. Our model predicts that motors with reduced stall force, even when in the minority, can determine emergent transport properties of average velocity and run-length. Under appropriate external loads, our analysis predicts that motor ensembles containing mutant motors travel longer distances, potentially contributing to the dysregulation of coordinated cargo transport, impairment of neuronal function and the onset of neurodegeneration. These results are enabled by development of a novel semi-analytic methodology to study cargo transport by multiple motors with distinct transport properties. This method is computationally less extensive than existing Monte-Carlo based approaches, easy to implement, and holds potential for understanding how individual motor proteins and properties contribute to the coordination of transport by motor ensembles.
| Motor proteins- kinesin, dynein and myosins- are nanoscale machines that are the main effectors of intracellular transport. They play a critical role in the growth and sustenance of healthy cells by enabling a transport of intracellular cargo over networks of microtubules [1]. Disruption of the functions performed by these molecular motors is linked to neurodegenerative diseases such as Huntington’s, Parkinson’s and Alzheimer’s Disease [2, 3], muscular disorders such as heart disease, uterine complications and high blood pressure. The mechano-chemical behavior of a single motor moving along the microtubule substrate to transport a cargo is relatively well understood [4]. There is considerable prior work on understanding how a single-motor transports a cargo on a microtubule. In [5] the Fokker–Planck equation for a single particle moving in a one-dimensional potential is obtained and used to determine the probability distribution of the position of the particle in a single repeat length of the polymer track via discretization. The underlying approach was extended to a motor connected to a cargo by a linear spring [6, 7] while travelling under the influence of a tilted periodic potential. In vivo it is known that multiple motors work in teams to transport a common cargo [16]; how multiple motor-proteins coordinate to transport a common cargo is not well understood [16–18]. Many studies employ a probabilistic description of the behavior of a single motor protein to construct models that describe how multiple motors transport a common cargo [16, 19–21]. In [22], Gross and coworkers employed Monte-Carlo simulation studies built on a model of a single kinesin in to explore how multiple identical kinesin motors might interact to transport a cargo against a hindering load force. Their work indicates, counter intuitively, the existence of a form of strain-gating, where the motors of an ensemble share loads unequally enabling cargo transport over longer distances. Xu and coworkers examined the effects of ATP concentration on the transport of cargo carried by single and two motors in [23]. At decreased levels of ATP concentration, the velocities of cargoes transported by single and two motor proteins decreases. Coincidentally at decreased ATP concentration there was an an appreciable increase in the run-length of cargoes transported by two motors, while no such effect was seen in the case of transport by one motor. The authors proposed that the increased run-length observed in the presence of two motors results from the lowered dissociation of each motor from the microtubule at decreased ATP concentrations and the increased probability that the cargo stayed bound to the microtubule. Studies such as [21, 29] using probabilistic models of single motors have also predicted in that an ensemble of kinesin motors is a robust system and the robustness increases under high loads [37].
The study of cargo transport by a heterogeneous ensemble of motor proteins composed of both wild type and mutant motors is important to inform our understanding of how mutant motors impact intracellular transport and lead to an onset of diseases. Recent studies have implied that alterations in the kinesin-1 motor domain may have a role in impaired axonal transport. Phosphorylation of a mammal kinesin motor domain by kinase c-Jun N-terminal kinase-3 (JNK3) at a conserved serine residue (Ser-176 in A and C isoforms and Ser-175 in B isoform) is implicated in Huntington’s disease [24]. However the mechanisms affected by Ser-175 phosphorylation are not well understood. An experimental study by Selvin and coworkers in [25] reported that a negative charge at Ser-175, acquired through mutation or phosphorylation, leads to a lower stall force and decreased velocity under external loads of 1 pN or more, while leaving the ATPase, microtubule-binding affnity and processivity unchanged.
Using a semi-analytical method, we reveal surprising emergent transport behaviors arising when a cargo is transported by multiple motor proteins, some of which are mutated and some are not. In particular we analyze the impact of Ser-175 kinesin mutation such as those reported in [25] on cargo transport in the presence of wild type motors. The detailed investigation made possible by our method leads us to hypothesize that under certain conditions, a ‘mooring mechanism’ is activated where proteins moor the cargo to the microtubule and prevent it from being lost. While these mooring proteins do not contribute to the motion of the cargo, they enhance the probability of attachment of other cargo-bound motor proteins to the microtubule that subsequently contributes to an increase in average cargo displacement. The activation of mooring mechanism depends on a number of external factors such as load force and ATP concentration. However, it is also determined by intrinsic properties of the motor protein such as the stall force of the individual motors. Remarkably, mutant motors that have stalling forces matched to the external load force can act dominantly and determine emergent transport properties such as longer run-length, even when they are outnumbered in the ensemble by wild-type motors. Such mechanisms could potentially point to how diseased states emerge and progress coincident with the accumulation of the mutant motor species.
A separate contribution of the article is a semi-analytical method for determining the probability distribution of various configurations of a cargo carried by multiple number and types of motor-proteins. A detailed experimental study of the various modalities of transport by multiple motor proteins (homologous or otherwise) requires significant instrumental resolution than what is needed to investigate single motor behavior. As a consequence, observing the transport dynamics of multiple motors is experimentally challenging [26, 27]. It is further compounded by the combinatorial complexity introduced by the multitude of scenarios possible when many motors and motor types participate in transport. Such challenges motivate the use of analytical and computational tools. The mean-field approach in [19, 21] makes use of simplifying assumptions, such as equal load sharing among all motors, to achieve analytical results, thereby sacrificing significant detail for computational benefits. The Monte Carlo approach in [22], provides better fidelity where complex models can be employed; however, the accuracy of results depend on the number of iterations and on the rarity of the events that occur.
Unlike the Monte-Carlo simulations or any currently implemented simulation methods that study multi-motor ensembles, our Master Equation based method [28] analytically solves for the probability distributions of all possible scenarios at any time point. The methodology is uniquely powerful and enables the calculation of various biologically relevant quantities such as average velocity and run-length, for reasonably sized ensembles and with high accuracy while using lesser computational resources and time as compared to Monte Carlo simulations (the methodology proposed in [6, 7] also obtains results without resorting to Monte Carlo simulations with a single motor attached to the cargo). The underlying concepts behind our methodology are motivated by earlier work reported in [29]. The key enabling concept is that of ‘relative configuration’ of motor proteins, determining the transition probabilities between relative configurations from the transition probabilities of the absolute configuration space and subsequently determining the biologically relevant quantities from the relative configuration space. The computation engine is implemented using MATLAB and can be used to simulate cargo transport by any two unidirectional species.
Our method provides a general platform to study the transport of cargo by multiple motors of two different types where each type of motor protein can be individually characterized by a probabilistic model describing its stepping, detachment and attachment rates. For this article, the technique has been utilized to introduce alterations of various parameters from the nominal ones to model a mutation of the serine residue and compute the effect of such a mutation on cargo carried by a mixture of wild-type and mutant motor-proteins. In summary, we developed a simulation engine to study the transport of cargo by multiple motor proteins with distinct properties that in concert can exhibit emergent transport behaviors.
In this study, we investigate the impact of a previously reported kinesin mutation on the transport of a motor ensemble and its attached cargo. The mutation, located within the motor domain of the kinesin, mimics the phosphorylated state of Ser-175 [25]. The motor domain phosphomimetic mutation does not affect the ATPase, microtubule-binding affinity or processivity of the motor, but does reduce the stall force and velocity of kinesin under a load force. The in vitro phosphorylation of Ser-175 for a full length kinesin similarly reduces stall force and velocity of the motor. Both the mutant and in vitro phosphorylated kinesin showed no other aberrant single motor behavior.
Here we use computational modelling to analyze the impact of the Ser-175 mutant kinesin on a heterogenous ensemble of motors and its transported cargo. The wild-type and mutant kinesin motors have a different stalling force; the wild type kinesin has a stall force of Fs = 6pN and the mutated kinesin has a reduced stall force of F ¯ s = 5 . 5 p N(see [25] for more detail). In our analysis we considered cargoes transported by the following motor ensembles: cargo with two wild-type (WW) motors, one wild-type and one mutant (WM) motors, two mutant (MM) motors, three wild-type (WWW) motors, two wild-type and one mutant (WWM) motors, one wild-type and two mutant (WMM)motors, and three mutant (MMM) motors.
The knowledge of the steady state conditional probability distributions allows one to compute biologically relevant quantities such as average velocity and run-length. The results obtained for ensembles containing two and three motors for varying load forces are reported here.
The Master Equation based methodology used to obtain the aforementioned results is described in this section. This method is used to study emergent properties of an ensemble of multiple motors of two types, that can each take a step on, detach from or reattach onto the microtubule. The knowledge of transition probabilities of stepping, detachment and attachment enable the determination of transition rates between various absolute configurations of the motors, allowing for the calculation of transition rates between the corresponding relative configurations. These rates enable the calculation of the probability distribution of the various ensemble configurations, thereby facilitating the computation of several biologically relevant quantities such as average velocity, run-length and number of attached motors.
We begin with the construction of the relative state space, along with calculations necessary to arrive at several biologically relavent quantities. Then, a general methodology to obtain the transition rates between absolute configurations given the knowledge of the probability rates of stepping, detachment and reattachment to the microtubule for a wild-type motor(PS, PD, PA) and probability rates for a mutant motor ( P ¯ S , P ¯ D , P ¯ A ) is presented. Finally, the model used to determine the probability rates for kinesin motor proteins is detailed.
Consider a cargo that is carried by both wild-type and mutant motor proteins on a microtubule. The microtubule is modeled as directed linear lattice formed by equally sized dimers with dimension d. Here the kth dimer is located at location a ¯ k = k d and indexed by the set of integers I = {…, −2, −1, 0, 1, 2, …}. Each motor protein bound to the cargo can attach, take a forward step or detach from the microtubule. The absolute configuration Ω≔{ Ωh,kΩd,k }k∈I of motor-protein arrangement on the microtubule specifies the number, Ωh,k, of wild-type motor proteins and the number, Ωd,k, of mutant motor proteins at the kth location on the microtubule.
For example, the absolute configuration of the ensemble of motors illustrated in Fig 16 is given by
Ω = [ ⋯ Ω - 1 Ω 0 Ω 1 Ω 2 Ω 3 Ω 4 Ω 5 Ω 6 ⋯ ] = ⋯ 0 1 0 1 0 0 0 1 ⋯ ⋯ 0 0 0 1 0 1 0 0 ⋯ .
The relative configuration of an ensemble of motors is represented using a string of three symbols. Given an absolute configuration we first identify the rearguard motor which is the motor that is attached to the microtubule and lags behind all the other motors on the microtubule. Using the location of the reargurad motor as a reference, the relative configuration ϑ is obtained as a string of three symbols ‘Mh’, ‘Md’ and ‘|’, where ‘Mh’ and ‘Md’ denote wild-type and mutant motors respectively, with ‘|’ denoting a separator that distinguishes different microtubule locations. The motor located the farthest from the rearguard motor on the microtubule is identified as the vanguard motor.
For example, the relative configuration of the ensemble in Fig 17(a) is the string ‘|Mh Md||Md||Mh|’. The configuration that results after the furthermost mutant motor in Fig 17(a) takes a step is shown in Fig 17(b) which has a relative configuration given by ‘|Mh Md|||Md|Mh|’.
Both the mutant as well as the wild-type motor proteins are characterized via their own set of stepping, attachment, and detachment probabilities(for wild-type (PS, PD, PA) and for mutant ( P ¯ S , P ¯ D , P ¯ A )). The individual motors for both species are modeled as hookean springs when stretched that offer no resistance when compressed. A single motor is assumed to have a linkage rest length L0 and spring stiffness constant Ke. Motors of both the species are assumed to not step backward and are bound to the cargo particle irreversibly. More complex models of motor-proteins can be easily accommodated. It is further assumed that the there exists a force Fstall called the stalling force, where if the force on the motor protein F ≥ Fstall then the motor does not take a forward step and stepping probability is zero [31]. The stalling force can be that is also used to estimate how many motors are carrying the cargo. For an ensemble of wild-type and mutant motors carrying a cargo, the following result holds:
Result 1: Given an ensemble of M molecular motors attached to a common cargo that is subjected to a load force Fload, the distance between the rearguard and the vanguard motor is bound by
n
=
m
a
x
(
M
+
1
)
m
a
x
(
F
s
,
F
¯
s
)
-
F
l
o
a
d
K
e
+
d
,
F
l
o
a
d
K
e
+
2
L
0
(1)
where Fs is the minimum load force for which the stepping probability of the wild-type motor protein becomes zero (i.e. the stalling force for the wild-type motor protein), F ¯ s is the minimum load force for which the stepping probability of the mutant motor protein becomes zero (i.e. the stalling force for the mutated motor protein), L0 is the rest length of the motor linkage, Ke is the linkage stiffness and d is the step-size of the motor.
A detailed derivation is provided in the S1 Text.
It is to be noted that the absolute configuration space admits infinitely many representations as there is always a small probability of finding the cargo at any location on the microtubule. However, the above result concludes, that given a stall force for both wild-type and mutant motors the relative configuration space is finite, since there are no motors beyond n units away from the rearguard motor in any relative configuration.
In this section we present a general scheme for determining transition rates between absolute configurations. We begin with a structural model for single motor protein that consists of motor head, stalk and cargo binding tail domain. The linkage between the motor-heads and tail for single motor is modeled as a hookean spring when stretched, that has a rest length L0. It offers no resistance when compressed [22]. The motor heads move along the microtubules exerting a force F on a cargo that is expressed as a function of its length L by,
F ( L ) = K e ( L - L 0 ) if L ≥ L 0 , 0 if | L | < L 0 , K e ( L + L 0 ) if L ≤ - L 0 .
Zeq is the mean position of the cargo that is its equilibrium position determined by the forces exerted by the motors on the cargo through their linkages and the load force Fload on the cargo. The mean cargo position for a fixed Fload > 0 is a function of the absolute configuration i.e. Zeq = Zeq(Ω). If the cargo position is assumed to follow a truncated Gaussian distribution Θ(z) with variance σ, its probability density Θ(z) for |z| < 3σ is given by,
Θ ( z ) = ( e - z 2 2 σ 2 ) / ( 2 ∫ 0 3 σ e - z 2 2 σ 2 d z ) .
The effect of thermal noise can be incorporated by determining the steady state variance σ of the cargo position. We assume that when a motor in the ensemble takes a forward step or detaches, the probability distribution of the cargo position reaches a new distribution with negligible transient. Here, the time scale of the cargo dynamics is assumed to be faster than the rate of change of motor configurations.
A transition to another configuration Ω′ occurs if either the wild type or mutant motor at a location a ¯ k steps forward to a ¯ k + 1, detaches from the location a ¯ k or reattaches to the location a ¯ k on the microtubule. By representing Ω′ as Ω + S, S is a sequence that corresponds to the type of transition(step, detach or attach) and the type of motor(wild-type or mutant) that has transitioned. The transition rate from Ω to Ω′ is determined by averaging the associated probability rate over the position of the cargo.
The model of a single motor-protein is specified via the probability PS(F) of the motor taking a step, the detachment probability, PD(F), of the motor detaching from the microtubule, and the probability of attachment PA of an unattached motor-protien to the microtubule, per second. Here F is the force acting on the motor which is considered positive if it is directed opposite to the motor stepping direction (e.g. kinesin forward stepping is towards the mictorubule + end). Here in order to calculate the transition rates between absolute configurations it is assumed that the probability rates of step, detachment and attachment are known; later we illustrate a way to compute these probabilities for kinesin motors.
The results put forth in this article correspond to an instantiation of our methodology for kinesin motor protein. The probability rates of stepping, detachment and attachment of a single kinesin motor are determined using several available studies [16, 21, 30, 31].
|
10.1371/journal.pcbi.1000409 | Spatio-Temporal Dynamics of Yeast Mitochondrial Biogenesis:
Transcriptional and Post-Transcriptional mRNA Oscillatory Modules | Examples of metabolic rhythms have recently emerged from studies of budding
yeast. High density microarray analyses have produced a remarkably detailed
picture of cycling gene expression that could be clustered according to
metabolic functions. We developed a model-based approach for the decomposition
of expression to analyze these data and to identify functional modules which,
expressed sequentially and periodically, contribute to the complex and intricate
mitochondrial architecture. This approach revealed that mitochondrial
spatio-temporal modules are expressed during periodic spikes and specific
cellular localizations, which cover the entire oscillatory period. For instance,
assembly factors (32 genes) and translation regulators (47 genes) are expressed
earlier than the components of the amino-acid synthesis pathways (31 genes). In
addition, we could correlate the expression modules identified with particular
post-transcriptional properties. Thus, mRNAs of modules expressed
“early” are mostly translated in the vicinity of
mitochondria under the control of the Puf3p mRNA-binding protein. This last
spatio-temporal module concerns mostly mRNAs coding for basic elements of
mitochondrial construction: assembly and regulatory factors. Prediction that
unknown genes from this module code for important elements of mitochondrial
biogenesis is supported by experimental evidence. More generally, these
observations underscore the importance of post-transcriptional processes in
mitochondrial biogenesis, highlighting close connections between nuclear
transcription and cytoplasmic site-specific translation.
| In bacterial and eukaryotic cells, gene expression is regulated at both the
transcriptional and translational levels. In eukaryotes these two processes
cannot be directly coupled because the nuclear membrane separates the
chromosomes from the ribosomes. Although the transcription levels in different
cellular conditions have been widely examined, genome-wide post-transcriptional
mechanisms are poorly documented and therefore, the connections between the two
processes are difficult to explain. In this work, the time-regulated expression
of the genes involved in the construction of the mitochondrion, an important
organelle present in nearly all the eukaryotic cells, was scrutinized both at
transcriptional and post-transcriptional levels. We observed that temporal
transcriptional profiles coincide with groups of genes which are translated at
specific cellular loci. The description of these relationships is functionally
relevant since the genes which are transcribed early in mitochondria cycles are
those which are translated to the vicinity of mitochondria. In addition, these
early genes code for essential assembling factors or core elements of the
protein complexes whereas the peripheral proteins are translated later in the
cytoplasm. Also, these observations support the concerted action of important
regulatory factors which control either the gene transcription level
(transcription factors) or the mRNA localization (mRNA-binding proteins).
| Cell construction requires the tight linking of various molecular processes, from
nuclear transcription to the site-specific production of proteins. The control of
the orchestration of these processes remains poorly understood. In classical
experimental conditions, coordinated waves of transcription are difficult to observe
because of the metabolic asynchrony of the cells in growing cultures. A yeast system
with properties avoiding these difficulties was recently described. In well-defined
continuous cultures of Saccharomyces cerevisiae, the oxygen
consumption rate oscillates with a constant period [1], implying that
cell-to-cell signaling synchronizes oxidative and reductive functions in the
culture. The gene-expression dynamics of the yeast metabolic cycle is therefore a
useful model system for studies of the lifecycle of groups of transcripts in
eukaryotic cells [2]. Indeed, microarray studies have demonstrated
periodicity in the expression of the yeast genome, and consequently the existence of
similar temporal expression patterns in functionally connected groups of genes [3]. Genes
specifying functions associated with energy appeared to be expressed with
exceptionally robust periodicity, consistent with the variations in the amount of
dissolved oxygen in the medium of synchronized culture. In pioneering studies [4], it was
shown that yeast mitochondrial morphology oscillates in response to energetic
demands driven by the ultradian clock output.
In this work, our purpose was to distinguish temporal gene clusters, which may allow
describing a biologically relevant scenario of mitochondria biogenesis. Depending on
the addressed points and on the quality of the microarray data, several methods such
as SVD (Singular Value decomposition), PCA (Principal Components Analysis),
self-organizing maps, wavelet multiresolution decomposition and FFT (Fast Fourier
Transform) have been used to analyze relevant transcript data [5]. We decided to use a
model-based approach [6] to decomposition of published expression data for
the 626 oscillating nuclear genes encoding mitochondrial proteins. We established a
classification of these genes into temporal groups, which cover the 5-hour long
metabolic cycle, and present a dynamic and global picture of mitochondrial
biogenesis. These temporal groups correlate both with particular functional
properties of the corresponding proteins and with specific translational sites in
the cell. This global description of mitochondrial transcriptome clusters in
temporal phases is consistent with the concept of RNA regulons, according to which
post-transcriptional RNA operons may constitute an important element of eukaryotic
genome expression [7],[8].
Microarray data from the study by Tu et al. [3] were
collected from the Gene Expression Omnibus database [9], under accession
number GSE3431. This dataset comprised the normalized gene expression values,
i.e. the median of each array (all data points) equal 1,
used by Tu et al.
[3] in
their pioneering analysis. Tu et al.
[3]
performed microarray experiments at 25-minute intervals, over three consecutive
metabolic cycles (the length of one cycle is ∼300 minutes). For each
gene, expression measurements were thus available for 36 successive time points.
We considered only those genes for which expression measurements were available
and which (i) displayed significant periodic patterns, as
defined by Tu et al.
[3]
(∼3552 genes with a confidence level greater than 95%) and
(ii) were identified as involved in mitochondrial
biogenesis, as defined by Saint-Georges et al.
[10] (∼794 genes). The resulting expression
matrix comprised data for 626 genes (the complete list is available in Dataset
S1).
To cluster genes whose RNA level peaks at the same time points in the yeast
metabolic cycle (YMC), we used the ω
-values
obtained for each gene using EDPM algorithm (see previous paragraph).
Pearson correlation coefficients (
r
) were calculated between all
W
vector pairs, and hierarchical cluster analysis was applied. This
classical clustering method can be summarized as follows: (1) Distances (
d
) between all
W
vector pairs is calculated using Pearson's correlation analysis (
d = 1−r
); (2) The resulting distance matrix is thoroughly inspected to find the
smallest distance; (3) The corresponding genes are joined together in the tree
and form a new cluster; (4) The distances between the newly formed cluster and
the other genes are recalculated; (5) Steps 2, 3 and 4 are repeated until all
genes and clusters are linked in a final tree.
We searched for cis-acting signals in 3′ and
5′UTR sequences, using motifs predicted by the MatrixREDUCE algorithm
[11]. For 3′UTR signals, we tested several
motifs identified in previous studies [10],[12] as
possible binding sites for mRNA stability regulators in Saccharomyces
cerevisiae. For 5′ UTR signals, we examined upstream
regions between nucleotide positions −600 and −1 and
searched for motifs between 1 and 7 nt long. We assessed whether any of the
signals were observed at a frequency greater than that expected by chance, by
calculating p-values as described in [13] (hypergeometric
distribution). We then search the YEASTRACT database for transcription factors
with DNA-binding sites matching the motifs identified with MatrixREDUCE [14].
The EDPM algorithm was implemented in R programming language (http://cran.r-project.org/) and functions were numerically
minimized using the quasi-Newton method (R function available in the BASE
package). Hierarchical clustering was carried out with the
“hclust” function (also available in R programming
language), with the “ward” method for gene agglomeration.
MatrixREDUCE source code is freely available online from http://bussemaker.bio.columbia.edu/software/MatrixREDUCE/ and
was used for analyses of upstream sequences with default parameters (see the
documentation available online for more information).
All the strains used in this study are isogenic to BY4742 (MATα;
his3Δ1; leu2 Δ0; lys2 Δ0; ura3Δ0) from the
Euroscarf gene deletion library.
This analysis leads to the prediction that unknown cluster A genes translated in
the vicinity of mitochondria in a Puf3p-dependent way (class I) are likely to be
involved in early steps of mitochondria biogenesis. To test this experimentally,
we examined the properties of nine strains carrying deletions of uncharacterized
cluster A/class I genes (Figure
6B and Dataset S3). For each mutant strain, we checked the ability to grow
on non fermentable carbon sources and tested the assembly of respiratory
complexes III and IV by recording cytochrome spectra (see Methods). Disturbance of early steps of mitochondrial
biogenesis —for example replication of mitochondrial DNA,
mitochondrial transcription and translation— can affect maintenance of
mitochondrial DNA [18], we also tested whether these mutant strains
retained the mitochondrial chromosome by measuring the production of petite
cells (rho−). The phenotypes of these deleted strains are
presented in Figure 6C.
Strikingly, seven out of the nine gene-deleted strains displayed severe
respiratory dysfunctions (poor growth on non-fermentable media) and/or
alterations in their cytochrome spectra. These phenotypes strongly suggest that
most of the unknown phase A/class I genes have functions in mitochondrial
transcription/translation or assembly of respiratory complexes. This is strongly
in favour of the idea that during this short period (phase A lasts only 25
minutes, Figure 4A), there
is a surge in the abundance of mRNAs important for mitochondrial biogenesis and
that they are translated at particular subcellular localization.
It was recently observed [3],[19] that yeast cells
can be synchronized and exhibit synchronous waves of storing and then burning
carbohydrates. Using microarrays, it was shown that many nuclear genes coding
for mitochondrial proteins, have their mRNAs which oscillate and peak at a time
when highest rate of respiration has passed. It was suggested [19]
that cells are either rebuilding or duplicating their mitochondria at this time.
We took advantage of these data to better analyze the mitochondria rebuilding
program and identified new gene clusters reflecting spatio-temporal groups of
gene expression. Our findings are entirely consistent with the notion of RNA
regulons [7],[8], according to which
mRNA-binding proteins (RBP) play an important role, coordinating the various
post-transcriptional events. We show here that 262 mRNAs coding for important
mitochondrial proteins (assembly factors, ribosomal proteins, translation
regulators) are coordinately and periodically present in increasing amounts
early in the mitochondrial cycle (phase A = 25
minutes). In addition, most of these mRNAs are specifically localized in the
vicinity of mitochondria under the control of the protein Puf3p. This suggests
that during this particular time-window, Puf3p acts in the control of mRNA
localization/translation. During the rest of the mitochondrial cycle, Puf3p may
function (possibly in association with other RBPs) either in mRNA degradation
[20] or in the control of bud-directed mitochondrial
movement [21]. Following this early phase A, phases B (50
minutes) and C (50 minutes) concern elements of the fundamental mitochondrial
machineries (respiratory chain complexes, TCA cycle, etc.). Undoubtedly, this
chronology of events should reflect the logic of mitochondria construction.
This point can be illustrated with the well-documented assembly process of
cytochrome c oxidase (COX) [22],[23], a fascinating
process involving the sequential and ordinate addition of 11 subunits to an
initial seed consisting of Cox1p (Table 1, “core” and “shield
proteins”). In addition to the structural subunits, a large number of
accessory factors are required to build the holoenzyme. Unexpectedly, we found
that all the mRNAs for these accessory factors are relatively abundant early in
mitochondrial biogenesis, that is during phase A. Cluster A includes genes whose
expression is essential for a preliminary step, consisting of the synthesis of
all the elements (RNA polymerase, ribosomes, translation factors) required for
mitochondrial production of Cox1p; this step is followed by the construction of
the core enzyme (Cox1p+Cox2p+Cox3p). We also observed that the
mRNAs coding for the 18 assembly factor transcripts involved in COX assembly
[22],[24] are mostly found
during phase A (Table 1,
“assembly factors”) and, in addition, all but one are
translated in the vicinity of mitochondria under the control of Puf3p (MLR class
I, [10]). The situation is very different for structural
COX proteins (shield proteins of the complex). Except for Cox5A, all the
corresponding mRNAs are found in phase B, indicating that the corresponding
genes are expressed after those of phase A. Unlike phase A mRNAs, they are all
translated on free cytoplasmic polysomes (MLR class III, [10]). This
scenario agrees with the previous biochemical description of short intermediates
[23]; especially interesting is the observation that
Cox5Ap, found here in phase A, was previously identified as the first structural
protein added to the S2 complex [23]. The properties
of COX assembly described here are common to the other respiratory chain
complexes. The mRNAs for assembly factors mostly peak in phase A and they are
translated close to mitochondria, under the control of Puf3p; they initiate the
formation of respiratory complexes by the successive addition of structural
proteins whose mRNAs mostly peak in phase B. This is the first evidence that, at
least in the conditions described in [3], the construction of the
respiratory chain is one of the first steps of mitochondrial biogenesis; indeed,
all the production machinery (assembly factors, translation, etc.) are available
in phase A to produce and assemble the protein complexes in phase B.
Genes coding for mitochondrial proteins can be classified into two different
regulatory systems. This dichotomy is well illustrated in the case of OXPHOS
complexes coding genes. The first class corresponds to mRNAs translated to the
vicinity of mitochondria, mainly present in phase A and which code, for
instance, for assembly factors. Genes of the second class code for structural
proteins, and are found mainly in phases B or C during which transcription
regulation is the major mechanism. Previous studies suggested that genes coding
for assembly factors are not transcriptionally regulated [25]. We confirmed
and extended these preliminary observations by showing that genes encoding
assembly factors: (i) are expressed before genes encoding
structural proteins, (ii) have a functional Puf3p binding site
which controls localization/translation to the vicinity of mitochondria and may
thus generate discrete foci on the matrix face of the mitochondrial membrane,
and (iii) do not contain any evident signals in their
5′UTR, a feature which distinguishes them from the genes encoding
structural proteins. The mRNAs for translation and assembly factors are all
expressed only during phase A, but mRNAs for structural proteins are found
during phases A, B and to a lesser extent C. This is likely to reflect the
timing of the building of the various complexes. Thus, for instance, COX
assembly requires an intact functional ATPase [22], which is in
agreement with the fact that mRNAs for ATPase structural proteins are mostly
found in phase A (see Dataset S4) whereas the COX equivalents are
mostly in phase B (see Dataset S4). Also, unlike genes encoding
assembly factors, genes coding for structural proteins of the respiratory chain
complexes are mainly controlled transcriptionally. According to the
environmental conditions, either Hap4p (depending on carbon availability [26],[27]) or Hap1p
(depending on oxygen concentration [28]), regulate the transcription of nuclear genes
coding for structural proteins. Binding sites for these two transcription
factors are present significantly more frequently than expected from a random
distribution in the genes of clusters A, B and C (Figure 5A). In addition, the amounts for both
HAP4 mRNA and HAP1 mRNA also oscillate and
peak in phases A and C, respectively (Figure 5C). HAP1 mRNA
variation is interesting because Hap1p can repress its own transcription and may
act either as a repressor or as an activator, depending on oxygen levels [29].
It was observed that fluctuating levels of O2 dissolved in the
culture, indicates changing activities of mitochondrial oxygen consumption and
cellular redox switching [30]. Thus, Hap1p, which is an oscillating redox
sensor, is an excellent candidate to signal the transition between
non-respiratory rebuilding and respiratory phases (Figure 5D).
Overall, we report a comprehensive picture of the biogenesis of yeast
mitochondria and illustrate spatio-temporal differences between groups of
nuclear genes. The unexpected finding that transcriptionally or
post-transcriptionaly regulated groups of genes are expressed both at different
times and translated in different places may be of relevance to mitochondria in
other species. Indeed, mammalian β F1-ATPase mRNA is found in
the outer membrane and is translated, under the control of 3′UTR
signals and RNA-binding proteins [31], only during
cell cycle phase G2/M [32]; this gives credence to the general
applicability of our observations. Studies with human cells are currently
underway to assess the similarities and differences between yeast and human
cells regarding these aspects of mitochondrial biogenesis.
|
10.1371/journal.ppat.1002016 | Human Cytomegalovirus IE1 Protein Elicits a Type II Interferon-Like
Host Cell Response That Depends on Activated STAT1 but Not
Interferon-γ | Human cytomegalovirus (hCMV) is a highly prevalent pathogen that, upon primary
infection, establishes life-long persistence in all infected individuals. Acute
hCMV infections cause a variety of diseases in humans with developmental or
acquired immune deficits. In addition, persistent hCMV infection may contribute
to various chronic disease conditions even in immunologically normal people. The
pathogenesis of hCMV disease has been frequently linked to inflammatory host
immune responses triggered by virus-infected cells. Moreover, hCMV infection
activates numerous host genes many of which encode pro-inflammatory proteins.
However, little is known about the relative contributions of individual viral
gene products to these changes in cellular transcription. We systematically
analyzed the effects of the hCMV 72-kDa immediate-early 1 (IE1) protein, a major
transcriptional activator and antagonist of type I interferon (IFN) signaling,
on the human transcriptome. Following expression under conditions closely
mimicking the situation during productive infection, IE1 elicits a global type
II IFN-like host cell response. This response is dominated by the selective
up-regulation of immune stimulatory genes normally controlled by IFN-γ and
includes the synthesis and secretion of pro-inflammatory chemokines.
IE1-mediated induction of IFN-stimulated genes strictly depends on
tyrosine-phosphorylated signal transducer and activator of transcription 1
(STAT1) and correlates with the nuclear accumulation and sequence-specific
binding of STAT1 to IFN-γ-responsive promoters. However, neither synthesis
nor secretion of IFN-γ or other IFNs seems to be required for the
IE1-dependent effects on cellular gene expression. Our results demonstrate that
a single hCMV protein can trigger a pro-inflammatory host transcriptional
response via an unexpected STAT1-dependent but IFN-independent mechanism and
identify IE1 as a candidate determinant of hCMV pathogenicity.
| Human cytomegalovirus (hCMV) is a leading cause of birth defects and severe
disease in people with compromised immunity. Disease caused by hCMV is
frequently linked to inflammation, and the virus has been shown to induce
numerous host genes many of which encode pro-inflammatory proteins. However,
little is known about the contributions of individual viral proteins to these
changes in cellular transcription. We systematically analyzed the effects of the
hCMV immediate-early 1 (IE1) protein, a major viral transcriptional activator,
on expression of >28,000 human genes. Following expression under conditions
mimicking the situation during hCMV infection, IE1 elicited a transcriptional
response dominated by the up-regulation of pro-inflammatory and immune
stimulatory genes normally induced by the secreted signaling protein
interferon-γ. However, IE1-mediated gene expression was independent of
interferon induction, yet required the activated form of signal transducer and
activator of transcription 1 (STAT1), a central mediator of interferon
signaling. Indeed, STAT1 moved to the nucleus and became associated with IE1
target genes upon expression of the viral protein. Our results demonstrate that
a single hCMV protein can trigger a pro-inflammatory host cell response via an
unexpected mechanism and suggest that IE1 may contribute to hCMV disease in more
direct ways than previously thought.
| Human cytomegalovirus (hCMV), the prototypical β-herpesvirus, is an extremely
widespread pathogen (reviewed in [1]). Primary hCMV infection is invariably followed by
life-long viral persistence in all infected individuals. The groups most evidently
affected by hCMV disease are humans with acquired or developmental immune deficits
including allograft recipients receiving immunosuppressive drugs, human
immunodeficiency virus-infected individuals, cancer patients undergoing intensive
chemotherapy, and infants infected in utero (reviewed in [2]). In
immunologically normal hosts, clinically relevant symptoms rarely accompany acute
infections (reviewed in [3]), but viral persistence may contribute to chronic disease
conditions including atherosclerosis, cardiovascular disease, inflammatory bowel
disease, immune senescence, and certain malignancies (reviewed in [4], [5], [6], [7], [8]).
The pathogenesis of disease (e.g., pneumonitis, retinitis, hepatitis, enterocolitis,
and encephalitis) associated with acute hCMV infection in immunocompromised people
is most readily attributable to end organ damage either directly caused by
cytopathic viral replication or by host immunological responses that target
virus-infected cells. In contrast, chronic disease associated with persistent hCMV
infection in immunocompetent individuals as well as in the allografts of transplant
recipients is most likely related to prolonged inflammation (reviewed in [9]). In fact, hCMV
has been frequently detected in the midst of intense inflammation, and a myriad of
studies from transplant recipients and normal hosts have presented a strong case for
this virus as an etiologic agent in chronic inflammatory processes, particularly
those resulting in vascular disease (reviewed in [4]). At the molecular level, this
is reflected by the fact that, in both human cells and animal models,
cytomegalovirus infections activate numerous host genes many of which encode growth
factors, cytokines, chemokines, and adhesion molecules with pro-inflammatory and
immune stimulatory activities [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23]. A number of these virus-induced proteins are released
from infected cells forming the viral “secretome” [4], [24], [25].
A large proportion of human genes that undergo activation during hCMV infection are
normally controlled by interferons (IFNs) (reviewed in [26], [27]). The IFNs constitute a
distinct group of cytokines synthesized and released by most vertebrate cells in
response to the presence of many different pathogens including hCMV. They are
divided among three classes: type I IFNs (primarily IFN-α and IFN-β), type
II IFN (IFN-γ), and type III IFNs (IFN-λ or interleukin 28/29). The type I
IFNs share many biological activities with type III IFNs, especially in host
protection against viruses. IFN-γ, the sole type II IFN, is one of the most
important mediators of inflammation and immunity exerting pleiotropic effects on
activation, differentiation, expansion and/or survival of virtually any cell type of
the immune system (reviewed in [28]). A significant body of research has identified the
primary IFN pathway components and has characterized their roles in
“canonical” signaling (reviewed in [29], [30]). In this pathway, IFNs bind to
their cognate cell surface receptors to induce conformational changes that activate
the receptor-associated enzymes of the Janus kinase (JAK) family. The
post-translational modifications that follow this activation create docking sites
for proteins of the signal transducer and activator of transcription (STAT) family
with seven human members. In turn, the STAT proteins undergo JAK-mediated
phosphorylation at a single tyrosine residue (Y701 in STAT1), which triggers their
transition to an active dimer conformation. The STAT dimers accumulate in the
nucleus where they may recruit additional proteins, and these complexes then bind
sequence-specifically to short DNA motifs termed IFN-stimulated response element
(ISRE) or gamma-activated sequence (GAS). ISREs are usually bound by a ternary
complex composed of a STAT1-STAT2 heterodimer and IFN regulatory factor (IRF) 9,
which forms upon induction by type I and type III IFNs and is referred to as
IFN-stimulated gene factor 3 (ISGF3). In contrast, type II IFN typically signals via
STAT1 homodimers that associate with GAS elements. Finally, promoter-associated STAT
proteins stimulate transcription of numerous IFN-stimulated genes (ISGs) via their
carboxy-terminal transcriptional activation domain. Within this domain,
phosphorylation of a serine residue (S727 in STAT1) can augment STAT transcriptional
activity. To some extent, the complex responses elicited by type I, type II, and
type III IFNs are redundant as a consequence of partly overlapping ISGs.
Since many ISGs, especially those induced by type I IFNs, exhibit potent anti-viral
activities most viruses have evolved escape mechanisms that mitigate IFN responses.
In fact, both hCMV and murine cytomegalovirus (mCMV) are known to disrupt IFN
pathways at multiple points (reviewed in [26], [27]). For example, JAK-STAT
signaling is inhibited by the hCMV 72-kDa immediate-early 1 (IE1) gene product [31], [32], [33], a key
regulatory nuclear protein required for viral early gene expression and replication
in fibroblasts infected at low input multiplicities [34], [35], [36]. IE1 orthologs of mCMV and
rat cytomegalovirus (rCMV) also contribute to replication and virulence in the
respective animals [37], [38]. The hCMV IE1 protein counteracts virus- or type I
IFN-induced ISG activation via complex formation with STAT1 and STAT2 resulting in
reduced binding of ISGF3 to ISREs [31], [32], [33], [39]. STAT2 interaction contributes to hCMV type I IFN
resistance and to IE1 function during productive infection [33], but the viral protein
undergoes many additional host cell interactions (reviewed in [2], [40], [41]). For example, IE1 targets
subnuclear structures known as promyelocytic leukemia (PML) bodies or nuclear domain
10 (ND10) ([42], [43], [44]; reviewed in [45], [46], [47], [48]). In addition, IE1 associates
with chromatin [49] and interacts with a variety of transcription regulatory
proteins [50],
[51], [52], [53], [54], [55], [56], [57]. Consequently,
IE1 stimulates expression from a broad range of viral and cellular promoters in
transient transfection assays. However, IE1-mediated activation or repression of
merely a few single endogenous human genes has been demonstrated so far [58], [59], [60], [61], [62], [63], [64].
Here we present the results of the first systematic human transcriptome analysis
following expression of the hCMV IE1 protein. Surprisingly, the predominant response
to IE1 was characterized by activation of pro-inflammatory and immune stimulatory
genes normally controlled by IFN-γ. We further demonstrate that IE1 employs an
unusual mechanism, which does not require induction of IFNs but nonetheless depends
on activated (Y701-phosphorylated) STAT1, to up-regulate a subset of ISGs.
The hCMV IE1 protein exhibits complex activities, and results obtained from
experiments with IE1 mutant virus strains are inherently difficult to interpret.
In fact, regarding the phenotype of IE1-deficient viruses at low input
multiplicities, it seems almost impossible to discriminate between effects
directly linked to any of the IE1 activities and indirect effects caused by
delays in downstream viral gene expression and replication. On the other hand,
following infection at high multiplicity, many consequences of absent IE1
expression are compensated for by excess viral structural components, such as
tegument proteins and/or DNA, and therefore undetectable ([35], [36]; reviewed in [2], [40], [41]). Thus,
it is apparent that cells with inducible expression of functional IE1 at
physiological levels would be highly useful by allowing a definite assessment of
the viral protein's activities outside the confounding context of
infection. Furthermore, such cells would avoid potential difficulties typically
associated with transient transfection, including variable frequency of positive
cells and protein accumulation to non-physiologically high levels. Importantly,
an inducible expression system would also preclude cells from adapting to
long-term IE1 expression. In fact, the continued presence of IE1 is reportedly
incompatible with genomic integrity and normal cell proliferation [65], [66], [67].
We used a tetracycline-dependent induction (Tet-on) system built into lentivirus
vectors to generate cells in which IE1 expression can be synchronously induced
and compared to cells not expressing the viral protein. The first component of
this system is a lentiviral vector (pLKOneo.CMV.EGFPnlsTetR; [68], [69], [70]) that
includes a hybrid gene encoding the tetracycline repressor (TetR) linked to a
nuclear localization signal (NLS) derived from the SV40 large T antigen and the
enhanced green fluorescent protein (EGFP) to produce an EGFPnlsTetR fusion
protein [68]. In addition, this vector encodes neomycin
resistance. The second component is a lentivirus vector (pLKO.DCMV.TetO.cIE1)
conferring puromycin resistance, in which a fragment of the hCMV
promoter-enhancer drives expression of the IE1 (Towne strain) cDNA. In this
vector, tandem tetracycline operator (TetO) sequences are present immediately
downstream of the TATA box. For the lentivirus transductions, we chose MRC-5
primary human embryonic lung fibroblasts, because they support robust wild-type
hCMV replication, whereas IE1-deficient virus strains exhibit a severe growth
defect after low multiplicity infection of these cells ([31], [33] and Figure 1 C). Initially, low passage MRC-5
cells were transduced with lentivirus prepared from plasmid
pLKOneo.CMV.EGFPnlsTetR, and a neomycin-resistant polyclonal cell population
(named TetR) was isolated in which almost all cells expressed the EGFP fusion
protein located in the nucleus (data not shown). Next, TetR cells were
transduced with lentivirus prepared from pLKO.DCMV.TetO.cIE1 and a mixed cell
population (named TetR-IE1) exhibiting both neomycin and puromycin resistance
was selected. Finally, fluorescence-activated cell sorting was performed to
collect cells with high levels of EGFPnlsTetR and, consequently, low levels of
IE1 in the absence of inductor.
To characterize the newly generated cells, TetR-IE1 cells were treated with
doxycycline for 24 or 72 h and examined for IE1 expression by indirect
immunofluorescence microscopy (Figure 1 A). Before induction, the majority (67.0%) of cells
was IE1 negative, and most other cells expressed barely detectably levels of the
viral protein. Interestingly, in the latter proportion of cells IE1 was present
in a predominantly punctate nuclear pattern. This likely reflects stable
co-localization between IE1 and ND10 due to viral protein levels insufficient to
disrupt the nuclear structures. At 24 h following induction only 2.8% of
cells were negative for IE1 expression and >97% stained positive for
the viral protein. In almost all positive cells IE1 exhibited a largely diffuse
nuclear staining indicating complete disruption of ND10. Very similar results
were obtained for IE1 expression and localization 72 h post induction.
Importantly, the observed temporal and spatial pattern of IE1 subnuclear
localization in TetR-IE1 cells closely resembles that observed during productive
hCMV infection in fibroblasts where initial colocalization between IE1 and ND10
is succeeded by ND10 disruption and diffuse nuclear distribution of the viral
protein [43],
[44], [71].
To compare the relative levels of IE1 expressed during hCMV infection and after
induction of TetR-IE1 cells, TetR cells were infected with the hCMV Towne
strain, and samples collected before or 3 h, 6 h, 12 h, 24 h, 48 h and 72 h
after infection were analyzed for IE1 steady-state protein levels in comparison
with samples of TetR-IE1 cells that had been treated with doxycycline (Figure 1 B). The timing of IE1
induction in TetR-IE1 cells was remarkably similar to the kinetics of IE1
accumulation in hCMV-infected cells. In addition, the IE1 levels detected at 24
to 72 h post induction were comparable to the protein amounts that had
accumulated by 24 h post hCMV infection.
To confirm that TetR-IE1 cells express fully active IE1, replication of wild-type
and IE1-deficient hCMV strains was compared by multi-step analyses conducted in
doxycycline-treated TetR and TetR-IE1 cells (Figure 1 C). To this end, we employed a
bacterial artificial chromosome (BAC)-based recombination approach to generate a
“markerless” mutant virus strain (TNdlIE1) lacking
the entire IE1-specific coding sequence. For details on the construction of
TNdlIE1 and a revertant virus (TNrvIE1)
see Materials and Methods. As expected, the
replication of two independent TNdlIE1 clones was strongly
attenuated in TetR cells, with a ∼2 to >3 log difference in titers
between mutant and revertant virus strains. It is important to note that our
previous work has shown that TNrvIE1 and the parental wild-type
strain (TNwt) exhibit identical replication kinetics [33]. However, induced TetR-IE1
cells were able to support wild-type-like replication of the
TNdlIE1 viruses demonstrating that the viral protein
provided in trans can fully compensate for the lack of IE1
expression from the hCMV genome during productive infection. Interestingly, even
the titers of TNrvIE1 were reproducibly up to ∼20-fold
higher in TetR-IE1 as compared to IE1-negative cells between 3 and 12 days post
infection.
Taken together, these results show that in TetR-IE1 cells expression of IE1 can
be synchronously induced from the autologous hCMV major IE (MIE) promoter
resulting in fully functional protein at levels present during the early stages
of hCMV infection. Thus, TetR/TetR-IE1 cells present an ideal model to study the
activities of the IE1 protein outside the complexity of infection, yet under
physiological conditions.
The capacity of hCMV IE1 to activate transcription from both viral and cellular
promoters has long been appreciated ([72]; reviewed in [2], [40], [41]).
However, most reports on IE1-regulated host gene transcription have relied on
transient transfections and promoter-reporter assays. To our knowledge,
regulation of endogenous cellular transcription by IE1 has so far only been
studied sporadically and at the level of single genes.
To comprehensively assess the impact of IE1 on the human transcriptome, we
performed a systematic gene expression analysis using our TetR/TetR-IE1 cell
model and Affymetrix GeneChip Human Gene 1.0 ST Arrays covering 28,869 genes
(>99% of sequences currently present in the RefSeq database, National
Center for Biotechnology Information). We compared the gene expression profiles
at 24 h and 72 h post induction in induced versus non-induced TetR-IE1 cells and
in induced TetR-IE1 versus induced TetR cells. Expression from the vast majority
(99.9%) of genes represented on the arrays was not significantly affected
by IE1. However, mRNA levels of 38 human genes differed by a factor of two or
more (p>0.01) in both the induced TetR-IE1/non-induced
TetR-IE1 and the induced TetR-IE1/induced TetR comparisons. For 32 (84%)
of the 38 genes, changes in mRNA levels were only observed after 72 h (but not
24 h) of IE1 expression, and only six (16%) were differentially expressed
at both 24 h and 72 h. Moreover, 13 (34%) of these genes were
down-regulated by a factor between 2.0 and 5.5 (data not shown) and 25
(66%) were up-regulated by a factor between 2.0 and 41.9 (Table 1). For the present
work, we concentrated on the set of genes that was found to be up-regulated by
expression of IE1.
We utilized the Gene Ontology (GO) classification system (http://www.geneontology.org) to identify attributes which
predominate among IE1-activated gene products regarding the three GO domains
“biological process”, “molecular function”, and
“cellular component”. Furthermore, we employed a set of analysis
tools to construct maps that visualize overrepresented attributes on the GO
hierarchy (Figure 2).
According to GO, the most significantly enriched “biological
process” terms with respect to the 25 IE1-activated genes are:
“immune system process”, “immune response”,
“inflammatory response”, “response to wounding”,
“response to stimulus”, “defense response”,
“chemotaxis”, “taxis”, and “regulation of cell
proliferation” (Figure 2
A). In fact, virtually all IE1-induced genes with assigned functions
have been implicated in adaptive or innate immune processes including
inflammation. Moreover, 7 (28%) of the 25 genes encode known cytokines or
other soluble mediators, namely the chemokine (C-X-C motif) ligands CXCL9,
CXCL10 and CXCL11, the chemokine (C-C motif) ligand CCL11, endothelin 1 (encoded
by EDN1), and the tumor necrosis factor (TNF) superfamily members 4 (TNFSF4,
also known as OX40 ligand) and 18 (TNFSF18, also known as GITR ligand). This
observation is also illustrated by the fact that, according to GO, the most
significantly enriched “molecular function” terms in the
IE1-activated transcriptome are: “cytokine receptor binding”,
“cytokine activity”, “chemokine activity”,
“chemokine receptor binding”, and “G-protein-coupled receptor
binding” (Figure 2 B).
Furthermore, the top “cellular component” category is
“extracellular space” (Figure 2 C). For a more thorough assessment of overrepresented GO
terms among IE1-induced genes, see Supporting Tables S1,
S2
and S3.
Surprisingly, the genes induced by IE1 are generally associated with stimulatory
rather than inhibitory effects on immune function including inflammation (Figure 2 A and Supporting
Table
S1). For example, some of the gene products are involved in the
proteolysis (cathepsin S encoded by CTSS), intracellular transport (TAP1
transporter) or cell surface presentation (HLA-DRA) of antigens (reviewed in
[73]). The
chemokines CXCL9, CXCL10, and CXCL11 mediate leukocyte migration (see Discussion; reviewed in [73], [74], [75]). CD274 (also known as
PDL1), TNFSF4, and TNFSF18 are co-stimulatory molecules which promote leukocyte
(including T and B lymphocyte) activation, proliferation and/or survival
(reviewed in [73], [76], [77], [78], [79]). Indoleamine 2,3-dioxygenase 1 (IDO1) and IRF1 have
also been linked to T lymphocyte regulation, but they have additional functions
in innate immune control of viral infection (reviewed in [73], [80], [81], [82], [83], [84], [85]. Likewise, GBP1 and murine
GBP2 exhibit antiviral activity [86], [87], [88], [89].
Out of the 25 IE1-activated genes, 14 were selected for validation by qRT-PCR.
The selected genes were representative of the entire range of expression
kinetics and induction magnitudes measured by microarray analysis. The PCR
approach confirmed expression of all tested genes typically reporting similar or
larger fold increases compared to the array data (Figure 3 A–B and Figure 4 A). For example, in induced (72 h)
versus non-induced TetR-IE1 cells the CXCL10 mRNA was found to be increased
24.6-fold by array analysis (Table 1) and 68.0-fold by PCR (Figure 3 A). Under the same conditions, the
GBP4 transcript was induced 13.5-fold by array analysis (Table 1) as compared to 19.1-fold by PCR
(Figure 3 A). The
corresponding data for TAP1 were 2.1-fold (array analysis; Table 1) and 2.3-fold (PCR;
Figure 3 A). Largely
concordant results regarding induction magnitudes between array and PCR analyses
were also obtained for CCDC3, CCL11, HES1, SERTAD4, TNFSF4, and TNFSF18 (Figure 3 B) as well as for
CXCL9, CXCL11, IDO1, IFIT2, and IRF1 (Figure 4 A). In addition to the extent of
gene activation, the precise timing of induction was exemplary investigated for
CXCL10, GBP4 and TAP1 (Figure 3
A). A substantial increase in mRNA production from all three genes
was evident at 72 h (and to a lesser extent at 48 h) but only minor effects were
detected between 6 h and 24 h post IE1 induction consistent with the array data
(Table 1).
Tubulin-β (TUBB) gene expression, which is not affected by IE1, served as a
negative control for the PCR experiments. Finally, the chemokines CXCL9 and
CXCL11 were exclusively detected in supernatants from TetR-IE1 but not TetR
cells (Figure 3 C).
Moreover, the levels of CXCL10 protein were drastically increased in TetR-IE1
compared to TetR cells. This demonstrates that for these genes elevated mRNA
levels also translate into enhanced protein synthesis and secretion.
The fact that increased expression of all tested IE1-activated genes was
detectable with two or three alternative approaches strongly suggests that
essentially all genes identified within the given experimental framework and
data analysis settings are truly differentially expressed upon induction of IE1.
Moreover, the activation of at least a subset of IE1-responsive genes appears to
be temporally coupled.
A plethora of past studies has established that immune regulatory genes are
preferential targets of IFN-based regulation [28], [29], [30]. Intriguingly, at least 21
(84%) of the 25 IE1-activated human genes identified by microarray
analysis turned out to be bona fide ISGs (Table 2) according to
informations retrieved from the Interferome database (http://www.interferome.org
[90]) and
other sources including our own qRT-PCR analyses (Figure 4 A and Supporting Table S4).
Several of these ISGs cluster in certain chromosomal locations (e.g., 1p22,
4q21, and 10q23-q25; Table
2) apparently reflective of their co-regulation.
An initial assessment mainly based on the Interferome data revealed that
IE1-activated ISGs are normally induced by either only IFN-γ or by both type
II and type I IFNs (Table
2). To confirm this assignment and to further discriminate between
type I and type II ISGs, we treated TetR and TetR-IE1 cells with exogenous
IFN-α or IFN-γ and analyzed the effects on mRNA accumulation from a
select subset of IE1-responsive ISGs. The transcript levels of all tested ISGs,
namely CXCL9–11, GBP4, IDO1, IFIT2, IRF1, and TAP1 (Figure 4 A) as well as CCL11 (Supporting
Table
S4) were not only increased by IE1 expression (TetR-IE1 relative to
TetR cells) but also by IFN-γ treatment of TetR cells, although to varying
degrees (∼2 to >30,000-fold; Figure 4 A). Notably, there was a significant positive correlation
(Pearson's correlation coefficient = 0.81) between the
magnitudes of IE1- and IFN-γ-mediated ISG induction. In contrast, the same
genes were substantially less susceptible (CXCL9–11, GBP4, IDO1, and
IFIT2) or entirely unresponsive (CCL11, IRF1, and TAP1) to IFN-α (Figure 4 A), and there was no
correlation (Pearson's correlation
coefficient = −0.04) between IE1 and IFN-α
responsiveness. For comparison, three typical type I ISGs, the genes encoding
eukaryotic translation initiation factor 2α kinase 2 (EIF2AK2, also known as
PKR), myxovirus (influenza virus) resistance 1 (Mx1, also known as MxA), and
2′,5′-oligoadenylate synthetase (OAS1), were strongly induced by
IFN-α but barely by IFN-γ or IE1 (Figure 4 B). Although no obvious synergistic
or additive effects between IE1 expression and IFN-γ treatment were observed
in these assays (Figure 4
A–B), IFN-α induction of type I ISGs was severely
compromised in TetR-IE1 as compared to TetR cells (Figure 4 B). The latter observation is
consistent with our previous work which has demonstrated that IE1 blocks
STAT2-dependent signaling resulting in inhibition of type I ISG activation [31], [33].
Hence, it appears that expression of IE1 selectively activates a subset of ISGs
and ISG gene clusters which are primarily responsive to IFN-γ indicating
that the viral protein elicits a type II IFN-like transcriptional response.
ISG activation typically requires synthesis, secretion and receptor binding of
IFNs (reviewed in [26], [27], [29], [30]). IFN-α is encoded by a multi-gene family and is
mainly expressed in leukocytes although some members are stimulated by IFN-β
in fibroblasts [91]. However, neither of 12 IFN-α (IFNA) and three
alternative type I IFN coding genes (IFNE, IFNK, and IFNW1 encoding IFN-ε,
IFN-κ, and IFN-ω, respectively) was noticeably induced by IE1 as judged
by our microarray results (Supporting Table S5). In contrast to IFN-α,
IFN-β is encoded by a single gene (IFNB) and is produced by most cell types,
especially by fibroblasts (IFN-β is also known as “fibroblast
IFN”). However, previous work has shown that IE1 expression does not
induce transcription from the IFN-β gene in fibroblasts [31], [32], [92].
Consistently, our microarray data did not reveal appreciable differences in
IFNB1 mRNA levels between TetR and TetR-IE1 cells (Supporting Table S5).
The single human IFN-γ gene (IFNG) is expressed upon stimulation of many
immune cell types but not usually in fibroblasts, and our microarray results
indicate that IE1 does not activate expression from this gene. Likewise, none of
the known type III IFN genes (IL28A, IL28B, and IL29 encoding IFN-λ2/IL-28A,
IFN-λ3/IL-28B, and IFN-λ1/IL-29, respectively) was significantly
responsive to IE1 expression in this system (Supporting Table S5).
For the IFN-β and IFN-γ transcripts, these results were confirmed by
highly sensitive qRT-PCR from doxycycline-treated TetR-IE1 and TetR cells.
Levels of the two IFN mRNAs did not significantly differ between TetR-IE1 and
TetR cells at any of ten post induction time points (0 h–96 h) under
investigation (Supporting Figure S1 and Supporting Table S6).
Thus, IE1 does not seem to induce expression from the IFN-γ or any other
human IFN gene.
To further rule out the possibility that ISG activation is a result of low level
IFN production or secretion of any other soluble mediator from IE1 expressing
cells, culture supernatants from TetR-IE1 cells induced with doxycycline for 24
h or 72 h were transferred to MRC-5 cells. As expected, MRC-5 cells did not
undergo ISG induction 3 h to 72 h following media transfer (data not shown).
Furthermore, we set up a transwell system with TetR cells in the top and
TetR-IE1 cells in the bottom chamber (Figure 5). Following addition of IFN-γ to
the lower chamber, we observed substantially increased mRNA levels of three
IE1-responsive indicator ISGs (CXCL9, CXCL11, and GBP4) in both TetR and
TetR-IE1 cells (Figure 5 A).
In contrast, addition of doxycycline caused up-regulation of ISG mRNA levels in
TetR-IE1 but not TetR cells (Figure
5 B). These results indicate that ISG induction is restricted to IE1
expressing cells and that a diffusible factor is not sufficient to mediate gene
activation by the viral protein.
Finally, we performed experiments adding neutralizing antibodies directed against
IFN-β and IFN-γ to the cell culture media (Figure 6). ISG-specific qRT-PCRs from TetR
cells treated with a combination of antibodies and high doses of the respective
exogenous IFN confirmed that cytokine neutralization was both highly effective
and specific. At the same time, neither the IFN-β- nor the
IFN-γ-specific neutralizing antibodies had any significant negative effect
on IE1-mediated ISG induction. These results strongly support the view that ISG
activation by IE1 is independent of IFN-β, IFN-γ, and likely other
IFNs.
Homodimeric STAT1 complexes are the central intracellular mediators of canonical
IFN-γ signaling (reviewed in [26], [27], [28], [29], [30]). Interestingly, previous work
has shown that the IE1 protein interacts with both STAT1 and STAT2, although
STAT2 binding appeared to be more efficient [31], [32], [33], [39]. STAT2 has also been
implicated in certain IFN-γ responses ([93], [94]; reviewed in [95]), although
some (hCMV-mediated) activation of ISG transcription appears to occur entirely
independent of STAT proteins ([96]; reviewed in [26], [27]).
To investigate whether ISG activation by IE1 requires the presence of STAT1
and/or STAT2, we employed siRNA-based gene silencing individually targeting the
two STAT transcripts. Following transfection of MRC-5, TetR and/or TetR-IE1
cells with two different siRNA duplexes each for STAT1 and STAT2, we monitored
endogenous STAT expression by immunoblotting (Figure 7 A) and qRT-PCR (Figure 7 B). An estimated ≥80%
selective reduction in STAT1 and STAT2 protein accumulation was observed 2 days
following siRNA transfection, and even after 5 days significantly lower protein
levels were detected compared to cells transfected with a non-specific control
siRNA (Figure 7 A). The
knock-down of STAT1 and STAT2 was also evident at the level of mRNA accumulation
(86 to 95% for STAT1 and 51 to 95% for STAT2 at day 5 post
transfection; Figure 7 B).
The knock-down specificity was verified by confirming that STAT1 siRNAs do not
significantly reduce STAT2 mRNA levels and vice versa.
Moreover, none of the STAT-directed siRNAs had any appreciable effect on IE1
expression (Figure 7 B).
Again, expression from the CXCL10 and GBP4 genes was strongly up-regulated in
doxycycline-treated TetR-IE1 versus TetR cells. However, STAT1 knock-down caused
the CXCL10 and GBP4 genes to become almost entirely resistant to IE1-mediated
activation in induced TetR-IE1 cells. In contrast, depletion of STAT2 had no
negative effect on IE1-dependent ISG induction (Figure 7 B) although it diminished basal and
IFN-α-induced type I ISG (OAS1) expression (Supporting Figure S2).
These results demonstrate that STAT1, but not STAT2, is an essential mediator of
the cellular transcriptional response to IE1 expression and suggest that the
viral protein might mediate ISG activation via activation of JAK-STAT
signaling.
The activation-inactivation cycle of STAT transcription factors entails their
transition between different dimer conformations. Unphosphorylated STATs can
dimerize in an anti-parallel conformation, whereas tyrosine (Y701)
phosphorylation triggers transition to a parallel dimer conformation resulting
in increased DNA binding and nuclear retention of STAT1 (reviewed in [29], [30], [97]). In
addition, serine (S727) phosphorylation is required for the full transcriptional
and biological activity of STAT1 [98]. In order to investigate
whether IE1 promotes STAT1 activation, we compared the levels of Y701- and
S727-phosphorylated STAT1 in doxycyline-induced TetR and TetR-IE1 cells (Figure 8 A). Total STAT1
steady-state protein levels were very similar in TetR and TetR-IE1 cells. In
contrast, Y701-phosphorylated forms of STAT1 were only detectable in the
presence of IE1 unless cells were treated with IFN-γ. In addition, IE1 was
almost as efficient as IFN-γ in inducing STAT1 S727 phosphorylation. These
results strongly suggest that IE1 expression triggers the formation of Y701- and
S727-phosphorylated, transcriptionally fully active STAT1 dimers.
To examine whether STAT1 Y701 and/or S727 phosphorylation is an essential step in
IE1-mediated ISG activation, we set up a “knock-down/knock-in”
system designed to study mutant STAT1 proteins in a context of diminished
endogenous wild-type protein levels. We constructed an
“siRNA-resistant” STAT1 coding sequence, termed STAT1*,
containing two silent nucleotide exchanges in the sequence corresponding to
siRNA STAT1 #146 (Figure 7
A). The STAT1* sequence was used as a substrate for further
mutagenesis to generate siRNA-resistant constructs encoding mutant STAT1
proteins with conservative amino acid substitutions that preclude tyrosine or
serine phosphorylation (Y701F or S727A, respectively; reviewed in [99], [100]). A
retroviral gene transfer system based on vector pLHCX was utilized to
efficiently express the different STAT1 proteins in TetR-IE1 cells. All STAT1
variants (STAT1*, STAT1*Y701F, and STAT1*S727A) were overexpressed
to levels undiscernible from the wild-type protein and mRNA (Figure 8 B–C). In
comparison to transfections with a non-specific control siRNA (#149), siRNA #146
severely reduced the levels of endogenous and overexpressed wild-type STAT1
without negatively affecting expression of the siRNA-resistant STAT1 variants or
IE1 (Figure 8 B–C). As
expected, the Y701F and S727A mutant STAT1 proteins did not undergo tyrosine or
serine phosphorylation, respectively, upon stimulation by IFN-γ.
Interestingly, while the S727A protein could still be tyrosine-phosphorylated,
the Y701F mutant was defective for both tyrosine and serine phosphorylation
(Figure 8 B). This
observation is in agreement with previous findings showing that
IFN-γ-dependent S727 phosphorylation occurs exclusively on
Y701-phosphorylated STAT1 [101]. Ectopic expression of wild-type STAT1, STAT1*,
and STAT1*S727A but not STAT1*Y701F in addition to the endogenous
protein enhanced IE1-mediated activation of CXCL10 and GBP4 transcription.
Conversely, siRNA-mediated depletion of endogenous STAT1 strongly reduced this
response. Importantly, expression of STAT1* in cells depleted of endogenous
STAT1 rescued ISG induction by IE1 almost completely. STAT1*S727A expression
also compensated for the lack of endogenous STAT1, although slightly less
efficiently compared to STAT1*, whereas STAT1*Y701F was unable to rescue
IE1-mediated ISG activation (Figure
8 C).
Thus, although IE1 appears to trigger phosphorylation of STAT1 at both Y701 and
S727, only the former modification is required for ISG activation. Nonetheless,
STAT1 S727 phosphorylation may augment IE1-dependent gene activation.
Y701 phosphorylation usually causes a cytoplasmic to nuclear shift in
steady-state localization and efficient sequence-specific DNA binding of STAT1
dimers (reviewed in [29], [30], [97]). Accordingly, immunofluorescence microscopy revealed
that the presence of IE1 strongly promotes nuclear accumulation of STAT1, very
similar to what was observed following addition of IFN-γ (Figure 9 A). In contrast,
significant amounts of nuclear STAT2 were only detected after treatment of cells
with IFN-α but not upon IE1 expression. These results were confirmed by
nucleo-cytoplasmic cell fractionation (Figure 9 B). In these assays, IE1 induction
for 72 h was as efficient in promoting STAT1 nuclear accumulation as treatment
with type I or type II IFNs for 1 h. IFN treatment also strongly induced the
nuclear accumulation of STAT2. However, the levels of nuclear STAT2 increased
only marginally upon expression of IE1.
Finally, we asked whether IE1 may direct STAT1 to promoters of type II ISGs.
Chromatin immunoprecipitation (ChIP) analyses demonstrated that the viral
protein potentiates the recruitment of STAT1 to certain IFN-γ- and
IE1-responsive ISG promoters (e.g., TAP1) but not to promoters of several
non-ISGs (e.g., GAPDH; Figure 10
A). Moreover, there was a positive correlation between the magnitude
of STAT1 chromatin association induced by IE1 and IFN-γ. At the same time,
IE1 had no effect on association of STAT2 with these promoters (Figure 10 B). These results
are in agreement with the fact that a previous global ChIP-sequencing study has
experimentally demonstrated STAT1 association with 14 (56%) out of the 25
IE1-responsive gene promoters identified in this study ([102] and Supporting Table S7).
In addition, 22 (88%) of these promoter sequences (all except EDN1, HBG1,
and HLA-DRA) carry one or more (up to six) predicted STAT1β binding sites
(GAS elements) according to the PROMO tool (version 3.0.2, default settings with
15% maximum matrix dissimilarity rate, http://alggen.lsi.upc.es),
which predicts transcription factor binding sites as defined by position weight
matrices derived from the TRANSFAC (version 8.3) database [103], [104]. Similar results were
obtained with other in silico promoter analysis tools (data not
shown).
Based on these findings we propose that IE1 activates a subset of ISGs at least
in part through increasing the nuclear concentration and sequence-specific DNA
binding of phosphorylated STAT1 thereby modulating host gene expression in an
unanticipated fashion.
The transcriptional transactivation capacity of the hCMV MIE proteins has been
recognized for decades ([72]; reviewed in [2], [40], [41]). For example, it has long
been established that the 72-kDa IE1 protein can stimulate transcription from its
own promoter-enhancer [36], [105], [106]. IE1 also activates at least a subset of hCMV early
promoters therein collaborating with the viral 86-kDa IE2 protein [34], [35], [53], [71], [72], [107], [108], [109]. Furthermore,
IE1 or combinations of IE1 and IE2 can stimulate expression from a variety of
non-hCMV promoters. In fact, numerous heterologous viral and cellular promoters are
responsive to IE1 or combinations of IE1 and IE2 [50], [51], [52], [57], [60], [61], [71], [72], [110], [111], [112], [113], [114], [115], [116], [117]. IE1 may accomplish
transcriptional activation via interactions with a diverse set of cellular
transcription regulatory proteins thereby acting through multiple DNA elements [50], [51], [52], [54], [55], [56], [57], [58], [59], [105], [106], [109], [110], [111], [112], [113], [117], [118], [119], [120], [121], [122], [123], [124], [125], [126] as well as
epigenetic mechanisms including histone acetylation [53], [59], [127]. More recently, IE1 has also
been implicated in transcriptional repression [31], [32], [33], [57], [62], [63], [64]. Our own work ([31] and this study,
Figure 4 B) and a report by
Huh et al. (2008) has demonstrated that IE1 can inhibit the hCMV-
or IFN-α/β-dependent activation of human ISGs including ISG54, MxA, PKR, and
CXCL10. The mechanism of inhibition appears to involve physical interactions of IE1
with the cellular STAT1 and STAT2 proteins that result in diminished DNA binding of
the ternary ISGF3 complex to promoters of type I ISGs ultimately interfering with
transcriptional activation [31], [32], [33]. Despite this plethora of studies, our understanding of
the true transcriptional regulatory capacity of IE1 is still limited. This is mainly
due to the fact that IE1-regulated transcription has almost exclusively been studied
at the single gene level. Moreover, much of the past work has relied on
transfection-based promoter-reporter assays, and IE1-dependent up- or
down-regulation of only very few endogenous human genes has been demonstrated so
far.
The present work constitutes the first systematic analysis of IE1-specific changes to
transcription from the human genome. Importantly, to minimize cellular compensatory
effects and to closely mimic the situation during hCMV infection, all experiments
were based on short-term (up to 72 h) induction of IE1 expression from its
autologous promoter (Figure 1
A–B). Just over 0.1% (25 out of 28,869) of all human
transcripts under examination were found to be significantly up-regulated by IE1
under stringent analysis conditions (Table 1). This figure may be unexpected in the light of the reported
interactions of IE1 with several ubiquitous transcription factors and its reputation
as a “promiscuous” transactivator. However, rather than causing a broad
transcriptional host response, IE1-specific gene activation was largely restricted
to a subset of ISGs that are primarily responsive to IFN-γ (Table 2, Figure 4 and Supporting Table S4).
Thus, IE1 appears to activate certain ISGs (typically type II ISGs) while
simultaneously inhibiting the activation of other ISGs (typically type I ISGs).
Importantly, more than half (at least 14 out of the 25) IE1-activated genes
identified in this study were previously shown to be induced during hCMV infection
of fibroblasts and/or other human cell types (Table 3). This strongly suggests that many if not
all IE1-specific transcriptional changes observed in our expression model may be
relevant to viral infection. On the other hand, our preliminary results indicate
that the conditional replication defect of IE1 knock-out viruses in human
fibroblasts [35],
[36] may not
result from an inability to initiate an IFN-γ-like response (data not shown). In
fact, additional viral gene products are known or expected to contribute to ISG
activation during hCMV infection (reviewed in [26], [27]) and may compensate for IE1
in this respect, at least during productive infection of fibroblasts.
In addition to being distinctively responsive to IFN-γ, most IE1-activated genes
appear to share similar kinetics of induction (Table 1 and Figure 3), and many cluster in certain genomic
locations (Table 2) suggesting
a common underlying mechanism of activation. Specific siRNA-mediated STAT1 (but not
STAT2) knock-down inhibited IE1-dependent activation of several target ISGs almost
completely (Figure 7 A).
Conversely, STAT1 overexpression proved to enhance ISG activation in IE1 expressing
cells (Figure 8 C). Moreover,
defective IE1-activated ISG transcription in cells depleted of endogenous STAT1 was
efficiently rescued by ectopic STAT1 expression (Figure 8 C). These results demonstrate that the
STAT1 protein is a critical mediator of the cellular transcriptional response to
IE1. Moreover, this response appears to strictly depend on the Y701-phosphorylated
form of STAT1 which is induced by IE1 expression (Figure 8). Although recent work has shown that
some STAT1 functions are executed by the non-phosphorylated protein (reviewed in
[97], [99], [100]), it is the
Y701-phosphorylated form that preferentially accumulates in the nucleus and binds to
DNA with high affinity (reviewed in [29], [30]) providing a mechanism for
IE1-dependent ISG activation. IE1 also induces S727 phosphorylation of STAT1 (Figure 8 A), but this modification
is dispensable merely serving an augmenting function in ISG activation triggered by
the viral protein (Figure 8 C).
Phosphorylation of S727 is thought to be required for the full transcriptional
activity of STAT1 by recruiting histone acetyltransferase activity [98], [128], [129].
Interestingly, the hCMV IE1 protein can promote histone acetylation [53] suggesting it
might compensate for S727 phosphorylation by binding to DNA-associated STAT1.
Our prior work has shown that IE1 physically interacts with STAT1 during hCMV
infection and in vitro, and the two proteins co-localize in the
nuclei of transfected cells treated with IFN-α [31]. The results of Figure 9 extend these observations
by demonstrating that the viral protein facilitates nuclear accumulation and DNA
binding of STAT1 in the absence of IFNs. The STATs were initially described as
cytoplasmic proteins that enter the nucleus only in the presence of cytokines.
However, it has now been established that STATs constantly shuttle between nucleus
and cytoplasm irrespective of cytokine stimulation (reviewed in [97], [130], [131]). Thus, complex
formation between nuclear resident IE1 and STAT1 passing through the nucleus may be
sufficient to impair STAT1 export to the cytoplasm resulting in nuclear retention
and increased DNA binding of the cellular protein. In this scenario, IE1 may
increase the levels of Y701-phosphorylated STAT1 by interfering with nuclear
dephosphorylation of the cellular protein. In fact, DNA binding was shown to protect
STAT1 from dephosphorylation, which normally occurs at a step preceding export to
the cytoplasm [132], [133]. This one-step “nuclear shortcut” model
assumes that small amounts of Y701-phosphorylated STAT1 enter the nucleus in the
absence of IFNs and any potential IE1-induced mediators of STAT1 activation.
Conceivably, human fibroblasts (TetR cells) may constitutively release small amounts
of soluble inducers (e.g., certain growth factors; see below) that maintain residual
levels of activated STAT1 undetectable by immunoblotting (Figure 8 A). Moreover, we cannot rule out that
the fetal calf serum used for cell culture media may contain factors causing a
limited number of STAT1 molecules to undergo Y701 phosphorylation. In contrast,
increased S727 phosphorylation in the presence of IE1 may result from higher levels
of DNA-targeted STAT1, as this modification is preferentially or exclusively
incorporated into the nuclear chromatin-associated cellular protein, at least during
the normal IFN-γ response [101].
Alternatively, IE1 may actively induce STAT1 Y701 phosphorylation thereby promoting
nuclear import of STAT1 dimers. This phosphorylation event is typically mediated by
cytoplasmic JAK family kinases upon ligand-mediated activation of IFN receptors.
However, our results demonstrate that IE1 does not induce the expression of human
IFN genes, and we found no evidence for IFN-γ or IFN-β secretion from IE1
expressing cells (Supporting Table S5, Figure 6 and data not shown). Moreover, our
transwell and media transfer experiments indicate that cytokines or other soluble
mediators that may constitute a hypothetical IE1 “secretome” are not
sufficient to stimulate ISG expression (Figure 5 and data not shown). However, this does not rule out the
possibility that IE1 may cooperate with one or more soluble factors to trigger the
observed transcriptional response. In fact, 80% of all IE1 target genes were
not found activated within the first 24 h after induction of IE1 expression despite
the fact that the viral protein had reached almost peak levels by this time (Figure 1 B and Table 1). Instead, up-regulation
typically started at 48 h and increased until at least 72 h following IE1 expression
(Table 1 and Figure 3 A). This timing of
induction is compatible with a two-step model in which IE1 first initiates
de novo synthesis and secretion of an unidentified cellular
gene product required to trigger STAT1 Y701 phosphorylation (step 1). Besides IFNs,
STAT1 signaling can be induced by several interleukins (e.g., IL-6) some of which
are known to be up-regulated by IE1 [58], [60], [61], [110]. However, STAT1 Y701
phosphorylation can also occur independently of cytokines (reviewed in [134]). In
fact, growth factors including the epidermal growth factor and certain hormones are
also able to induce STAT1 Y701 phosphorylation [135], [136], [137], [138], [139]. In addition, tumor necrosis
factor (TNF) has been shown to signal through activated STAT1 [140] raising the intriguing
possibility that the soluble protein products of TNFSF4 and/or TNFSF18, two TNF
family members belonging to the few genes already activated by 24 h following IE1
induction (Table 1), may be
involved in IE1-mediated Y701 phosphorylation of STAT1. However, activation of one
or more of these IFN-independent pathways may not produce enough activated nuclear
STAT1 to trigger efficient ISG expression and may therefore be required but not
sufficient for IE1-mediated gene induction. In accordance with this possibility, the
levels of Y701-phosphorylated STAT1 were much higher in IFN-γ-treated as
compared to IE1 expressing cells (Figure 8 A). Thus, on top of low level Y701 phosphorylation,
IE1-dependent nuclear retention of STAT1 through complex formation between the viral
and cellular protein (as outlined for the one-step model; see above) may be
necessary in order to elicit a significant transcriptional response (step 2).
Although activated STAT1 is clearly a key mediator of IE1-dependent ISG induction,
additional factors may be involved. In fact, not all known STAT1-activated human
genes seem to be included in the IE1-specific transcriptome implying that additional
gene products likely contribute to target specificity. One of the candidate
co-factors that has been repeatedly linked to IE1 function is NFκB. In fact, IE1
was shown to activate the NFκB p65 (RelA) and RelB promoters [55], [112], [113], [121], to facilitate
expression of the NFκB RelB subunit and/or NFκB post-translational
activation [58],
[113], [119], [121], and to activate
transcription through NFκB binding sites [58], [105], [106], [113], [119], [126]. At the same time, NFκB has
been implicated in IFN-γ-induced activation of a subset of ISGs including CXCL10
and GBP2 ([141],
[142], [143], [144], [145]; reviewed in
[146], [147]). However, we
did not observe nuclear translocation of NFκB following induction of IE1 in
TetR-IE1 cells. Moreover, siRNA-mediated knock-down of NFκB p65 had no
significant impact on IE1-activated CXCL10 and GBP4 expression in these cells (data
not shown). These observations indicate that the transcriptional response to IE1 is
largely independent of NFκB, at least within our experimental setup. IRF1 is
another transcription factor that contributes to the activation of certain ISGs
including CTSS, GBP2, and TAP1 ([128], [148], [149], [150]; reviewed in
[80], [81], [82]). IRF1 might
enhance IE1-mediated ISG activation, especially since its mRNA is up-regulated by
expression of the viral protein (Table 1 and Figure 4
A).
A key feature of the IE1 protein appears to be its ability to target to and disrupt
subnuclear multi-protein structures known as PML bodies or ND10 during the early
phase of hCMV infection and upon ectopic expression [42], [43], [44]. The mechanism of IE1-dependent
ND10 disruption most likely involves binding to the PML protein, a major constituent
of ND10 [54]. We have
not specifically investigated the role of PML in IE1-mediated gene induction.
Nonetheless, our results are compatible with the possibility that ND10 disruption is
required for the transcriptional response to IE1 since the nuclear structures were
confirmed to be disintegrated at both post-induction time points (24 h and 72 h) of
our microarray analysis (data not shown). Although the exact function of ND10
remains unclear, the structures have been implicated in a variety of processes
including inflammation [151] and anti-viral defense (reviewed in [45], [46], [47], [48]). Besides a
proposed role of ND10 in viral gene expression, they may also function in
transcriptional regulation of certain cellular genes. Several examples of selective
associations between ND10 and genes or chromosomal loci, especially regions of high
transcription activity and/or gene density, have been reported (reviewed in [152]). For example,
immunofluorescent in situ hybridization analyses demonstrated that
the major histocompatibility (MHC) class I gene cluster on chromosome 6 (6p21) is
non-randomly associated with ND10 in human fibroblasts [153]. Transcriptional activation in
the presence of IFN-γ correlates with the relocalization of this locus to the
exterior of the chromosome 6 territory in a process that appears to involve DNA
binding of Y701-phosphorylated STAT1, changes in chromatin loop architecture, and
histone hyperacetylation [154], [155], [156]. Interestingly, many IE1-activated genes cluster in
certain genomic locations (Table
2). This includes the HLA-DRA and TAP1 genes located within the
ND10-associated MHC locus at 6p21. Together these observations raise the intriguing
possibility that, through a combination of PML disruption and STAT1 activation, IE1
might cause higher order chromatin remodeling of entire chromosomal loci resulting
in transcriptional activation.
One of the most surprising findings of the present study concerns the fact that most
IE1-induced cellular genes are generally associated with stimulatory rather than
inhibitory effects on immune function and inflammation (Table 1, Figure 2 and Supporting Tables S1,
S2). It
has been proposed that certain inflammatory and innate defense mechanisms launched
by the host to limit hCMV replication may actually facilitate viral dissemination,
for example by increasing target cell availability and/or by creating an environment
conducive to virus reactivation (coined “no pain, no gain” by Mocarski
[157]).
Thus, it is plausible that hCMV not just attenuates host immunity through the
numerous immune evasion mechanisms ascribed to this virus (reviewed in [158]), but rather
aims at counterbalancing the effects of the innate and inflammatory response in
restricting and facilitating viral replication. This strategy may be crucial in
allowing for what has been termed “mutually assured survival” of both
virus and host [159].
The functional group of IE1-induced pro-inflammatory proteins potentially involved in
viral target cell recruitment is best represented by the chemokines CXCL9, CXCL10,
and CXCL11. All three proteins are not only induced by IE1 (Table 1 and Figures 3–7) but also during hCMV infection of various cell
types, and they represent major constituents of the viral secretome ([4], [18], [24], [160], [161], [162], [163], [164], [165] and Table 3). By binding to a common
receptor, termed CXCR3, the three chemokines have the ability to attract subsets of
circulating leukocytes to sites of infection and/or inflammation (reviewed in [74], [75]). Although
CXCR3 is preferentially expressed on activated T helper 1 cells, the receptor
protein is also present on many other cell types including CD34+ hematopoietic
progenitors [166]
which are preferential sites of hCMV latency [167], [168], [169], [170], [171], [172]. CXCR3 and its ligands have
been implicated in a large variety of inflammatory and immune disorders (reviewed in
[74], [75]). For example,
cells expressing CXCR3 are found at high numbers in biopsies taken from patients
experiencing organ transplant dysfunction and/or rejection [173], [174], [175], [176], [177], [178], [179], [180], [181]. Moreover, CXCL9 [175], [176], [177], [179], [180], CXCL10 [173], [174], [175], [176], [177], [179], [180], and CXCL11 [175], [176], [177], [178], [179], [180], [181] mRNA and
protein levels are increased in tissues of organs undergoing rejection. Importantly,
the levels of CXCR3-positive cells and CXCR3 ligand mRNA in the biopsy samples
frequently correlate with the grade of graft rejection [174], [176], [177], [178], [180] suggesting a causative role of
this pathway. Up-regulation of CXCL10 and other chemokines also correlated with
transplant vascular sclerosis and chronic rejection in an rCMV cardiac allograft
infection model [4], [182], [183]. In addition to CXCL9, CXCL10, and CXCL11, IE1 also
up-regulates expression of CCL11 (Table 1), another CXCR3-interacting chemokine [184]. Through activation of the
CXCR3 axis, IE1 might contribute to hCMV dissemination and pathogenesis in
unexpected ways.
The IE1 protein has long been suspected to be a key player in the events leading to
reactivation from hCMV latency although this view has recently been challenged by
functional analysis of the mCMV and rCMV IE1 orthologs in mouse and rat models of
infection, respectively [37], [185]. Nonetheless, inflammatory (including allogeneic) immune
responses are believed to be efficient stimuli for hCMV reactivation. In fact,
stimulation of latently infected monocytes or myeloid progenitor cells with
pro-inflammatory cytokines including IFN-γ can reactivate viral replication
([186], [187], [188],
[189]; reviewed
in [190], [191], [192]). IFN-γ
may aid hCMV reactivation by affecting cellular differentiation ([193]; reviewed in
[28], [190], [191], [192]) and/or by
activating transcription through GAS-like elements present in the viral MIE
promoter-enhancer [194]. These GAS-like elements were shown to be required for
efficient hCMV transcription and replication, at least after low multiplicity
infection, and IFNs enhanced MIE gene expression [194]. Conceivably, the IE1
protein may phenocopy the effect of IFN-γ in activating both cellular ISGs and
the viral MIE promoter thereby facilitating viral reactivation. Conversely, along
the lines of the “immune sensing hypothesis of latency control” proposed
by Reddehase and colleagues [195], episodes of IE1 expression may promote maintenance of
viral latency not only through providing antigenic peptides (reviewed in [196]) but also
by concomitantly activating critical immune effector functions including antigen
transport (TAP1), processing (CTSS) and presentation (HLA-DRA) as well as immune
cell recruitment (CXCL9, CXCL10, CXCL11, CCL11; see above) and co-stimulation
(TNFSF4, TNFSF18 and CD274).
Current anti-hCMV strategies are directed against viral DNA replication, but
sometimes fail to halt disease. This may be due to virus-induced “side
effects” that are not correlated to production of virus particles and lysis of
host cells. In fact, in hCMV pneumonitis and retinitis, disease symptoms were
repeatedly found in the absence of replicating virus or viral cytopathogenicity
[197], [198]. Similarly, in
mouse models of viral pneumonitis mCMV replication per se was not
sufficient to cause disease [197], [199], [200]. Conversely, mCMV disease could be triggered
immunologically without inducing viral replication [201]. Here we have shown that out
of >160 different hCMV gene products, a single protein (IE1) is sufficient to
alter the expression of human genes with strong pro-inflammatory and immune
stimulatory potential without the requirement for virus replication. The present
work supports the idea that the hCMV MIE gene and specifically the IE1 protein may
play a direct and predominant role in viral immunopathogenesis and inflammatory
disease [202],
[203], [204], [205]. Thus, the IE1
protein should be considered a prime target for the development of improved
prevention and treatment options directed against hCMV.
The pMD2.G and psPAX2 packaging vectors for recombinant lentivirus production
were obtained from Addgene (http://www.addgene.org;
plasmids 12259 and 12260, respectively). Plasmids pLKOneo.CMV.EGFPnlsTetR,
pLKO.DCMV.TetO.cICP0, and pCMV.TetO.cICP0 were kindly provided by Roger Everett
(Glasgow, UK). pLKOneo.CMV.EGFPnlsTetR contains the complete hCMV MIE promoter
upstream of a sequence encoding EGFP fused to an NLS and TetR [68], [69], [70]. In the
pLKO.1puro derivative pLKO.DCMV.TetO.cICP0, expression of the herpes simplex
virus type 1 infected cell protein 0 cDNA (cICP0) is under the control of a
tandem TetO sequence located downstream of a truncated version of the hCMV MIE
promoter (DCMV) [69], [70]. To generate pLKO.DCMV.TetO.cIE1, the IE1 cDNA of the
hCMV Towne strain was PCR-amplified from pEGFP-IE1 [71] with upstream primer #483
containing a HindIII site and downstream primer #484 containing
an EcoRI site (the sequences of all primers used in this study
are listed in Supporting Table S8). The IE1 sequence was subcloned
into the HindIII and EcoRI sites of
pCMV.TetO.cICP0. The NdeI-EcoRI fragment of
the resulting plasmid pCMV.TetO.IE1 was verified by sequencing and used to
replace the ICP0 cDNA in pLKO.DCMV.TetO.cICP0 thereby generating plasmid
pLKO.DCMV.TetO.cIE1.
QuikChange site-directed mutagenesis of plasmid pRc/CMV-hSTAT1p91 (kindly
provided by Christian Schindler, New York, USA) with oligonucleotides #660 and
#661 resulted in pCMV-STAT1* encoding a STAT1 variant mRNA resistant to
silencing by the STAT1-specific siRNA duplex #146 (the sequences of all siRNAs
used in this study are listed in Supporting Table S9).
The plasmids pCMV-STAT1*Y701F and pCMV-STAT1*S727A were generated by
QuikChange mutagenesis of pCMV-STAT1* with primer pairs #662/#663 and
#664/#665, respectively. BamHI-EcoRV fragments
of pRc/CMV-hSTAT1p91, pCMV-STAT1*, pCMV-STAT1*Y701F, and
pCMV-STAT1*S727A were treated with Klenow fragment and ligated to the
HpaI-digested, dephosphorylated retroviral vector pLHCX
(Clontech, no. 631511) resulting in plasmids pLHCX-STAT1, pLHCX-STAT1*,
pLHCX-STAT1*Y701F, and pLHCX-STAT1*S727A, respectively. The correct
orientations and nucleotide sequences of the inserted STAT1 cDNAs were verified
by sequencing.
Human MRC-5 embryonic lung fibroblasts (Sigma-Aldrich, no. 05011802), the human
p53-negative non-small cell lung carcinoma cell line H1299 (ATCC, no. CRL-5803
[206]),
and Phoenix-Ampho retrovirus packaging cells (from Garry Nolan, Stanford, USA
[207]) were
maintained in Dulbecco's Modified Eagle's Medium supplemented with
10% fetal calf serum, 100 units/ml penicillin, and 100 µg/ml
streptomycin. All cultures were regularly screened for mycoplasma contamination
using the PCR Mycoplasma Test Kit II from PromoKine. Where applicable, cells
were treated with 1,000 U/ml recombinant human IFN-α A/D (R&D Systems,
no. 11200), 10 ng/ml recombinant human IFN-β 1a (Biomol, no. 86421), or 10
ng/ml recombinant human IFN-γ (R&D Systems, no. 285-IF) for various
durations. Neutralizing goat antibodies to human IFN-β (no. AF814) or
IFN-γ (no. AF-285-NA) and normal goat IgG (no. AB-108-C) were purchased from
R&D Systems and used at concentrations of 1 µg/ml (anti-IFN-β) or
2 µg/ml (anti-IFN-γ, normal IgG). Transwell assays were performed in
tissue-culture-treated 100-mm plates with polycarbonate membrane and 0.4
µm pore size (Corning, no. 3419).
During the week prior to transfection, Phoenix-Ampho cells were grown in medium
containing hygromycin (300 µg/ml) and diphtheria toxin (1 µg/ml).
Production of replication-deficient retroviral particles, retrovirus infections,
and selection of stable cell lines were performed according to the pLKO.1
protocol available on the Addgene website (http://www.addgene.org/pgvec1?f=c&cmd=showcol&colid=170&page=2)
with minor modifications. Retroviral particles were generated by transient
transfection of H1299 cells (pLKO-based vectors) or Phoenix-Ampho cells
(pLHCX-based vectors) using the calcium phosphate co-precipitation technique
[208].
Recombinant viruses were collected 36 h and 60 h after transfection, and were
used for transduction of target cells by two subsequent 16 h incubations. To
generate TetR cells, MRC-5 fibroblasts at population doubling 19 were infected
with pLKOneo.CMV.EGFPnlsTetR-derived lentiviruses and selected with G418 (0.2
mg/ml). To generate TetR-IE1 cells, TetR cells were transduced by
pLKO.DCMV.TetO.cIE1-derived lentiviruses and selected with puromycin (1
µg/ml). Cells with high level EGFPnlsTetR expression (and low IE1
background) were enriched by fluorescence-activated cell sorting in a FACSCanto
II flow cytometer (BD Biosciences). TetR cells were maintained in medium
containing G418 (0.1 mg/ml), while TetR-IE1 cells were cultured in the presence
of both G418 (0.1 mg/ml) and puromycin (0.5 µg/ml). To induce IE1
expression, cells were treated with doxycycline (Clontech, no. 631311) at a
final concentration of 1 µg/ml. To generate TetR-IE1 cells with stable
expression of ectopic STAT1 proteins, uninduced TetR-IE1 cells were transduced
with pLHCX-derived retroviruses encoding STAT1, STAT1*, STAT1*Y701F, or
STAT1*S727A.
The EGFP-expressing wild-type Towne strain (TNwt) of hCMV was derived from an
infectious BAC clone (T-BACwt [209]) of the viral genome. Allelic exchange to generate
IE1-deficient viruses (TNdlIE1) and corresponding
“revertants” (TNrvIE1) utilized the following
derivatives of transfer plasmid pGS284 [210]:
pGS284-TNIE1kanlacZ, pGS284-TNMIEdlIE1,
pGS248-TNMIE, and pGS284-TNMIErvIE1. Plasmid
pGS284-TNIE1kanlacZ contains the kanamycin resistance gene
(kan) and the lacZ gene cloned between
sequences flanking the IE1-specific exon four of the hCMV TN MIE transcription
unit. The ∼1000-bp flanking sequences were obtained by PCR amplification
using primers #136 and #137 (downstream flanking sequence) or #139 and #140
(upstream flanking sequence; for PCR primer sequences, see Supporting Table S8)
and T-BACwt as template. The amplified downstream flanking sequence was cloned
into pGS284 via BglII and NotI sites present
in both the PCR primers and target vector sequences. Following addition of
adenosine nucleotide overhangs to the 3′-ends of the PCR product, the
upstream flanking sequence was first subcloned into vector pCR4-TOPO
(Invitrogen) and subsequently inserted via NotI sites into
pGS284 carrying the downstream flanking sequence. The kanlacZ
expression cassette was released from plasmid YD-C54 [211] and cloned into the
PacI sites (introduced through PCR primers #137 and #139)
located between the hCMV flanking sequences in the pGS284 derivative described
above. Plasmid pGS284-TNMIEdlIE1 contains an MIE fragment
lacking 1,413 bp between the AccI sites upstream and downstream
of exon four. The exon four-deleted MIE fragment was obtained from T-BACwt by
overlap extension PCR as previously described [212]. The primer pairs used
for PCR mutagenesis were #348/#349 (upstream fragment), #350/#351 (downstream
fragment), and #348/#351 (complete fragment). The final PCR product was cloned
via BglII and NotI sites into pGS284. For the
construction of pGS248-TNMIE (previously termed pGS248-MIE; [33]), a
∼3000-bp sequence of the MIE region was amplified by PCR using template
T-BACwt and primers #155 and #156. After phosphorylation, the PCR product was
first inserted into the SmaI site of pUC18 and then excised
from this vector via FseI and NotI sites. The
FseI-NotI fragment was subsequently cloned
into the same sites of pGS284-TNMIEdlIE1 thereby repairing the
exon four deletion in this plasmid to generate
pGS284-TNMIErvIE1. DNA sequence analysis was completed on all
hCMV-specific PCR amplification products to confirm their integrity. Allelic
exchange was performed through homologous recombination in Escherichia
coli strain GS500 as previously described [33], [210], [211]. First, the BAC
pTNIE1kanlacZ was generated by recombination of T-BACwt
with pGS284-TNIE1kanlacZ followed by selection for kanamycin
resistance and LacZ expression. After that, the BACs pTNdlIE1
and pTNrvIE1 were made through recombination of
pTNIE1kanlacZ with pGS284-TNMIEdlIE1 and
pGS284-TNMIErvIE1, respectively, followed by selection for
the loss of kanamycin resistance and LacZ expression. The BAC constructs were
analyzed by EcoRI digestion. The BACs pTNdlIE1
and pTNrvIE1 were used for electroporation of MRC-5 cells to
reconstitute viruses TNdlIE1 and TNrvIE1,
respectively, as has been described previously [211]. Cell- and serum-free virus
stocks were produced upon BAC transfection of MRC-5 fibroblasts (TNwt and
TNrvIE1) or TetR-IE1 cells (TNdlIE1), and
the titers of the wild-type TN and revertant preparations were determined by
standard plaque assay on MRC-5 cells. Titration of TNdlIE1
stocks was performed by quantification of intracellular genome equivalents [33]. Multistep
replication analysis of recombinant viruses on TetR and TetR-IE1 cells has been
described previously [33].
For global transcriptome analysis, 1.9×106 TetR or TetR-IE1
cells of the same passage number were seeded on 10-cm dishes. When cells reached
confluency (three days after plating), the medium was replaced, and cells were
growth-arrested by maintaining them in the same medium for seven days before
they were collected for transcriptome analysis. During the last 72 h or 24 h
prior to collection, cultures were treated with doxycycline at a final
concentration of 1 µg/ml or were left untreated. Total RNA was isolated
using TRIzol reagent (Invitrogen) and Phase Lock Gel Heavy (Eppendorf) according
to the manufacturers' instructions. A second purification step with
on-column DNase digestion was performed on the isolated RNA using the RNeasy
Mini Kit from Qiagen. All subsequent steps were performed at the
Kompetenzzentrum für Fluoreszente Bioanalytik (Regensburg, Germany). Total
RNA (100 ng) was labeled using reagents and protocols specified in the
Affymetrix GeneChip Whole Transcript (WT) Sense Target Labeling Assay Manual
(P/N 701880 Rev. 4). Quantity and quality of starting total RNA, cRNA, and
single-stranded cDNA were assessed in a NanoDrop spectrophotometer (Thermo
Fisher Scientific) and a 2100 Bioanalyzer (Agilent Technologies), respectively.
Samples were hybridized to Affymetrix Human Gene 1.0 ST Arrays which interrogate
28,869 well-annotated genes and cover >99% of sequences present in the
RefSeq database (National Center for Biotechnology Information). We probed a
total of 18 microarrays, which allowed us to monitor three biological replicates
for each experimental condition (TetR and TetR-IE1 cells without and with 24 h
and 72 h of doxycycline treatment). For creation of the summarized probe
intensity signals, the Robust Multi-Array Average algorithm [213] was
used. Files generated by the Affymetrix GeneChip Operating 1.4 and Expression
Console 1.1 software have been deposited in Gene Expression Omnibus (GEO,
National Center for Biotechnology Information [214]) and are accessible through
GEO Series accession number GSE24434 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE24434).
In order to determine steady-state mRNA levels by qRT-PCR, total RNA was isolated
from 3 to 4×105 fibroblasts using Qiagen's RNeasy Mini Kit
and RNase-Free DNase Set according to the manufacturer's instructions.
First-strand cDNA was synthesized using SuperScript III and
Oligo(dT)20 primers (Invitrogen) starting from 2 µg of
total RNA. Unless otherwise noted, first-strand cDNA was diluted 10-fold with
sterile ultrapure water, and 5 µl were used to template 20-µl
real-time PCRs performed in a Roche LightCycler 1.5 [33]. The instrument was
operated with a ramp rate of 20°C per sec using the following protocol:
pre-incubation cycle (95°C for 10 min, analysis mode: none), 40 to 50
amplification cycles with single fluorescence measurement at the end of the
extension step (denaturation at 95°C for 10 sec, primer-dependent annealing
at 66 to 56°C for 10 sec, primer-dependent extension at 72°C for 8 to 10
sec, analysis mode: quantification), melting curve cycle with continuous data
acquisition during the melting step (denaturation at 95°C for 0 sec,
annealing at 65°C for 60 sec, melting at 95°C for 0 sec with a ramp rate
of 0.1°C/sec, analysis mode: melting curves), cooling cycle (40°C for 30
sec, analysis mode: none). The PCR mix was composed of 9 µl PCR grade
water, 1 µl forward primer solution (10 µM), 1 µl reverse
primer solution (10 µM), and 4 µl 5× concentrated Master Mix
from the LightCycler FastStart DNA MasterPLUS SYBR Green I kit. The
sequences of the high pressure liquid chromatography-purified PCR primers are
listed in Supporting Table S8. All samples were quantified at
least in duplicate, and each analysis included positive, minus-RT, and
non-templated controls. The second derivate maximum method with arithmetic
baseline adjustment (LightCycler Software 3.5) was used to determine
quantification cycle (Cq) values. Cq values were further validated by ensuring
they meet the following criteria: (i) corresponding melting peaks of the
generated PCR products, calculated using the polynomial method with digital
filters enabled, had to match the single peak of the positive control sample,
(ii) standard deviations of Cq values from technical replicates had to be below
0.33, (iii) Cq values had to be significantly different from minus-RT controls
(CqCq-RT-1), and (iv) Cq values had to be within
the linear quantification range. The linear quantification range was
individually determined for each primer pair by generating a standard curve with
serial dilutions of first-strand cDNA from the sample with the highest
expression level. PCR efficiency (E) was calculated from the
slope of the standard curve according to equation (1):(1)The
relative expression ratio (R) of the target
(trgt) and reference (ref) gene in an
experimental (eptl) versus control (ctrl)
sample was calculated using the efficiency-corrected model shown in equation
(2):(2)
Control samples of all experiments had reference and target gene expression
levels well above the limits of detection. The tubulin-β gene (TUBB) was
chosen as a reference, because (i) expression levels did not change upon IE1
induction, IFN treatment, siRNA transfection, or hCMV infection, (ii) it allowed
for RNA-specific detection with no spurious product generation in minus-RT
controls, and (iii) it exhibited similar expression levels compared to the
target genes under investigation, which were generally expressed at levels lower
than TUBB in the absence and at similar or higher levels relative to TUBB in the
presence of IE1 expression, IFN treatment, or hCMV infection.
CXCL9, CXCL10, and CXCL11 chemokine concentrations in cell culture supernatants
were determined using commercially available colorimetric sandwich enzyme
immunoassay kits (Quantikine Immunoassays no. DCX900, DIP100, and DCX110 from
R&D Systems) following the manufacturer's instructions.
The sequences of siRNA duplexes used for mRNA knock-down experiments are listed
in Supporting Table S9. They were introduced into cells at 30 nM final
concentration using the Lipofectamine RNAiMAX Reagent (Invitrogen) following the
manufacturer's instructions. Briefly, exponentially growing cells were
seeded either in 12-well dishes at 2.5×105 cells/well for RNA
analyses or in 6-well dishes at 5×105 cells/well for protein
analyses. Transfections were performed in Opti-MEM I Reduced Serum Medium
(Invitrogen) with 2 µl or 5 µl of RNAiMAX Reagent for 12- or
6-wells, respectively.
Cells (3.8×106) on 10-cm dishes were collected with trypsin/EDTA
and then centrifuged for 5 min at 500× g and 4°C. Supernatants were
removed and cells resuspended in 100 µl CSK buffer (10 mM PIPES [pH
6.8], 300 mM sucrose, 100 mM NaCl, 3 mM MgCl2, 1 mM EDTA,
0.1% (v/v) Igepal CA-630) with freshly added protease and phosphatase
inhibitor cocktails. Lysates were centrifuged for 1 min at 1,300× g and
4°C, and the supernatants (cytoplasmic extracts) were transferred to clean
pre-chilled tubes and combined with one volume of 2× protein sample buffer
(100 mM Tris-HCl [pH 6.8], 4% (w/v) SDS, 20% (v/v)
glycerol, 200 mM β-mercaptoethanol, 0.1% (w/v) bromophenol blue). The
insoluble (pellet) fractions containing nuclei were washed once with 500
µl CSK buffer before they were suspended in 200 µl 2× protein
sample buffer and sonified in a Bioruptor (Diagenode; “H” setting;
30 sec on-off interval) for 15 min. Samples were centrifuged for 10 min at
20,000× g and 4°C, and the supernatants (nuclear extracts) were
transferred to clean pre-chilled tubes. Cytosolic and nuclear extracts were
heated to 95°C for 5 min before immunoblot analysis. Generation of whole
cell extracts, sodium dodecyl sulfate-polyacrylamide gel electrophoresis,
immunoblotting, and (immuno)fluorescence microscopy were performed according to
previously published protocols [33], [53], [215]. Immunodetection employed primary mono- or
polyclonal antibodies directed against hCMV IE1 (1B12; [216]) or human GAPDH (Abcam, no.
ab9485), histone H2A (Abcam, no. ab13923), STAT1 (no. sc-464 for immunoblotting
and no. sc-346 for immunofluorescence, both from Santa Cruz), STAT1α (Santa
Cruz, no. sc-345), STAT2 (Santa Cruz, no. sc-22816), and phosphorylated STAT1
(Y701-specific antibody no. 9171 and S727-specific antibody no. 9177, both from
Cell Signaling Technologies). The secondary antibodies used were
peroxidase-conjugated goat anti-mouse (no. 115-035-166) or goat anti-rabbit IgG
(no. 111-035-144) from Dianova for immunoblotting, and highly cross-adsorbed
Alexa Fluor 594- or Alexa Fluor 633-conjugated goat anti-mouse (no. A-11032 or
no. A-21052, respectively) and Alexa Fluor 546-conjugated goat anti-rabbit IgG
(no. A-11035) from Invitrogen for immunofluorescence.
ChIP was performed essentially as described by Nelson et al.
[217], [218]. Resting
cells on a 15-cm dish were cross-linked by treatment with 1% (v/v)
formaldehyde for 10 min at 37°C. Isolated chromatin was sonified for 15 min
in a Bioruptor (Diagenode; “H” setting, 30 sec on-off interval) and
cleared by centrifugation for 20 min at 20,000× g and 4°C. Sheared
chromatin from 7×106 cells (0.7 ml) was subjected to
immunoprecipitation for 16 h at 4°C with gentle rotation using 10 µg
of antibody. Two different polyclonal rabbit antibodies each against STAT1 (no.
sc-3454 and sc-346 from Santa Cruz) and STAT2 (no. sc-476 and sc-839 from Santa
Cruz) were used. After the antibody incubation step, insoluble material was
removed by centrifugation (10 min at 20,000× g and 4°C) and 0.63 ml
(90%) supernatant was transferred to a clean pre-chilled tube.
Antibody-antigen complexes were isolated by sedimentation following incubation
with 60 µl of Protein A Agarose/Salmon Sperm DNA slurry (Millipore) for 60
min at 4°C. PCR-ready DNA was prepared using Chelex-100 and duplicate
samples of 5 µl (25% of the final reaction volume) each were used
for DNA quantification by qPCR as described above and in recent publications
[33],
[215].
The PCR primer sequences are listed in Supporting Table
S8.
|
10.1371/journal.pcbi.1007205 | Bayesian hypothesis testing and experimental design for two-photon imaging data | Variability, stochastic or otherwise, is a central feature of neural activity. Yet the means by which estimates of variation and uncertainty are derived from noisy observations of neural activity is often heuristic, with more weight given to numerical convenience than statistical rigour. For two-photon imaging data, composed of fundamentally probabilistic streams of photon detections, the problem is particularly acute. Here, we present a statistical pipeline for the inference and analysis of neural activity using Gaussian Process regression, applied to two-photon recordings of light-driven activity in ex vivo mouse retina. We demonstrate the flexibility and extensibility of these models, considering cases with non-stationary statistics, driven by complex parametric stimuli, in signal discrimination, hierarchical clustering and other inference tasks. Sparse approximation methods allow these models to be fitted rapidly, permitting them to actively guide the design of light stimulation in the midst of ongoing two-photon experiments.
| There are many sources of noise in recordings of neural activity, and the first challenge in neural data analysis is to separate this noise from experimentally relevant variation. This is particularly problematic for two-photon imaging data. Two-photon imaging uses fluorescent indicators to measure changes in the concentration of molecules involved in cell signalling, and adds a variety of numerical, biological and optical noise sources. We present a method for disentangling this signal and noise using Gaussian processes, a family of probabilistic models which provide a principled way of inferring mean activity and variability. In addition to signal recovery, we show that these models can test the evidence for whether and where two signals are different and that these tests can be used to look for groups in sets of signals. We explore how these models can be extended to predict how signals will change under different experimental conditions, and that these predictions can be used to select new conditions for further exploration.
| Over the last two decades, two-photon (2P) imaging has become one of the premier tools for studying coding in neural systems from the population level down to individual neural compartments [1]. The resulting data is highly variable due to the inherent variability of neurons and technical sources of noise in the imaging process [2, 3]. Yet we typically assume that beneath the noisy signals which are observed there is a smooth latent function describing the activity of a neuron or a neural compartment. In a typical analysis pipeline for 2P data, we attempt to recover this function by grouping noisy observations from pixels into regions of interest (ROIs), which cover the soma or different compartments of a neuron, temporally interpolating them to a common frame rate and averaging across repetitions of the same stimulus (see also Box 1). Each stage is intended to smooth the observations and get closer to the “true” underlying activity function of the neuron. To measure the uncertainty about this latent activity function, often the variance between repetitions of the same experimental condition is used, with little assessment of whether this reflects the actual uncertainty given measurement and neural variability.
Here, we propose a different approach based on Gaussian Process (GP) regression [4] to infer signals from 2P recordings in a statistically principled manner, propagating the uncertainty all the way from the measurements to the desired inference. This regression procedure recovers an estimate of the true activity of the neuron, whether changes in calcium or glutamate concentration, from observations with experimental noise. This is facilitated by modelling explicitly the change in the signal over time and as a function of stimulus parameters. Gaussian processes are probabilistic models, which describe the functional relationship between a set of predictors and a set of observations (see Box 2 for a mathematical primer). In contrast to typical pre-processing pipelines, the statistical properties of the observed signal are considered explicitly as part of the model optimisation. Recently developed sparse GP approximations allow us to apply these models to comparatively large datasets with several thousand observations, as are common in 2P experiments [5].
Using 2P recordings of calcium and glutamate dynamics in isolated mouse retina, we demonstrate how these models can be used to construct probabilistic representations of neural activity. We treat several use cases: First, we show that GP-based analysis of 2P recordings can be used to perform comparisons between the responses of a given cell under different conditions, allowing one to identify parts of the response with significant differences. Second, we exploit the properties of the GPs to perform a hierarchical clustering of cell responses and provides quantitative criteria for deciding how many clusters to keep. In addition, we use the framework to test which stimulus parameters influence neural activity in an ANOVA-like framework. Finally, we explore how the representation of uncertainty can be exploited for experimental design, informing the choice of parameters to optimally reduce the uncertainty about the neural response.
We applied a Bayesian framework based on Gaussian Process (GP) regression to efficiently infer neural activity with uncertainty estimates from recordings of light stimulus-driven activity in the mouse retina. The retina decomposes a stream of images into parallel channels representing salient stimulus features. The central circuit of this network is a feedforward pathway relaying the initial signal from the photoreceptors through the intermediate bipolar cells to retinal ganglion cells (RGCs), and from there through the optic nerve to the rest of the visual system. Inhibitory interneurons called horizontal and amacrine cells play key roles in the adaptation and feature extraction (for review, see [6]). In the datasets analysed here, we measured three stages of the excitatory pathway: Firstly, the presynaptic calcium signal in the axon terminals of a bipolar cell using the synthetic indicator dyes Oregon-Green BAPTA-1 (OGB-1) and GCaMP6f (the latter data previously published in [7]). Secondly, the glutamate release from these terminals, as measured by the genetically-encoded biosensor iGluSnFR [8, 9]. Finally, the calcium signal in RGC somata loaded with OGB-1 through bulk electroporation [10].
In our framework, a GP model (see Box 2 for mathematical primer) infers an estimate of the activity of a neuron from the observed fluorescence of each ROI in a scan field, which are typically cell somata or axon terminals. The estimated function models the “true” activation state of the neuron, i.e. the concentration of calcium or glutamate. In addition, we model the uncertainty about the estimated function including the observation noise, and the latent uncertainty about the activity function, once the observation noise is removed.
Our first objective was to infer the neural activity function and its associated uncertainty from our observations of the activity of single ROIs, located on individual synaptic axon terminals of a bipolar cell. In bipolar cells injected with the calcium indicator OGB-1, it is possible to image the anatomy of the cell by recording 3D stacks of x-y images at regular intervals along the vertical (z) axis (Fig 1a). High resolution scans allowed us to identify individual axon terminals (Fig 1b). Faster scans with lower spatial resolution are required to resolve neural activity, although the required reduction in resolution is substantially less for spiral configurations relative to classical linear configurations (Fig 1c and 1d). Although the scan patterns are highly regular, the spatial organisation of the neural structures results in irregular sampling over time (Fig 1e–1h).
We recorded bipolar cell calcium and glutamate signals measured during the presentation of a spatially homogeneous light stimulus including a light step and variations in temporal frequency and contrast (Fig 2a, chirp stimulus), as used in previous studies [7, 11]. We used the observed activity of a ROI (Fig 2b), and inferred a signal for each repeat using frame-averaging and cubic-spline interpolation (Fig 2c), corresponding to the classical way of inferring these functions (i.e. [7, 11]).
We then fitted a GP with a radial basis function (RBF) kernel for the time dimension to the observed activity (Fig 2d). We monitored the computation time and calculated the likelihood of an out-of-sample test set to determine a suitable number of inducing inputs. Surprisingly, this indicated that there was already little improvement in the performance of the model when more than 250 data points were used (S1a and S1b Fig), and that relatively few iterations of the fitting algorithm were required (S1c Fig).
To account for temporal non-stationarities in the neural response, we then compared the GP model to an extended model with input warping (see Methods). One assumption of classical GP models is that the function space has a stationary autocorrelation function, i.e. that its correlational structure does not change with respect to a predictor, such as time. However, light induced neural activity like responses to the chirp, which have highly non-stationary correlational structure, are likely to show a commensurate non-stationarity in the response. We computed a warping function which transforms the time dimension such that the stimulus input has a stationary autocorrelation structure (Fig 3a). We then used this warping function to transform the input to the GP model of the response, under the assumption that the correlational structure of the response matched that of the stimulus input [12, 13](Fig 3b–3d). By performing this extra processing step, we were able to fit a model which could vary in its autocorrelation.
Our results show a clear difference between the predictions of the warped GP model and the simpler stationary one. In the simpler model, the selected parameters reflect a trade-off between models which fit closely to each of the different stimulus components (i.e. steps vs. chirps), resulting in an inferred mean signal which appears noisy during the light step and poorly tracks the faster chirp oscillations. As a consequence, the inferred uncertainty was relatively stationary over time (Fig 3e). By contrast, in the warped GP model, the inferred mean signal during the light step was smoother and tracked the faster oscillations much more closely (Fig 3f). More importantly, in contrast to the interpolated signal derived by a classical pipeline, the warped GP infers a high level of uncertainty during periods of rapid oscillation which are at, or close to, the sampling limit of the recording.
In practice, we found that the approach described above was more stable and faster than inferring the autocorrelation function directly from the observed activity. This appeared to be due to two factors: the irregular sampling distribution of the observed activity and the observation noise. Estimating the autocorrelation function separately for each ROI added a considerable computational burden to the pre-processing pipeline. In principle, the approach demonstrated can be applied to any stimulus with a known time-course if it is spatially homogeneous. Where this is not the case, the temporal statistics of the observed response may also be influenced by spatial integration and an alternative model, which explicitly accounted for this, would be appropriate.
The key benefit of the GP framework is that it provides an explicit estimate of the uncertainty about the neural activity which can be used to perform well calibrated statistical inference, e.g. for inferring which periods of neural activity differed between two conditions. This is in contrast to classical approaches, where typical analysis follows multiple smoothing steps and often only the inter-trial variability is considered, providing a poorly calibrated estimate of uncertainty.
In our framework, we use a GP equality test to identify whether two signals are statistically different [14]. As an example, we consider the response of bipolar cells to the chirp stimulus as a function of the spatial extent of the light spot. This is known to modulate bipolar cell responses, with the difference being induced by lateral inhibition [7, 15–17]. We compared the calcium and glutamate signals of bipolar cells presented with chirp stimuli whose light spots differed in size (100μm and full field). We fitted a GP model with time warping to each of the sets of observations (Fig 4a and 4b), performed maximum likelihood estimation to optimise the GP parameters with respect to the data, and then computed the difference between the estimated latent functions.
We identified the periods of activity where the stimulus drives greater differences in the response than would be expected by chance (defined as the three standard deviations around the estimated difference function not including zero). We found the number of disconnected regions where the difference is greater than could be expected by chance, which is called the Euler characteristic (EC). It provides a measure of the strength of the difference between two signals (Fig 4c) and depends on the number of standard deviations chosen as a threshold. To estimate whether the EC was higher than expected by chance for a given threshold, we developed a bootstrap procedure for the GP models. We approximated a null distribution of the EC by shuffling the observed activity between the two conditions, and calculated an empirical p-value with respect to this null distribution of the EC.
If we assume a fixed threshold for calling two regions in the signal different (e.g. three s.d.), we did not find a significant difference between the two stimulus conditions (bootstrap: p ∼ 0.103, α = 0.01) for the calcium recordings, but for the glutamate recording (bootstrap: p ∼ 0, α = 0.01). Significant differences occurred during the light step and in both oscillatory sequences. The shuffle test can also be evaluated for the whole range of thresholds.
For comparison, we computed a similar test using the classical analysis pipeline, using inter-trial standard deviation as an estimate of the uncertainty associated with the mean signal. For the bootstrap procedure shuffled the interpolated data between the two stimulus conditions to approximate the null distribution. At the same threshold as above, for neither the calcium (bootstrap: p ∼ 0.062, α = 0.01) nor the glutamate (bootstrap: p ∼ 0.062, α = 0.01) recording was the EC found to be significantly elevated. It should be noted that the p-values estimated for the GP and classical pipeline are not directly comparable: the classical approach does not distinguish between observational and stimulus driven variability, rather identifying whether observed differences are greater than inter-trial variability.
The choice of a fixed threshold for inferring whether two signals are statistically distinct may result in overly conservative statistical estimates. While there were no thresholds for which the classical approach inferred an EC greater than expected from the null distribution, for the GP models of the calcium and glutamate signals there were a range of thresholds for which the EC was greater. These ranges differed between the two signals, which may relate to the effect of the physiological properties of these two signals or to the kinetics of their respective indicators on the dynamics of the observed signals. We recommend that the selection of a threshold be guided by the particular needs of the analysis task, not to keep to statistical conventions developed for other methods.
We next show how GP equality tests can be used to provide a principled criterion for choosing the number of clusters in a hierarchical clustering of light responses. For example, in a single imaging plane, one may wish to know whether the observed responses originate form distinct functional groups, perhaps due to the presence of multiple cells or cell types within the recording plane, or multiple neurites of the same cell acting independently [18].
This pipeline was composed of two stages, firstly identifying putative clusters, then evaluating the evidence for different cluster configurations. In the first stage, a GP was estimated for each ROI in GCaMP6f labelled bipolar cell axon terminals in a PCP2 mouse line (data previously published in [7]). Then, we hierarchically clustered the mean signals from each GP to identify putative clusters among the set of responses, using the Ward algorithm and Euclidean distance (Fig 5a).
Each node in the hierarchy then corresponded to a hypothesis about whether a particular cluster should be partitioned into two sub-clusters (Fig 5b). In the second stage, we start at the top of the clustering. At each node, we fit two GPs to the data from all ROIs assigned to each of the two clusters independently.
We then tested the hypothesis that the two clusters were different using a GP equality test with the EC as the measure of dissimilarity (Fig 5c–5f). A null distribution for the Euler Characteristic was approximated by a further bootstrap test, where the pair of signals for which the null distribution of the EC was calculated were drawn at random from the pooled observations at each node. This process continued iteratively through the hierarchy, terminating when the Euler Characteristic for a split in two new clusters was not greater than 99.5% of the null distribution at that node.
Interestingly, the first node (N0) separated ROIs belonging to two bipolar cells in the imaging field, with strong quantitative backup for the split (bootstrap: p ∼ 0 at three s.d., α = 0.01, Fig 5g). The split at the second node (N1) was also accepted (bootstrap: p ∼ 0, α = 0.01 at three s.d., Fig 5h), which separates ROIs of the left bipolar cell into two groups, indicating potential sub-clusters within the terminals of a single bipolar cell. Subsequent separations were rejected (N2, bootstrap: p ∼ 0.68, α = 0.01 at three s.d., Fig 5i). The difference observed within the terminal systems of these cells may reflect functional variation within the output of a single bipolar cell [19]. Were this difference to exist, it would likely be a consequence of differential inhibition from amacrine cells, and represent an additional layer of complexity in the functional parallelisation of retinal signalling. While our analysis is suggestive of this conclusion, verification is beyond the scope of this study.
We next extended our GP framework to study the effect of multiple stimulus parameters and their interactions on the latent neural activity in an ANOVA-like framework [20]. GP-ANOVA models posses multiple kernels, each of which models the effect of a predictor or an interaction between predictors. In contrast to classical ANOVA, the interaction effects can have non-linear structure [20], and it is possible to compute not merely the strength of particular effects but also an inference of the response of a ROI over time as a stimulus feature varies.
To demonstrate the usefulness of this extension, we fitted a GP model to predict the response of a single ROI to a light stimulus where light intensity was modulated as a sine wave of varying frequency and contrast (Fig 6a). The input for this model was a predictor matrix where each column corresponded to one of the stimulus parameters, including two columns jointly encoding phase as a circular feature, and one each for frequency and contrast (see Methods).
For the experiments, we selected 150 stimulus parameters using blue noise sampling, such that parameters were uniformly selected from the parameter space and excluded if they were below certain thresholds for frequency (< 1Hz) or contrast (< 10%) or too close to an already existing stimulus parameter. Although the frequency and contrast parameters are fixed during each one second trial of the sine stimulus, the model can accommodate parameters which vary continuously over time, by encoding this change in the columns of the predictor matrix.
As in a classical ANOVA, there are many possible ways in which the effects of these stimulus parameters can be incorporated into the model. In this case, stimulus features were encoded in the kernel (see Methods), either as additive independent effects of phase, contrast or frequency, or through multiplicative interactions between the features. The cost of adding more kernels with a fixed amount of data is that the uncertainty associated with each parameter increases as the number of parameters to be learned grows. To compensate for this, we performed kernel selection through a two stage iterative process (Fig 6b). The first stage identified the kernel which, when included, most strongly improved model performance, as measured by the log marginal likelihood. Once there were two or more parameters, each new kernel had to contribute a greater improvement to the model performance than could be expected by chance, as established by a likelihood ratio test (see Methods). If a kernel was accepted it was retained in the model in the consecutive iterations (for an overview of the models evaluated in this pipeline, see S1 and S2 Tables).
We fitted a GP model to the glutamate signal of a single ROI in the IPL in response to the sine stimulus using this procedure. After three iterations the improvement in model performance was less than the required ratio. We tested one further iteration which also returned a negative result, and the process ceased. The kernels which were accepted included an interaction kernel between all three parameters and a frequency-contrast kernel (Λ = 11.25, p < 0.001). A frequency kernel (Λ = 4.01946568, p = 0.045) was rejected in the third stage, and a phase kernel was rejected (Λ = 2.29, p = 0.130).
We then used the model to predict neural activity for unseen parameter combinations and quantified how uncertain our predictions about the activity in response to these were [21–23]. Intuitively, the model should have the least uncertainty about stimulus parameters which had been observed. Uncertainty then should increase as a function of the distance from the observed parameters. We quantified uncertainty by computing the expected response of the ROI and taking the sum of the latent variance (Fig 6c).
For the studied cell, calcium recordings to stimulation with the chirp stimulus were also available (Fig 6d), and we compared the model fitted directly to the chirp response data to predictions from the model fitted to the sine data. There were some qualitative similarities between the two models, such as the overall amplitude of the signal, and the decrease in signal amplitude as the frequency of the stimulus increased. The prediction that the signal amplitude would slightly increase with contrast was not reflected in the chirp data, where the relationship was more ambiguous. The quality of prediction of the activity at high frequencies was difficult to evaluate, as there is a high-level of uncertainty about the mean signal at those frequencies. One factor to consider with regards to this direct comparison is that the differences in the chirp responses may be due to temporal dependencies over time.
A critical advantage of our framework is that we can use it for Bayesian experimental design. This is useful, as in 2P imaging experiments time is usually severely limited. For example, isolated mouse retinal tissue becomes unresponsive to light stimulation in a matter of hours, and single recording fields often bleach within half an hour of recording. To efficiently explore the space of possible stimulus features under severe time constraints is thus a critical problem, which GP models can be used to address [21, 22].
To show how this works, we performed an experiment using GP models to guide parameter selection. In retinal tissue expressing iGluSnFR we selected a single ROI, likely representing a single bipolar cell axon terminal. We used two control stimuli to evaluate the parameter selection: a local chirp stimulus playing over three trials, to which we fitted a warped GP, and a sinusoidal stimulus with 90 parameters uniformly sampled from the parameter space (Fig 7a), to which we fitted a GP with the kernels derived in the previous likelihood ratio procedure. We performed three rounds of active parameter selection, starting with 30 uniformly sampled parameters in the first iteration, fitting the GP and using parameters selected by identifying 30 peaks in the uncertainty map in the subsequent two iterations. We then used the models from each iteration to predict how the ROI would respond during the oscillatory components of the chirp stimulus (Fig 7b).
Parameters selected using the active approach were more broadly distributed across the parameter space, although we noted that the peak finding algorithm was biased away from the edges. In the purely random design procedure, parameters often clustered and there were large empty regions, resulting in high uncertainty in these regions. Neither the random nor the active parameter procedure inferred a good prediction of the contrast-varying chirp component, which in the case of the active parameter inference was likely due to the bias away from the periphery of the parameter space, resulting in very few samples in the proximity of the 8Hz parametric edge. At lower frequencies the experimental design algorithm seemed better able to capture qualitative aspects of the chirp response, such as the decrease in response amplitude as the frequency increased, though again the lack of samples at the very highest frequencies resulted in a high level of uncertainty.
We finally constructed a model which combined stimulus effect modelling and hierarchical clustering into a single framework. We fitted the model to calcium recordings of RGC activity in response to a bright bar moving in different directions on a dark background. RGCs show different response polarities and a large range of response kinetics to this stimulus [11] and some modulate the response amplitude as a function of stimulus direction. The model incorporated the stimulus features of time and direction as additive effects, alongside with a time-direction interaction effect (Fig 8a and 8b). The data were then sorted using hierarchical clustering (Fig 8c and 8d; S2 Fig) and for the purpose of demonstration the first three nodes of the hierarchy were tested using GP equality tests (Fig 8e–8h).
The algorithm first separated ON and OFF responses into separate clusters (N0, bootstrap: p ∼ 0 at three s.d., Fig 8i). The ON cluster was then further divided into sustained and transient responses (N1, bootstrap: p ∼ 0 at three s.d., Fig 8j). The sustained ON responses were finally separated into direction selective and non-direction selective clusters (N2, bootstrap: p ∼ 0.01 at three s.d., Fig 8k). We did not test further splits for significance.
Here we presented a data analysis pipeline for 2P imaging data based on Gaussian Processes. The advantage of this framework is that uncertainty about the underlying latent neural activity can be propagated through the analysis pipeline, so statistical inference can be performed in a principled way. We applied our pipeline to recordings of mouse retinal bipolar and ganglion cell activity driven by light stimuli, showing how: (1) to determine whether and when two response functions are statistically distinct; (2) to evaluate the strength of the evidence for a partition of data into functional clusters; (3) to determine the relevant stimulus effects to incorporate into a model of neural responses; and (4) to guide the choice of stimulus parameters for iterations of a closed loop adaptive experiment.
Accurately characterising the variability of neural responses is essential for understanding neural coding. Noise manifests itself throughout sensory systems and presents a fundamental problem for information processing [2]. While imaging ex-vivo retinal tissue does not present some of the challenges as in vivo cortical recordings (where movement is a significant source of variability), two-photon imaging in ex-vivo tissue is still subject to many sources of variance, due to fluctuations in biosensor excitation and photon detection, among other factors. This issue may be particularly acute for two-photon imaging of retinal tissue, where it is necessary to keep the excitation energy low to avoid laser-evoked responses, which may result in lower overall fluorescent signals relative to in vivo recordings of non-light-sensitive tissue. Computational processing can introduce further variability, e.g. due to the discretisation of the measured signal. This is rarely acknowledged, perhaps due to the convenience of standard approaches. In principle, splines in combination with generalized additive models (GAMs, e.g. [24]) provide an alternative framework to perform uncertainty aware analysis of calcium imaging data. Exploring and contrasting this to the GP framework introduced here is beyond the scope of this paper.
Classical GP models can be computationally costly due to the need to compute the inverse of the kernel matrix involving all training data [4]. To make practical use of GPs for modelling large 2P recordings, we capitalized on recent advances in sparse approximations for GPs that work with a limited number of inducing points [5] and demonstrated their applicability for a real world task. In addition, we only performed point-estimates for hyperparameters instead of fully Bayesian inference and pre-determined kernels before statistical evaluation. This was important for two reasons: firstly, a processing pipeline should not be excessively computationally costly, so as to make them impractical for general use with larger datasets; and secondly, the application of these models in closed-loop imaging experiments was only possible if one complete iteration of the process (data acquisition; pre-processing; prediction; parameter selection) could be completed in a few minutes. In principle, our approach could be extended to a fully Bayesian framework with hyperpriors on the model parameters, although this introduces additional difficulties for sparse approximation and still entails a greater computational burden [25]. While our work solely addressed Gaussian distributed data, the models can be readily extended to point processes as well. There, sparse approximation techniques overcome the computational intractability of the model, and allow inference on relatively large datasets [26].
Although we demonstrated the potential for using GP models during 2P experiments, there were several limitations to our approach. We were able to reduce the time per iteration of our active experiments to less than five minutes, addressing a key practical concern. However, it emerged during the experiment that the peak-finding algorithm was biased away from the periphery of the stimulus space, which made the chirp stimulus unsuitable as our “ground truth” for model evaluation.
The parameter batch size may also have been too small for each iteration. Batch size is a critical consideration in active Bayesian experimentation. Where the cost per iteration is low, single parameters can be selected for each iteration, for which the objective function can be relatively easily defined and evaluated. In one recent publication, Charles et al. [23] used GPs to model the effect of inter-trial variability in monkey V1 neurons, using sequential parameter selection to optimise a coloured light stimulus. For experiments where iterations are prohibitively expensive, new parameters can be selected in batches, although this requires interactions between parameters to be taken into account, which can be computationally expensive to evaluate. In such cases, approximate methods provide an attractive method for reducing computational overheads (e.g. [22]). Batch parameter selection algorithms which account for, or approximate, parameter interactions would likely overcome simple peak finding methods.
Historical obstacles to the use of Bayesian methods such as the difficulty of their implementation and their computational cost have been reduced. Much research over the past decade has focused on the problem of minimising the computational complexity of the algorithms through sparse approximation methods and efficient parameter estimation (such as [5]). New libraries for popular coding languages—such as GPy [27], PyMC3 [28], and pySTAN [29] for Python 3—have lowered the barrier to entry. Likewise, we provide a collection of notebooks with this paper to allow straightforward application of our framework.
Taken together, our approach exploits the flexibility and extensibility of Gaussian process models to improve on classical approaches for two photon data analysis and addresses important analytical tasks in a way that preserves a representation of uncertainty propagated up from the underlying data. We feel that it will be particularly useful for disentangling the dynamics of neural circuits in the early visual system under complex, multivariate experimental conditions.
All animal procedures were performed according to the laws governing animal experimentation issued by the German Government. The documentation for the animal and tissue preparation was submitted in accordance with Mitteilung nach §4 Abs. 3 Tierschutzgesetz, and approved by the Regierungspräsidium Tübingen on 09.11.2016. Viral injection documentation Tierversuch Nr. AK6/13 was appraised by the ethics committee and approved by Regierungspräsidium Tübingen, on 05.11.2013.
For single-cell-injection experiments, we used one adult mouse cross-bred between transgenic line B6.Cg-Tg(Pcp2-cre)3555Jdhu/J (Tg3555, JAX 010536) and the Cre-dependent red fluorescence reporter line B6;129S6-Gt(ROSA)26Sortm9(CAG-tdTomato)Hze/J (Ai9tdTomato, JAX 007905). For glutamate-imaging, we used one adult C57BL/6J mouse. Owing to the exploratory nature of our study, we did not use blinding and did not perform a power analysis to predetermine sample size.
Animals were housed under a standard 12h day-night cycle. For recordings, animals were dark-adapted for ≤ 1h, then anaesthetised with isoflurane (Baxter) and killed by cervical dislocation. The eyes were removed and hemisected in carboxygenated (95% O2, 5% CO2) artificial cerebral spinal fluid (ACSF) solution containing (in mM): 125 NaCl, 2.5 KCl, 2 CaCl2, 1 MgCl2, 1.25 NaH2PO4, 26 NaHCO3, 20 glucose, and 0.5 L-glutamine (pH 7.4). Then, the tissue was moved to the recording chamber of the microscope, where it was continuously perfused with carboxygenated ACSF at ∼ 37°C. The ACSF contained ∼ 0.1μM sulforhodamine-101 (SR101, Invitrogen) to reveal blood vessels and any damaged cells in the red fluorescence channel. All procedures were carried out under very dim red (>650nm) light.
Sharp electrodes were pulled on a P-1000 micropipette puller (Sutter Instruments) with resistances between 70–100MΩ. Afterwards, the tip (∼ 500μm) of each electrode was bent on a custom-made microforge. Single bipolar cell somata in the inner nuclear layer were filled with the fluorescent calcium indicator Oregon-Green BAPTA-1 (OGB-1) by using the pulse function (500ms) of the MultiClamp 700B software (Molecular Devices). OGB-1 (hexapotassium salt; Life Technologies) was prepared as 15mM in distilled water. Immediately after filling, the electrode was carefully retracted. Imaging started after about 30 minutes after the injection to allow cells to recover and the dye to diffuse within the cell. At the end of the recording, a stack of images was captured for the cellular morphology, which was then traced semi-automatically using the Simple Neurite Tracer plugin implemented in Fiji [30].
For virus injections, we used adult wild-type mice (C57BL/6J). Animals were anesthetized with 10% ketamine (Bela-Pharm GmbH & Co. KG) and 2% xylazine (Rompun, Bayer Vital GmbH) in 0.9% NaCl (Fresenius). A volume of 1μl of the viral construct (AAV2.hSyn.iGluSnFR.WPRE.SV40, Penn Vector Core) was injected into the vitreous humour of both eyes via a Hamilton injection system (syringe: 7634-01, needles: 207434, point style 3, length 51mm, Hamilton Messtechnik GmbH) mounted on a micromanipulator (World Precision Instruments). Imaging experiments were performed 3 weeks after virus injection.
We used a MOM-type 2P microscope (designed by W. Denk, now MPI Martinsried; purchased from Sutter Instruments/Science Products). The design and procedures have been described previously [7, 11, 31]). In brief, the system was equipped with a mode-locked Ti:Sapphire laser (MaiTai-HP DeepSee, Newport Spectra-Physics), two fluorescence detection channels for OGB-1 or iGluSnFR (HQ 510/84, AHF/Chroma) and SR101/tdTomato (HQ 630/60, AHF), and a water immersion objective (W Plan-Apochromat 20x /1.0 DIC M27, Zeiss). The laser was tuned to 927nm for imaging OGB-1, iGluSnFR or SR101. For image acquisition, we used custom-made software (ScanM by M. Müller, MPI Martinsried, and T. Euler) running under IGOR Pro 6.3 for Windows (Wavemetrics), taking time lapsed 32 x 32 pixel image scans (at 15.625Hz) or 16-line “spiral” scans (at 31.25Hz). For documenting morphology, 512 x 512 pixel images were acquired with step size of 0.5μm along the Z axis.
To resolve transient changes in calcium concentration or glutamate release (i.e. with decay times of ∼100ms), scan rates of around 20Hz or more are wanted. Many scanning 2P microscopes use conventional (non-resonant) galvanometric scanners and are limited by the inertia of the scan mirrors, which introduce positional errors at high scan rates. This is especially critical for typical linear (image) scans, with their abrupt changes in direction when jumping between scan lines. For constant spatial resolution, faster scan rates are often realised by decreasing the scan area. However, it is possible to increase the spatio-temporal resolution by using non-linear “spiral scan” configurations. These overcome the key mechanical limitation of linear scans, that they incorporate sharp turns, rather than following smoother trajectories. Unlike linear scans, which are composed of single linear trajectories repeated along an axis at regular intervals, spiral scan configurations consist of radial trajectories moving away from a central point at a constant speed and rotation and permit rapid movement of the scan mirrors.
A regular radial grid can be constructed by generating a single spiral trajectory and successively rotating it around a central point. We used an Archimedean spiral is used to generate each trajectory (r = Θ1/a), where the radial distance r from the central point is a function of the angle Θ and a tightness parameter a which determines the rate of rotation around the centre. With a grid composed of 16 such curves we can resolve, for instance, axon terminals of retinal bipolar cells at twice the spatial and twice the temporal resolution of linear recordings. One can see the advantages of such scan configurations by showing how frequently the scan trajectory intersects with ROIs in a single frame. The times at which labelled structures are observed by these trajectories are both more frequent and more irregularly distributed in time than a typical linear scan, providing a superior temporal resolution.
For light stimulation, a modified LightCrafter (DLPLCR4500, Texas instruments; modification by EKB Technology) was focused through the objective lens of the microscope. Instead of standard RGB light-emitting diodes (LEDs), it was fitted with a green (576nm) and a UV (390nm) LED for matching the spectral sensitivity of mouse M- and S-opsins [32]. To prevent the LEDs from interfering with the fluorescence detection, the light from the projector was band-pass-filtered (ET Dualband Exciter, 380-407/562-589, AHF) and the LEDs were synchronised with the microscope’s scan retrace. Stimulator intensity was calibrated to range from 0.5 * 103 (“black” background image) to 20 * 103 (“white” full field) photoisomerisations P*/s/cone [7]. The light stimulus was centred before every experiment, such that its centre corresponded to the centre of the recording field. In linear scans, the stimulus is displayed while the trajectory moves between consecutive lines; while for the spiral scans this occurs while the trajectory returns from the periphery to the centre.
Light stimuli were generated using the QDSpy light stimulation software, which is written in Python 3 [33]. The chirp stimulus ran for 4 repeats of 32s each, with the stimulus extent alternating between a 800μm and a 100μm light spot. The moving bar stimulus consisted of a 300μm rectangular bar moving at 1000μm/s for 4 seconds along 8 evenly space directions, repeated three times for each direction. The sine stimulus consisted of a 100μm light spot, and ran for 45 1s-trials, with contrast and frequency varying in each trial. The contrast and frequency parameters were chosen by blue-noise sampling 150 parameters from the parameter space, between 10% and 100% contrast and 1Hz to 8Hz frequency. Later closed-loop experiments used a sine stimulus with 90 parameter sets sampled uniformly from the parameter space, in addition to 3x30 parameters sets, of which the first were chosen from random uniform sampling and the latter two sets by active Bayesian inference.
Initial data analysis was performed in IGOR Pro 6. Regions of Interest (ROIs) were defined manually using the SARFIA toolbox for IGOR Pro [34]. In the iGluSnFR recordings, a custom-script generated a correlation map [7], which defined structures for the ROI drawing. The observations were synchronised to the light stimuli using time markers which were generated by the stimulation software and acquired during imaging. Once the initial pre-processing was completed, the data was exported to HDF5 files, and all subsequent analysis was performed in Python 3.5.
Gaussian process (GP) models were used to infer the relationship between time, stimulus parameters and the observed activity of each ROI. Thus, the predictor matrix X was a function of the stimulus parameters and time, short hand referred to as θ. An introduction to the mathematics of GP regression is provided in Box 2. All GPs used the Radial Basis Function (RBF) kernel, with additive Gaussian noise.
k R B F , ϕ ( X , X ′ ) = σ s i g n a l 2 exp ( - ∥ X - X ′ ∥ 2 2 l 2 ) + I σ n o i s e 2 (10)
ϕ = { l , σ s i g n a l , σ n o i s e } (11)
The lengthscale l, signal variance σsignal and noise variance σnoise were learned as part of the model optimisation. Since the fluorescence measurements Fi,θ for ROI i were irregularly spaced in time, the mean μ and covariance Σ of the signal Fi were inferred for a new set of predictors X* where time is regularly spaced:
μ ϕ ( X * | X ) = k ϕ ( X * , X ) ( k ϕ ( X , X ) + I σ n o i s e 2 ) - 1 F i , θ (12)
Σ ϕ ( X * | X ) = k ϕ ( X * , X * ) - k ϕ ( X * , X ) ( k ϕ ( X , X ) + I σ n o i s e 2 ) - 1 k ϕ ( X , X * ) (13)
The additive noise component I σ n o i s e 2 was removed for statistical inference, and we refer to the resultant noise-free GP as the “latent function”, in line with the terminology in the GPy documentation [27]. Confidence intervals were calculated as
μ ϕ ( X * ) ± 3 * d i a g ( Σ ϕ ( X * ) ) (14)
The Gaussian process models were developed in the GPy framework [27]. Feature encoding, input warping, equality tests, parameter selection and closed-loop parameter selection were computed using custom scripts, which we provide as supplementary content to this document and online at https://github.com/berenslab/bayesian_2p_pipeline. Hierarchical clustering was performed using scripts from the Scipy library, using Euclidean distance, the Ward algorithm and maxclust as the criteria [35]. The Ward algorithm was chosen as it tends to infer balanced clusters across the hierarchy. Adaptive parameter selection used a local peak finding algorithm from the Scikit-Image library.
Since our datasets included several thousand observations, it was necessary to use sparse approximation methods to fit the GP models. The sparse approximation algorithm provided in GPy follows [5], whereby the kernel is approximated using a subset of the data, termed the inducing inputs. The selection of the inducing inputs is learned as part of the model optimisation, selecting the inputs which minimise the KL-Divergence between the approximation and the target distribution. Details are provided in [5].
The Gaussian process equality test establishes whether two functions modelled by GPs are equal [14]. It operates by computing the difference between the two distributions and identifying whether the credible region encompasses the zero vector across the complete domain of the predictors. If the zero vector is outside of these intervals, we say the two functions are distinct with probability 1 − a. The probability is calculated using the mean μ* and covariance k* of the posterior of our two functions, excluding their respective noise components from the estimate of the covariance.
μϕΔ(X*|X)μϕTΣϕΔ(X*|X)ΣϕμϕΔ(X*|X)≤χ2(1−a) (15)
μ ϕ Δ ( X * | X ) = μ ϕ 1 ( X * | X ) - μ ϕ 2 ( X * | X ) (16)
Σ ϕ Δ ( X * | X ) = Σ ϕ 1 ( X * | X ) + Σ ϕ 2 ( X * | X ) (17)
The total number of discrete, non-intersecting regions where two Gaussian processes differ more than could be expected by chance is termed the Euler characteristic (EC) [36]. The EC is a measure of the geometry of random fields which accounts for the smoothness of the underlying functions, and is well established in fMRI research, where it forms part of the broader literature on statistical mapping [37]. While the expected value of the EC can be analytically tractable under certain conditions, we wished to incorporate it into our pipeline in a manner which was not sensitive to the number of input dimensions and could handle non-stationary autocorrelation functions, and so inferred its null distribution through bootstrap resampling.
To evaluate whether the values of the EC, which were obtained from the GP equality tests, were statistically significant, we constructed an approximate null distribution by bootstrapping samples from pooled data and performing equality tests on these samples. The procedure was as follows: the observations from each of the signals being compared were pooled to form a larger set of observations; from this set, pairs of samples each 300 observations in size were drawn at random, without replacement; Gaussian processes were fitted to each of the samples in the pair; the difference between the two Gaussian process models was calculated; the Euler characteristic was calculated from this difference for varying thresholds. This was repeated 500 times to build an approximate null distribution.
We applied this bootstrap test in Figs 4, 5 and 8. For Fig 4, the observations were pooled from the responses to the stimulus; for Figs 5 and 8, for each node the observations were pooled from the two putative clusters. For the comparisons to the classical pipeline, the null distribution was computed by shuffling the observations between stimulus conditions, with a total of 500 shuffled pairs used for the estimation. Approximate p-values were computed by calculating the proportion of the N shuffled sample pairs which had a greater EC value than that calculated from the GP equality test.
To address non-stationarity of the chirp response data, we computed the autocorrelation function for the chirp stimulus in 500 ms windows spaced at 1/16 s intervals (512 windows total). As we used RBF kernels for our regression, we fitted a Gaussian curve to each autocorrelation function and retained the inferred lengthscale lt for each window. A further parameter A modulates the height of the function.
c o v ( f s t i m u l u s ) ∼ A e ( x - μ ) 2 2 l t 2 (18)
If the signal were stationary, we would observe that the lengthscale parameter was constant with respect to time.
By using the cumulative sum of the inverse of the lengthscale as the predictor, we could derive a warping function which transformed the predictors such that the stimulus autocorrelation was stationary.
We assumed that the autocorrelation of the observed signal was approximately equal to that of the light stimulus input, and used the warping function to transform the observations. This transformation could be inverted to visualise the fitted GP with respect to the original time base.
f w a r p e d ( x t ) ∼ 1 2 l t 2 (19)
GP models can also be used for functional Analysis of Variance [20]. These GP-ANOVA models disentangle the contribution and interaction of different predictors to the observed function. The GP models for the chirp stimulus data modelled the observed activity with time relative to the start of each stimulus trial as the predictor X. For the moving bar stimulus, a direction parameter was encoded as a 2D circular feature by converting the angle α in polar coordinates to an xy position in Cartesian coordinates (cos(2πα/360), sin(2πα/360)). Likewise, for the sine wave stimuli the phase of the oscillation was encoded as (cos(2πtf), sin(2πtf)), while frequency and contrast were encoded linearly.
In contrast to classical ANOVA, the effects and interactions can be non-linear. Flexible kernel composition makes such models comparatively simple to implement. Kernels can be combined in a number of ways [4], each expressing some belief about the effect of a parameter, most commonly by taking the sum or product of two kernels. Additive components represent effects of predictors which are independent of one another, while multiplicative kernels represent interactions between predictors. For example, for a kernel encoding two stimulus parameters xa and xb, with both additive and interactive effects and RBF kernels, the correlation function of the GP model would be:
k ϕ ( X , X ′ ) = k ϕ ( x a , x a ′ ) + k ϕ ( x b , x b ′ ) + k ϕ ( X a , b , X a , b ′ ) (20)
Here, Xa,b = (xa, xb). We estimated interaction effects of different stimulus parameters in our GP ANOVA models by including kernels which learned a single lengthscale parameter over multiple input dimensions. This inferred the joint effect of the parameters as a single function; where, since the parameters are z-scored, a change in the magnitude of one stimulus parameter would have the same effect as varying the other by an equal magnitude. This approach to GP ANOVA provides an efficient and principled way of choosing optimal hyperparameters to infer stimulus effects.
For the chirp stimulus, where there is one predictor, a single kernel encoding the autocorrelation of the signal over time was used. For the warped GPs, the warped time was used instead. For the moving bar and sine wave stimuli, additional kernels were included to model the effects of their respective parameters. The GP model for the moving bar responses included both additive effects for time and direction, and a time-direction interaction effect. Likelihood ratio tests were used to select kernels from the full set of additive and multiplicative stimulus effects:
χ 2 = - 2 l n ( L 0 L 1 ) (21)
Where LN is the likelihood of the fitted model, and the addition of the proposed parameter is rejected if the improvement in the likelihood is greater than chance with probability 1 − a. These tests were applied iteratively until a kernel was rejected. For the closed loop experimentation, we retained the model from the previous selection procedure with the randomly parameterised sine stimulus.
The data used throughout this paper and corresponding code used to compute the models will be provided as supplementary material alongside this paper.
|
10.1371/journal.ppat.1003889 | Detection of Host-Derived Sphingosine by Pseudomonas aeruginosa Is Important for Survival in the Murine Lung | Pseudomonas aeruginosa is a common environmental bacterium that is also a significant opportunistic pathogen, particularly of the human lung. We must understand how P. aeruginosa responds to the lung environment in order to identify the regulatory changes that bacteria use to establish and maintain infections. The P. aeruginosa response to pulmonary surfactant was used as a model to identify transcripts likely induced during lung infection. The most highly induced transcript in pulmonary surfactant, PA5325 (sphA), is regulated by an AraC-family transcription factor, PA5324 (SphR). We found that sphA was specifically induced by sphingosine in an SphR-dependent manner, and also via metabolism of sphingomyelin, ceramide, or sphingoshine-1-phosphate to sphingosine. These sphingolipids not only play a structural role in lipid membranes, but some are also intracellular and intercellular signaling molecules important in normal eukaryotic cell functions as well as orchestrating immune responses. The members of the SphR transcriptome were identified by microarray analyses, and DNA binding assays showed specific interaction of these promoters with SphR, which enabled us to determine the consensus SphR binding site. SphR binding to DNA was modified by sphingosine and we used labeled sphingosine to demonstrate direct binding of sphingosine by SphR. Deletion of sphR resulted in reduced bacterial survival during mouse lung infection. In vitro experiments show that deletion of sphR increases sensitivity to the antimicrobial effects of sphingosine which could, in part, explain the in vivo phenotype. This is the first identification of a sphingosine-responsive transcription factor in bacteria. We predict that SphR transcriptional regulation may be important in response to many sites of infection in eukaryotes and the presence of homologous transcription factors in other pathogens suggests that sphingosine detection is not limited to P. aeruginosa.
| Many opportunistic pathogens transition from an environmental niche into the host. To establish an infection, these bacteria must rapidly adapt their transcriptional profile to the conditions at the site of infection. We used the response of Pseudomonas aeruginosa to lung surfactant as a model to discover genes important for bacterial survival during mouse lung infection. Using this model we identified transcripts induced in response to host-derived sphingolipids, accomplished by detection of the core component sphingosine by a sphingosine-binding transcription factor, SphR. Deletion of this transcription factor in P. aeruginosa reduced bacterial survival, highlighting the importance of a proper response to host-derived sphingosine. We present evidence that impaired survival against the antimicrobial effects of sphingosine may explain part of the in vivo survival defect of mutants in this response system. This is the first description of a specific bacterial response to sphingosine and its precursors, some of which are important immune signaling molecules. Thus, P. aeruginosa is capable of intercepting and responding to host immune modulatory signals. The importance of this response during infection and the presence of similar systems in other pathogens opens up a new avenue for investigation and expands our understanding of bacterial metabolic interactions with the host.
| Pseudomonas aeruginosa is a common, Gram negative, environmental bacterium that can cause significant disease as an opportunistic pathogen, particularly in the lung. P. aeruginosa lung infections are prevalent in people with cystic fibrosis (CF) and chronic obstructive pulmonary disease (COPD), as well as individuals undergoing mechanical ventilation [1]–[4]. These infections cause significant morbidity and mortality and continue to be a major health care burden [5]–[8]. P. aeruginosa has a large genome by bacterial standards (∼6 Mbp) containing a high proportion of regulatory genes (∼8% are predicted to be transcriptional regulators) [9]. This large regulatory capacity is likely important for P. aeruginosa success as an opportunist, enabling it to rapidly alter gene expression in response to host-derived factors and environmental conditions. Understanding mechanisms by which P. aeruginosa detects and responds to the host could present new avenues to combat these devastating, and often antibiotic resistant [10], opportunistic infections. Our current understanding of P. aeruginosa response to the host come from transcriptional profiling using epithelial cells, mucus, or CF sputum [11], [12]. We were interested in the response of P. aeruginosa to the environment of the distal airway, particularly the response to mammalian pulmonary surfactant, the lipid rich mixture that coats the airway surface liquid of the lungs and participates in both respiratory physiology and host defense (reviewed in [13], [14]). This mixture is rich in phosphatidylcholine (∼75% by mass) but also has a substantial fraction of other phospholipids, cholesterol and its esters, and sphingolipids [15].
Sphingolipids constitute a class of molecules that are critical components of eukaryotic cell membranes. In addition to this structural role in membranes and their biophysical role in pulmonary surfactant, many sphingolipids have been shown to act as signaling molecules that play critical roles in regulation of diverse physiological processes. The broad importance of sphingolipid signaling in eukaryotic hosts has only recently been appreciated, and the rapidly expanding field has many recent reviews [16]–[19]. Sphingosine serves as a backbone component for all sphingolipids, which include the signaling molecules sphingosine-1-phosphate (S1P) and ceramide, as well as the structural lipid sphingomyelin. S1P, in particular, has been intensely studied in the past decade as a potent immune signaling molecule that plays a critical role in diverse immune functions such as lymphocyte trafficking, myeloid cell activation, and epithelial and endothelial barrier function, mediated by five G-protein coupled receptors [20]–[24]. Importantly, S1P is released by endothelial cells and platelets during the acute phase response and therefore plays an important role in the initial response to infection [25]–[28].
A specific transcriptional response to host derived sphingolipids and S1P has never been previously shown in bacteria. Here we have identified P. aeruginosa genes induced in response to mammalian pulmonary surfactant and subsequently characterized a subset of genes that are specifically and directly regulated by sphingosine or via metabolism of S1P, sphingomyelin, or ceramide to sphingosine. This response to sphingosine and its precursors is dependent on an AraC-family transcription factor in response to physiological levels of sphingosine and its precursors. This transcription factor binds sphingosine, which alters its association with DNA. A bacterial system to detect and respond to sphingosine may have broad implications in the modulation of host immune function and aid P. aeruginosa in altering host immune response in the human lung. In support of this prediction, deletion of the sphingosine-responsive transcription factor confers a survival defect during mouse lung infections.
Microarray studies were used to identify a group of P. aeruginosa transcripts that were induced when the bacteria were grown in minimal media supplemented with lung surfactant (Survanta). When wild type PAO1 was exposed to minimal media containing lung surfactant compared to minimal medium with pyruvate alone, 125 transcripts (both predicted open reading frames (ORFs) and intergenic regions) were changed more than 3-fold (p<0.05), with 96 being induced (Table S1) and 29 being reduced (Table S2). Of the induced transcripts, 56 were characterized and 40 were predicted or hypothetical, while in the reduced transcript group, 11 were characterized and 18 were predicted or hypothetical. The induced class was dominated by genes from the Anr-regulon (29 genes, 16 of which were recently demonstrated as induced in surfactant [29]) and the choline catabolic pathway (15 genes) [30]–[32]. One observation of note in the induced group is the preponderance of transcripts encoding stress-related proteins including the chaperones hslU, groEL, dnaK, and dnaJ, and the universal stress protein family members sspK, PA1789, PA4352, and PA5027. We were interested in using the response to lung surfactant to identify gene function and novel biology in P. aeruginosa, and thus we have focused on highly induced genes of unknown function.
The PA5325 transcript was induced ∼18 fold in the presence of lung surfactant and was the most highly induced transcript in these experiments (Table S1). The PA5325 gene is divergently transcribed from PA5324, which encodes a probable AraC-family transcription factor that we hypothesized could be the transcriptional regulator of PA5325 (Fig. 1A). The robust induction of PA5325 in the presence of lung surfactant (Fig. 1B) suggested a possible role of this gene in the early stages of lung infection by P. aeruginosa.
For the following studies, we generated two reporter plasmids; pAL5 contained both sphR and the PA5325-lacZYA reporter and pAL4 contained only the PA5325-lacZYA reporter. Unless specified, the reporter used was pAL5 as it resulted in more robust induction. In addition to verifying the microarray results with surfactant, the PA5325-lacZ reporter was also induced in response to mouse fibroblasts (L-cells) and defibrinated sheep's blood (Fig. 1B). To determine which component of these eukaryotic-derived mixtures was a specific inducer of PA5325, we extracted mouse fibroblasts into aqueous and organic fractions and tested induction of PA5325-lacZ. The organic fraction contained the inducing activity, suggesting a lipid or other hydrophobic compound (Fig. 1C). Mouse fibroblasts and sheep's blood both contain high percentages of sphingomyelin [33], [34], and sphingomyelin makes up ∼4% of lung surfactant. Therefore, we tested induction of PA5325 by sphingomyelin and related sphingolipids including ceramide, S1P, and sphingosine. PA5325 was induced by sphingomyelin, ceramide, S1P, and sphingosine, but not the likely degradation products of sphingosine: palmitate and glycine (Fig. 1D). Other common lipid components of surfactant such as phosphatidylcholine and cholesterol did not induce transcription of PA5325-lacZ, and neither did unsaturated fatty acids (Supplemental Fig. S1). The strong induction by sphingosine compared to the other sphingolipids led us to hypothesize that the specific inducer of PA5325 is sphingosine. We tested the sensitivity of our PA5325-lacZ reporter to sphingosine (Fig. 1E), which demonstrated PA5325-lacZ induction in a dose-dependent manner and showed response to physiological levels of sphingosine, which range from 200 nM (as S1P) in the serum and lymph up to 2–13 mM of free sphingosine in the skin and some epithelial surfaces [35], [36]. We did not reach saturation in this assay due to a combination of sphingosine insolubility, plastic binding, and bactericidal effects (discussed below).
We hypothesized that the reduced induction of PA5325-lacZ in response to sphingomyelin, S1P, and ceramide compared to sphingosine was due to a processing step required by P. aeruginosa to yield sphingosine. To test this hypothesis we generated a clean deletion in the neutral ceramidase encoded by PA0845 and measured enzyme activity from the PA5325-lacZ reporter construct pAL4. Ceramide fails to induce PA5325-lacZ in the absence of the neutral ceramidase, whereas the response to sphingosine was unaffected (Fig. 1F). This finding strongly supports our hypothesis that induction of PA5325 occurs in response to sphingosine. In addition, PA5325-lacZ is induced in response to S1P in P. aeruginosa (Fig. 1D), but not significantly induced by S1P in E. coli (Fig. 1G), although the reporter in E. coli could still be induced in the presence of sphingosine (Fig. 1G). This suggested that P. aeruginosa may be processing S1P and that E. coli does not possess an orthologous activity under these conditions. When S1P was pretreated with shrimp alkaline phosphatase, induction of PA5325-lacZ was partially restored in E. coli (Fig. 1G). Given the transcriptional control of PA5325 in response to sphingosine, we have renamed it sphingosine regulated gene A, sphA.
PA5324 encodes a predicted AraC-family transcription factor divergently transcribed from sphA (Fig. 1A). This arrangement led us to suspect that PA5324 was the transcriptional regulator of sphA. To confirm the requirement of PA5324 for induction of sphA we generated an in-frame deletion of PA5324. The PA5324 deletion strain carrying our sphA-lacZ reporter construct (pAL4) showed no induction in the presence of sphingosine (Fig. 2A). Insertion of PA5324 onto the chromosome at the attTn7 site restored induction in the deletion strain (Fig. 2A). Furthermore, PA5324 was necessary to induce sphA-lacZ in a heterologous E. coli system in response to sphingosine (Fig. 2B). Our data suggest that the sphingosine responsiveness via sphA transcription is dependent on PA5324, and PA5324 was sufficient to confer sphingosine responsiveness in an E. coli system, therefore we have renamed PA5324 as the Sphingosine-responsive Regulator, SphR.
Our a priori prediction was that SphR, controlling expression of a strongly induced gene by pulmonary surfactant, would be important for colonization and/or survival in the mammalian lung. To test this hypothesis, we examined bacterial survival 24 hours after infection in the mouse lung. The sphR deletion strain had significantly lower survival than wild type (7.7-fold decrease, Dunnett's multiple comparisons p<0.001) and the survival defect was complemented by addition of sphR at the attTn7 site (Fig. 3). In this comparison, wild type and ΔsphR both contained the empty attTn7 insertion cassette on the chromosome. The contribution of sphR to survival in the mouse lung led us to a more in-depth study of SphR and its target genes.
Deletion of sphR resulted in reduced P. aeruginosa survival in the mouse lung (Fig. 3), leading us to hypothesize that one or more of the genes in the SphR regulon were likely candidates for this phenotype. To identify SphR-regulated genes in addition to sphA, we conducted microarray transcriptome analyses to compare wild type and the sphR deletion mutant in the presence and absence of pulmonary surfactant. Using a two-fold change cutoff and a p-value <0.05, there are six genes that differ between wild type and ΔsphR in the presence of surfactant (Table 1). Transcripts that are induced in wild type but not in the sphR deletion mutant include sphA, the neutral ceramidase (PA0845), and a three gene operon convergently transcribed toward sphA, PA5328-PA5326. The argB gene (PA5323) was induced more strongly in the sphR deletion than in wild type, which we think is likely due to a cis effect of the sphR (PA5324) deletion, as these genes are convergently transcribed (Fig. 1A). To denote their placement in the SphR regulon, we have renamed the genes in the predicted PA5328-PA5326 operon as sphBCD. The sphB gene encodes a predicted periplasmic cytochrome and sphC and sphD encode a predicted flavin-dependent oxidoreductase and a predicted pyridoxalphosphate-containing threonine aldolase-like enzyme, respectively. The predicted functions of SphC and SphD suggest a potential two-step pathway for sphingosine degradation to glycine and a long chain aldehyde by oxidation to an aldol and subsequent cleavage by the aldolase, a prediction we are currently exploring. The neutral ceramidase (PA0845) was previously designated PaCD [37], which does not conform to standard bacterial nomenclature. We propose that PA0845 be renamed cerN for ceramidase, neutral. The induction of sphA by surfactant in wild type (17.8-fold) versus the difference of sphA induction between wild type and ΔsphR (5.9-fold) suggested altered regulation of sphA in the absence of sphR (Table 1). The relative induction of sphA in the sphR mutant compared to wt under pyruvate (non-inducing) conditions supports a de-repression of sphA transcription in the absence of sphR at baseline. The remaining genes in the operon appear solely regulated by SphR under these conditions, as their induction levels in wild type compared to the difference between wild type and ΔsphR are not different.
We used promoter mapping to identify the promoter proximal regions of the sphA, sphBCD, and cerN promoters that were important for sphingosine and sphR-dependent regulation. Using lacZ reporter fusions to each upstream region, we identified a portion of each promoter-proximal region required for responsiveness to sphingosine (Fig. 4A). The regions required for sphingosine responsiveness were aligned using KALIGN [38], which produced an alignment that highlights the general format of an AraC-family binding site (Fig. 4B). The MEME consensus for a single half-site is shown below the alignment (Fig. 4B). Bioinformatic search of the P. aeruginosa genome (DNA Motif Search [39]) turned up only one additional predicted binding site (two direct repeats of the consensus (TGNCCSNNRNNSNCC) separated by 6–8 bp) in the genome apart from those present in the three identified promoters. The additional binding site is in the intergenic region between PA0428 and PA0429, upstream of the PA0428 gene. We did not detect any change in the PA0428 transcript for wild type or ΔsphR in the presence of surfactant or in either strain in the absence of surfactant. Therefore, based on our microarray data and bioinformatic analysis, we predict that sphA, sphBCD, and cerN likely comprise the core SphR regulon. The upstream sequences for the SphR regulon members showing the predicted SphR binding sites, promoter elements, and ribosome binding sites are shown in Supplemental Figure S2.
To test both specificity and the importance of conserved consensus sequences we mutated the first two residues in the consensus sequence TG to AA in half-site 1 (sphA**) (Fig. 4B), and tested the ability of the mutant sequence to permit induction of the reporter gene in response to sphingosine. The mutant reporter was unable to support reporter induction in response to sphingosine (Fig. 4C), demonstrating the importance of these conserved binding site residues.
We conducted electrophoretic mobility shift assays (EMSAs) with purified MBP-SphR fusion protein to test if SphR directly bound the sphA, sphBCD, and cerN promoters. The binding of MBP-SphR to the sphA promoter probe was greatly enhanced by the addition of sphingosine to the binding reaction in a concentration-dependent manner, providing evidence that sphingosine was a direct ligand of SphR (Fig. 5A). In the presence of sphingosine, MBP-SphR specifically shifted the sphA, sphBCD, and cerN promoters in a protein concentration-dependent manner and the binding could be competed with unlabeled sphA promoter probe, which gives a sense of the relative affinities for each binding site (Fig. 5B). MBP-SphR did not shift the non-specific plcH probe (Fig. 5B). The plcH probe is a useful negative control and demonstrates the specificity of SphR binding, as it has a known binding site for the AraC-family transcription factor GbdR in P. aeruginosa and its regulation is well described [31], [40]–[43].
To test the predicted SphR binding site, 59-mer oligonucleotides containing the proposed SphR binding site from the sphA promoter were annealed and the resultant probe was used in binding reactions. MBP-SphR was able to shift the 59-bp sphA probe (Fig. 5C, left), but only in the presence of sphingosine. Based on the inability of the mutated consensus sequence (sphA**) to support sphingosine-dependent reporter expression (Fig. 4C), we predicted that an oligonucleotide carrying these mutations would also be unable to bind SphR. As shown in the right side of Figure 5C, MBP-SphR was unable to bind this mutated probe. Together with the reporter fusions, these data support both the specificity of SphR binding and the importance of the conserved residues in the consensus.
Based on the enhancement of SphR DNA binding in the presence of sphingosine and our genetic evidence, we predicted that SphR would directly bind sphingosine. We used 3H-sphingosine to test the ability of SphR to bind sphingosine (Fig. 6). The binding assay conditions were similar to those used for EMSA studies with MBP-SphR in the presence of 3H-sphingosine. Amylose resin beads were used to pull down the MBP-SphR, and bead-associated sphingosine was assayed by liquid scintillation counting. 3H-sphingosine was substantially enriched in the fraction containing amylose-bound MBP-SphR, while relatively little remained associated with the amylose beads alone, or beads bound to a non-specific MBP-tagged P. aeruginosa AraC-family transcription factor, CdhR (MBP-CdhR) [44]. These data, in combination with the EMSAs (Fig. 5), demonstrate direct interaction between sphingosine and SphR.
Because deletion of sphR led to reduced survival in the mouse lung, we were interested in determining which of the SphR regulon members contributed to survival in the lung. We generated deletions in cerN, sphA, and sphC and compared to wild type in our 24 hour lung infection model. Deletion of sphA led to a significant reduction in bacterial survival in the mouse lung (9-fold decrease, Dunnett's multiple comparisons p<0.001), while deletion of cerN or sphC had no impact on bacterial survival in vivo (Fig. 7). The sphA mutant phenotype could be complemented by supplying the sphA under its native promoter control at the attTn7 site (Supplemental Fig. S3). These data suggest an important role for sphA in survival during infection. We did not test deletions of sphB and sphD in the animal model, given their predicted coordinate role with sphC in sphingosine metabolism and their similar phenotype to an sphC deletion during in vitro sphingosine killing (Fig. 8 and Figure S4).
Sphingosine has previously been shown to have antimicrobial properties and is able to inhibit growth and kill many Gram positive and Gram negative bacteria [19]. Previous studies suggest that P. aeruginosa is not sensitive to killing by sphingosine [45]. We hypothesized that SphR might play a role in the response of P. aeruginosa to sphingosine and could regulate sphingosine resistance. Using a modified sphingosine killing assay, we show that the ΔsphR deletion strain is more sensitive to sphingosine compared to wild type (Fig. 8), an effect that could be complemented by sphR on a plasmid (Supplemental Fig. S4). Most of the sensitivity of the ΔsphR strain appears to be due to loss of sphA induction, as the ΔsphA strain is also more sensitive to sphingosine than wild type and is nearly as sensitive as ΔsphR (Fig. 8). The deletion phenotype of sphA could be complemented by sphA on a plasmid (Supplemental Fig. S4). Deletion of sphC and transposon insertions into sphD and sphB also led to small but reproducible decreases in survival on sphingosine, suggesting a minor role for this operon in the response to sphingosine (Fig. 8 and Supplemental Fig. S4). Deletion of cerN, befitting its known function as an extracellular ceramidase, had no effect on survival in sphingosine (data not shown).
The induction of ceramidase activity in response to sphingosine has been demonstrated in a few bacteria [37], but the mechanism of sphingosine detection and conversion into a response had not previously been elucidated. In this study we show that sphingosine is directly detected by the AraC-family transcription factor SphR (PA5324) leading to the induction of sphA, sphBCD, and cerN transcripts. Deletion of sphR or sphA resulted in survival defects in a mouse model of acute pneumonia, suggesting that the ability to detect and respond to host-derived sphingolipids is important for survival in the lung. Sphingolipids are abundant in mammals, plants, and fungi, constituting a diverse family of molecules that serve as essential structural components of eukaryotic cell membranes and as dynamic signaling molecules that mediate diverse cellular functions [16]–[19]. In particular, S1P has been implicated as a critical component of mammalian innate and adaptive immune function, particularly in the acute phase response to pathogens [25]–[28]. Interestingly, orthologs of SphR and some of the SphR-regulon members are present in other opportunistic pathogens including Acinetobacter haemolyticus and Burkholderia pseudomallei, as well as the professional pathogen Mycobacterium tuberculosis.
Sphingolipids play important roles in host-pathogen interactions, particularly S1P and ceramide signaling [46]–[48]. In addition to host modulation of sphingolipid pathways to combat infection, pathogens can modulate host sphingolipids. M. tuberculosis alters sphingolipid signaling in macrophages by undetermined mechanisms [49], and S1P levels in the lungs of patients infected with M. tuberculosis are significantly decreased [50]. Interestingly, M. tuberculosis has an AraC-family transcription factor that is 47% similar along the whole length to SphR (RV1395) that was identified though signature-tagged mutagenesis where the RV1395 transposon mutant strain had an ∼1.5 log reduced survival in a mouse lung infection model [51]. Similarity between RV1395 and SphR is not restricted to the helix-turn-helix DNA-binding domain, as the two proteins are 44% similar when the DNA-binding domain is removed from the alignment analysis. RV1395 was characterized and found to be an activator of a divergently transcribed cytochrome gene, however the signals that govern RV1395 activation and its direct contribution to virulence have yet to be determined [52]. Based on the similarity of RV1395 to SphR we predict that a sphingolipid, perhaps sphingosine, may be the inducing ligand of RV1395.
The AraC-family transcription regulators are one of the largest groups of regulatory proteins in bacteria, and are often involved in the regulation of catabolism, stress response, and virulence [53]. Many members of the AraC family have been shown to respond to host-derived chemical signals present at the site of infection, but relatively few inducing ligands have been demonstrated to bind directly to their cognate regulator [54]. We found that addition of sphingosine altered the binding of SphR to the sphA promoter in EMSA studies and observed a dose response curve of SphR DNA binding at physiologically relevant concentrations of sphingosine. Bioinformatic analysis suggest similarity of SphR to ToxT (44% similarity and 20% identity), which directly regulates the major virulence factors in Vibrio cholerae. ToxT activation is inhibited by unsaturated fatty acids found in bile [55]. Subsequently, the crystal structure of ToxT was solved revealing a bound 16-carbon fatty acid that alters the structure of ToxT to prevent DNA binding in the presence of these bile associated fatty acids [56]. The similar size and hydrophobic nature of the regulatory ligands (palmitate vs. sphingosine) coupled with the sequence similarity allows us to speculate that SphR may bind sphingosine in a manner analogous to ToxT binding of palmitate.
Ito et al. identified a neutral ceramidase encoded by PA0845 (renamed cerN in this study) that was induced in the presence of sphingomyelin, ceramide and sphingosine, however the regulatory mechanism was not reported [37]. The discovery of SphR control of neutral ceramidase allows us to expand a model of bacterial utilization of sphingomyelin by linking it to our previous work on regulation of the phospholipase C/sphingomyelinase PlcH. We previously characterized the AraC-family regulator GbdR that is integral to a positive feedback loop controlling PlcH expression in response to a metabolite of the choline headgroup of sphingomyelin [31]. Sphingomyelin hydrolysis by PlcH yields ceramide [57], which P. aeruginosa can further metabolize through the action of ceramidases [54]. Here we show that CerN is produced as part of an SphR-dependent positive feedback loop in response to the ceramide metabolite sphingosine, in a manner analogous to GbdR control of PlcH. Both of these positive feedback loops link induction of secreted catabolic enzymes not to the availability of the substrate itself, but to metabolic products derived from the substrate. In each case, this ensures that the positive feedback loop will robustly operate only if the substrate is being metabolized at sufficient rates.
Sphingolipids such as sphingosine have long been known to have antimicrobial properties and sphingosine is found in high concentration in the skin where it is thought to be part of the barrier function against microbial infections [58]–[61]. A variety of Gram positive and Gram negative bacteria are sensitive to sphingosine, including Staphylococcus aureus and Escherichia coli [62]. The precise bactericidal mechanism of sphingosine remains unknown. However, recent evidence suggests that sphingosine may directly damage bacterial membranes [63]. P. aeruginosa has recently been reported to be resistant to the bactericidal effects of sphingosine [45]. While none of the deletion strains generated in this study showed growth defects under normal conditions, we found that both the sphR and sphA deletion strains were susceptible to the antimicrobial effects of sphingosine compared to wild type in vitro. Strains with deletions in sphR and sphA were also shown to have reduced survival in the mouse lung. We hypothesize that the sensitivity of sphA and sphR mutants to sphingosine contributes to their observed reduced survival in vivo. It is interesting to note that the double deletion ΔcerNΔsphA strain did not survive better or worse than ΔsphA, minimally suggesting that if the defect is due to sphingosine sensitivity, it is not sphingosine derived from P. aeruginosa hydrolysis of host-derived ceramide; in other words, they are not causing their own death by sphingosine derived from sphingomyelin and ceramide hydrolysis. Therefore, while the in vitro sphingosine killing correlates well with the in vivo phenotypes, we currently do not know the mechanism governing reduced survival of the sphR and sphA mutants in the lung.
We speculate that SphR responds to sphingosine to induce transcripts encoding proteins that protect P. aeruginosa from the bactericidal effects of sphingosine by induction of membrane stabilizing factors and/or catabolism of sphingosine to non-bactericidal metabolites. Here we show that SphR binds to sphingosine to initiate transcription of sphA, sphBCD and cerN. sphA encodes a hypothetical protein with some homology to proteins involved in meta-pathway phenol degradation. Protein localization predictions for SphA using the structure similarity-based prediction of Phrye2 [64] suggests that SphA is an outer membrane porin. Perhaps P. aeruginosa responds to sphingosine by providing a porin for sphingosine import and subsequent degradation that could aid in protecting the outer membrane from the damaging effects of free sphingosine. Okino and Ito demonstrated sphingosine utilization by P. aeruginosa by measuring removal of sphingosine from the culture supernatants and cell fractions [54]. Based on bioinformatic predictions, SphB, SphC and SphD are most likely involved in the metabolism of sphingosine. The sphB gene encodes a predicted cytochrome, while sphC encodes an FMN-linked oxidoreductase, and sphD encodes a pyridoxalphosphate serine-threonine aldolase. The latter two activities could work in concert to oxidize carbon 1, generating an aldol, which SphD could hypothetically act upon, rendering a long chain aldehyde and glycine. Transposon insertion into the sphC coding sequence (PA5327) resulted in reduced bacterial survival in a chronic rat lung infection model [65], suggesting that while our sphC deletion strain did not show a phenotype in the acute mouse lung infection (Fig. 7), it nonetheless impacts survival in the mammalian lung.
The microarray data comparing wild type in the presence and absence of pulmonary surfactant suggests some interesting biology in the presence of surfactant. The first observation has been covered by Jackson et al., who recently analyzed the changes in transcript levels of P. aeruginosa exposed to pulmonary surfactant, and compared wild type to both plcH and gbdR mutants [29], but did not publish results of these strains in the absence of pulmonary surfactant. They noted a reduction in transcript levels for Anr-controlled genes in both the gbdR and plcH mutants grown in surfactant, as do we (Table S1). Given the high levels of phosphatidylcholine and sphingomyelin in pulmonary surfactant, it was not surprising that the transcripts encoding proteins from the choline catabolic pathway were also highly induced in the presence of surfactant (Table S1). In addition to the high proportion of transcripts encoding stress-related proteins (mentioned in the Results section), there are also a high proportion (∼8%) of transcriptional regulators: NalC, BetI, NirG, PsrA, NarL, CgrA, PA3458, and PA4596. It is possible that the effects of induction of these transcription factors is contained in our regulation data, however our transcriptome analyses were a snapshot of transcripts at four hours post-induction and effects from changes in these transcription factors may not have sufficiently accumulated in the transcriptome. Of the reduced transcripts (Table S2), we note that three of the pyrroquinoline quinine biosynthesis genes are down, suggesting a change in requirement for this cofactor between surfactant and pyruvate conditions.
The demonstration of sphingosine detection by P. aeruginosa also opens up the possibility that this bacterium, and others with similar detection systems, could alter sphingosine and related sphingolipid signals, including S1P in the host. We have not yet examined the contribution of host immune signaling effected by the SphR regulon, but the impact of altering such an important and tightly controlled signaling network by bacterial factors has not been elucidated and may be an important contributing factor to the survival of P. aeruginosa in vivo.
This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol for animal infection was approved by the University of Vermont Institutional Animal Care and Use Committee (Permit number A3301-01). All procedures were performed under pentobarbital anesthesia and all efforts were made to minimize animal suffering.
P. aeruginosa PAO1, isogenic mutant strains, and E. coli (Table 2) were maintained in LB-Lennox (LB) medium. Morpholinepropanesulfonic acid (MOPS) medium [66] supplemented with 25 mM sodium pyruvate, 5 mM glucose and 50 µg/ml gentamicin (for P. aeruginosa) or MOPS with 10% LB (v/v), 5 mM glucose, and 10 µg/ml gentamicin (for E. coli) was used to grow strains prior to transcriptional induction studies. See LaBauve and Wargo (2012) for further details on P. aeruginosa growth methods [67]. For bactericidal assays, 1% neopeptone was supplemented with varying sphingosine concentrations in ethanol to reach a final concentration of 6.25% (w/v) ethanol in the assay. All lipids were purchased from Avanti Polar Lipids and other chemicals were purchased from Sigma-Aldrich or Fisher.
We used the oropharyngeal route of mouse lung infection previously described [68], [69]. Briefly, P. aeruginosa PAO1 and isogenic strains were streaked onto LB plates from −80°C stocks. Colonies from the first plate were restreaked onto a new LB plate after 24 hours and incubated at 37°C for 24 hours. Cells from the second plate were used to start 3 ml cultures in LB that were grown for 16–18 hours at 37°C on a roller drum. From these overnight cultures, cells were collected by centrifugation, washed in Dulbecco's PBS (DPBS), and resuspended to give ∼1×107 viable P. aeruginosa in 40 µL, with actual inoculum determined by serial dilution and plate counting. Eight to twelve week old male C57Bl/6J mice (Jackson Labs) were inoculated with 40 µL of the bacterial suspension via oropharyngeal aspiration. Anesthesia, surgery, bronchoalveolar lavage fluid (BALF) collection, organ harvest, and organ homogenization were done as previously described [68], [69] at 24 hours post-infection. Viable bacterial counts in organs were determined by serial dilution plating onto Pseudomonas Isolation Agar (PIA) (BD-Difco) followed by incubation at 37°C for 24 hours.
Mouse experiments (Fig. 3 and 7 and Supplemental Fig. S3) show CFU counts from all animals from duplicate experiments with each replicate having 4–6 animals per experimental group. All informative comparisons: mutants versus wild type (both Figures) and mutant versus complementation strain (Fig. 3 and Supplemental Fig. S3) were conducted in at least one additional experiment, included with comparator strains from other studies. Therefore, all informative comparisons were assessed three times. All experiments met the same statistical criteria, i.e. all replicates were consistent with regards to effect size and significance of changes. Inoculation order and harvest order alternated between experiments to eliminate potential issues related to the difference between the duration of inoculation (∼20–30 min) and the duration of harvest (∼1.5 h). For group comparisons, data (log10 transformed CFU counts) were analyzed by ANOVA followed by Tukey's (Fig. 3 and Supplemental Fig. S3) or Dunnett's (Fig. 7) Multiple Comparisons tests. All calculations were done using GraphPad Prism.
P. aeruginosa PAO1 wild type and ΔsphR were grown overnight in MOPS media supplemented with 20 mM pyruvate and 5 mM glucose. Overnight cultures were collected by centrifugation and resuspended in either MOPS supplemented with 20 mM pyruvate alone or 20 mM pyruvate and a 1∶50 dilution of the bovine surfactant preparation Survanta (Abbott) and induced for 4 hours at 37°C. Bacteria were collected by centrifugation, resuspended in MOPS and RNA Protect Bacterial Reagent (Qiagen), and the resultant pellets stored overnight at −20°C. RNA was extracted using an RNeasy kit (Qiagen), and eluted samples were treated with DNase I followed by a second round of RNeasy purification including an on-column DNase I treatment. Purified RNA samples were checked for DNA contamination by PCR and RNA integrity scores based on Agilent Bioanalyzer analysis were indicative of little to no DNA contamination.
Microarray analysis was performed on a Pseudomonas aeruginosa PAO1 gene chip using raw oligonucleotide probes generated from each condition using the NuGen Pico system. Each condition was analyzed in duplicate (N = 2), and summarized in one probe intensity by the Vermont Genetics Network Microarray Facility using Affymetrix GCOS software. Information from multiple probes was combined to obtain a single measure of expression for each probe set and sample. Probe-level intensities were background-corrected, normalized, and summarized, and Robust Multichip Average (RMA) statistics were calculated for each probe set and sample as is implemented in Partek Genomic Suites, version 6.6 (Copyright 2009, Partek Inc., St. Louis, MO, USA). Sample quality was assessed based on relative log expression (RLE), and normalized unscaled standard error (NUSE). To identify differentially expressed genes, linear modeling of sample groups was performed using ANOVA as implemented in Partek Genomic Suites. The magnitude of the response (fold change calculated using the least square mean) and the p-value associated with each probe set and binary comparison were calculated. The data have been submitted to NCBI GEO with accession number GSE48982.
Deletion mutants were generated using the pMQ30 plasmid [70] carrying the flanking regions of each of the four genes, sphR, sphA, sphC, and cerN, using conjugation-mediated deletion as described previously [30], [69]. Primers for these constructs are listed in Table S3. Single cross-over mutants were selected on PIA with gentamicin and selection of double crossover deletion mutants were carried out on LB 5% sucrose plates prepared without NaCl. Unmarked deletion mutants were verified using PCR. Complementation was done by integration of the sphR or sphA coding sequence under control of their native promoter at the attTn7 locus using the pUC18-miniTn7T-Gm vector as we described previously [68], [69] using the method of Choi and Schweizer [71]. This allowed stable complementation in the absence of antibiotic. For complementation where reporter plasmids were used, the gentamicin resistance cassette was excised by FLP-mediated recombination [71]. All sphR::attTn7 and sphA::attTn7 complementation strains were compared with wild type or mutant strains carrying the empty attTn7 integration region from the pUC18-miniTn7T-Gm vector.
Two reporter constructs were generated in this study using yeast homologous recombination [70] to generate translational fusions to lacZYA. A target lacZYA-containing vector suitable for yeast cloning (pMW42) was generated by excising the lacZYA region from pMW5 [31] with HindIII and EcoRI and cloning into the similarly cut pMQ80 backbone [70], which removes egfp-mut3. Either the sphA promoter (pAL4), or the entire sphR gene and the sphA promoter (pAL5) were recombined with pMW42 linearized with KpnI and HindIII. P. aeruginosa strains were electrotransformed with the reporter constructs and grown overnight in MOPS media supplemented with 20 mM pyruvate, 5 mM glucose, and 50 µg/ml gentamicin prior to induction. Inductions were carried out in MOPS media supplemented with 20 mM pyruvate and the inducing compound and incubated at 37°C for 6 hours. β-galactosidase assays were done as previously described [31], [72], using the method of Miller [73]. Studies of heterologous sphA induction in E. coli were carried by transforming pAL4 and pAL5 into E. coli NEB5α. Resulting E. coli strains were grown overnight in MOPS media supplemented with 10% LB (v/v), 5 mM glucose and 10 µg/ml gentamicin. For induction assays with S1P in E. coli, 2.4 µg of S1P or sphingosine were pre-treated with or without 5 U shrimp alkaline phosphatase (SAP) in 100 µL of water with 1× SAP buffer (USB), and incubated at 37°C for 60 minutes. Induction assays were carried out in MOPS supplemented with 10% LB (v/v), treated inducing compounds, and 10 µg/ml gentamicin. All E. coli strains were induced for 8 hours prior to ß-galactosidase assays.
Full-length reporter constructs and truncations of sphA, sphB, and cerN promoters were cloned into pMW5 [31]. The resultant lacZYA reporter constructs were transformed into wild type P. aeruginosa and used to identify the region required for response to sphingosine. Inductions were carried out in MOPS media supplemented with 20 mM pyruvate and 150 µM sphingosine and incubated at 37°C for 6 hours followed by ß-galactosidase assays.
We constructed a maltose binding protein (MBP) fusion to SphR by using the pMALc2 vector system (NEB). The sphR gene was amplified from genomic DNA. The PCR product was gel purified and ligated into the pCR Blunt vector (Invitrogen). The insert was excised with KpnI and HindIII, gel purified, and ligated into a similarly digested pMALc2 vector to generate pAL11. E. coli NEB5α (New England Biolabs) carrying the pAL11 plasmid were grown overnight in LB supplemented with 120 µg/ml carbenicillin. The overnight culture was transferred to two 500 ml flasks containing 100 ml of LB-carbenicillin and shaken at 220 rpm for 5 hours. Isopropyl-β-D-thiogalactopyranoside (IPTG) was added to a final concentration of 1 mM, and the cells were induced for 3 hours. Cells were collected by centrifugation, lysed in column buffer (20 mM Tris-HCl, pH 7.5, 150 mM NaCl, 1 mM EDTA) supplemented with 3 mg/ml lysozyme and Halt protease inhibitor 1× cocktail (Thermo Scientific). Lysates were clarified by centrifugation, and the soluble fraction was applied to a column containing amylose resin (NEB). The column was washed with ten volumes of column wash buffer (20 mM Tris-HCl, 150 mM NaCl 1 mM EDTA pH 7.4), followed by elution with column wash buffer supplemented with 10 mM maltose. Elution fractions were run on 10% SDS-PAGE gels and visualized by Coomassie staining. Fractions containing the MBP-SphR were pooled and dialyzed against 20 mM Tris-HCl, pH 7.5 at 4°C in a 20,000 kDa cutoff Slide-A-lyzer cassette (Pierce). The full length MBP-SphR fusion protein was used in electrophoretic mobility shift assays, as the MBP tag did not prevent sequence specific DNA binding (Fig. 5) or binding to sphingosine (Fig. 6).
EMSA DNA probes were generated using PCR (Primers in Table S3) and were spot dialyzed against 2.5 mM Tris-HCl, 0.25 mM EDTA, pH 8.0. Labeled probes, generated using a primer with a covalently linked 5′ biotin tag (IDT), were used at 0.5 fmol/µl, and unlabeled competitor probes were used at a final concentration of 0.5 pmol/µl. EMSA was carried out using a Thermo Scientific Thermoshift kit. The final binding buffer was modified to contain 1× binding buffer (10 mM Tris-HCl, pH 7.5, 50 mM KCl, 1 mM dithiothreitol), 0.1 mM glycine betaine, and 2 µg/ml poly-dI-dC. Various concentrations of sphingosine dissolved in ethanol were added to reaction tubes and allowed to dry to eliminate ethanol prior to binding reactions. Binding reactions were carried out at 37°C for 15 minutes and electrophoresed on a 5% non-denaturing polyacrylamide gel then transferred to a BioDyne B membrane (Thermo Scientific). Detection was carried out using streptavidin-linked horseradish peroxidase according to the supplied protocol (Thermo Scientific).
Sphingosine association with SphR was measured by conducting binding reactions using 3H-D-erytho-sphingosine (Perkin-Elmer). Binding reactions were carried out as described for EMSA except 3H-D-erytho-sphingosine was used at a final concentration of 50 nM. Samples were incubated with and without either 10 µM MBP-SphR or 10 µM MBP-CdhR for 30 minutes then added to amylose resin. The amylose beads were collected by centrifugation and washed 3 times with amylose column wash buffer. After washes, amylose beads were resuspended in 200 µl of amylose wash buffer and transferred to a glass vial containing 10 ml of Biosafe II scintillation cocktail (RPI). Samples were quantified using a Tri-Carb 2910 TR liquid scintillation analyzer (Perkin-Elmer).
Killing assays were carried out as previously described [61]. Briefly, overnight P. aeruginosa strains were grown in trypticase soy broth (TSB) and diluted 1∶40. Diluted cultures (100 µl) were added to glass tubes containing 250 µl of 1% neopeptone supplemented with 50 µl of the appropriate sphingosine stock in ethanol or ethanol alone as the vehicle control. The cultures were shaken at 170 rpm for one hour. Survival was determined by serial dilution plating on PIA. Colonies were counted after 24 hour incubation and survival calculated by comparison to vehicle only controls.
|
10.1371/journal.pgen.1004006 | Multi-tissue Analysis of Co-expression Networks by Higher-Order Generalized Singular Value Decomposition Identifies Functionally Coherent Transcriptional Modules | Recent high-throughput efforts such as ENCODE have generated a large body of genome-scale transcriptional data in multiple conditions (e.g., cell-types and disease states). Leveraging these data is especially important for network-based approaches to human disease, for instance to identify coherent transcriptional modules (subnetworks) that can inform functional disease mechanisms and pathological pathways. Yet, genome-scale network analysis across conditions is significantly hampered by the paucity of robust and computationally-efficient methods. Building on the Higher-Order Generalized Singular Value Decomposition, we introduce a new algorithmic approach for efficient, parameter-free and reproducible identification of network-modules simultaneously across multiple conditions. Our method can accommodate weighted (and unweighted) networks of any size and can similarly use co-expression or raw gene expression input data, without hinging upon the definition and stability of the correlation used to assess gene co-expression. In simulation studies, we demonstrated distinctive advantages of our method over existing methods, which was able to recover accurately both common and condition-specific network-modules without entailing ad-hoc input parameters as required by other approaches. We applied our method to genome-scale and multi-tissue transcriptomic datasets from rats (microarray-based) and humans (mRNA-sequencing-based) and identified several common and tissue-specific subnetworks with functional significance, which were not detected by other methods. In humans we recapitulated the crosstalk between cell-cycle progression and cell-extracellular matrix interactions processes in ventricular zones during neocortex expansion and further, we uncovered pathways related to development of later cognitive functions in the cortical plate of the developing brain which were previously unappreciated. Analyses of seven rat tissues identified a multi-tissue subnetwork of co-expressed heat shock protein (Hsp) and cardiomyopathy genes (Bag3, Cryab, Kras, Emd, Plec), which was significantly replicated using separate failing heart and liver gene expression datasets in humans, thus revealing a conserved functional role for Hsp genes in cardiovascular disease.
| Complex biological interactions and processes can be modelled as networks, for instance metabolic pathways or protein-protein interactions. The growing availability of large high-throughput data in several experimental conditions now permits the full-scale analysis of biological interactions and processes. However, no reliable and computationally efficient methods for simultaneous analysis of multiple large-scale interaction datasets (networks) have been developed to date. To overcome this shortcoming, we have developed a new computational framework that is parameter-free, computationally efficient and highly reliable. We showed how these distinctive properties make it a useful tool for real genomic data exploration and analyses. Indeed, in extensive simulation studies and real-data analyses we have demonstrated that our method outperformed existing approaches in terms of efficiency and, most importantly, reproducibility of the results. Beyond the computational advantages, we illustrated how our method can be effectively applied to leverage the vast stream of genome-scale transcriptional data that has risen exponentially over the last years. In contrast with existing approaches, using our method we were able to identify and replicate multi-tissue gene co-expression networks that were associated with specific functional processes relevant to phenotypic variation and disease in rats and humans.
| The increasingly cheaper and rapid accumulation of large -omics datasets across several experimental conditions has prompted generation of a wealth of data on biological networks. This growth of network data now permits their large scale applications to biomedical research, including analysis of gene function, metabolic and signaling pathways, as well as disease-related or cell function-related networks [1], [2]. However, reconstructing and interpreting large biological networks, such as co-expression networks, protein-protein interaction networks or genetic networks, with different features (e.g., sparse or densely interconnected, etc.) poses many challenges, advocating efficient and flexible methods for network inference and pattern discovery. An important level of complexity in current network analysis regards its extension to multiple conditions, for instance different species [3], cell-types [4] or disease states [5], [6]. For example, reconstruction of networks across multiple disease-states is becoming a useful approach for efficient drug-target discovery, as networks can inform the “biological context” (e.g., pathways, cellular processes) where genes operate and therefore can help designing better therapeutic interventions [7]. In genetic studies of complex diseases researchers increasingly focus on groups of highly interconnected genes within larger networks (referred to as clusters, modules or subnetworks) to elucidate specific cellular and molecular processes that might represent functional disease mechanisms and pathological pathways [8]–[10].
While several computational tools for network analysis in single datasets or conditions are available, only few computationally efficient methods for genome-scale network analysis across multiple conditions have been developed to date. These methods can be broadly classified into two main categories: (i) methods to find the “difference” between networks across conditions or to pinpoint condition-specific networks [11]–[14], or (ii) methods to identify the common parts in networks across conditions [15]–[17]. More recently, tensor-based computational frameworks [15] or probabilistic Markov blanket search algorithms [18] have been proposed to learn network structures across conditions. However, these methods are either heavily influenced by the choice of input parameters (e.g., number of clusters, number of nodes within a cluster, cluster interconnectivity) [15] or, being based on probabilistic graphical modelling, they become prohibitively slow for high number of conditions since they are trying to learn the structure of large graphs [18].
Complementary to the above approaches, spectral methods, such as Singular Value Decomposition (SVD), have been also proposed to investigate patterns of connectivity between nodes within a single network [19], [20] or for comparing two networks [21]. Generally, any network can be described as a graph, which is denoted as comprising a set of vertices or nodes together with a set of edges [22]. The graph may be represented by a square, symmetric, real-valued matrix of size whose entries denote the relationship between the corresponding nodes. In the affinity matrix , the element , called weight, represents the strength of connection between vertices and . For instance, in gene regulatory (or co-expression) networks, the nodes might represent genes (or mRNAs expression) and edges represent the strength of gene-gene interactions (or mRNAs co-expression).
Generalized Singular Value Decomposition (GSVD) can be used to identify sub-network structures and for comparative analysis of genomic datasets across two conditions [11], [23]. Given two matrices and [24], [25], their GSVD is given by(1)where and have orthonormal columns, is invertible, with , with . The ratios are the generalized singular values of and . In this setup, the common factor is informative of the cluster structure shared across the two data matrices.
Recently, a novel mathematical formulation, higher-order GSVD (HO GSVD), which is constructed for more than two data matrices has been proposed [26]. Under this framework, the matrices , each with full column rank (i.e., the maximum number of linearly independent column vectors of is ), are decomposed as(2)where is composed of normalized left basis vectors, with and the latent factor matrix is composed of normalized right basis vectors. The HO GSVD can be also derived in the special case of square, symmetric, full rank affinity matrices, , where each element represents the weight of the edge between node and in the th condition. It has been previously employed to compare multiple datasets with identical column size in order to detect their common substructures of columns (i.e., observations) [26]. Yet, another useful application of the HO GSVD to genomics is to set it to discover gene networks across multiple conditions and pinpoint “common” and “differential” cluster structures.
In this paper, we build on the flexible HO GSVD mathematical framework and propose a new, parameter-free computational algorithm (Cross-Conditions Cluster Detection or C3D) for automatic detection of both similarity and dissimilarity clustering patterns in large weighted (and unweighted) networks across several conditions (). The original HO GSVD model has been employed for analysis of datasets that had varying number of genes (), the same number of observations () (i.e., arrays/time points in [26]) across conditions and with . As such, this illustrative application of the HO GSVD in genomics was aimed at the identification of common structures within the observations [26]. Here, we built on the initial HO GSVD to extract sub-structures (i.e., common and differential clusters) from genes across multiple conditions () by applying the decomposition to the transposed expression matrix . We show how this enables a more general application of the HO GSVD framework to genome-scale network analysis of genomic data (e.g., microarray, RNA-seq) in multiple conditions. Besides, a distinctive feature of our method is in its capability to take as an input either the raw expression matrices or co-expression matrices, allowing flexibility in the choice of the co-expression measures (e.g., Spearman, Kendall, mutual information, etc.).
Figure 1 illustrates the working principle of the C3D algorithm. The input data for C3D can be provided into different formats to be used by the HO GSVD: (i) the raw expression data matrices () or (ii) the co-expression data matrices (). In the former case, a first data initialization step is conducted where the input expression matrices, with the same number of genes are converted to co-expression matrices by scaling their variance to 1 and taking their quadratic form. In the second step (HO GSVD-based algorithm), an approximate HO GSVD is employed to identify a common basis , with representing the dimension of the GSVD common subspace, for the decomposition of the input datasets and identify the common and differential correlation structures. The HO GSVD-based algorithm computes a square matrix , which is built on the arithmetic mean of all pairwise quotients where denotes the Moore-Penrose inverse of the co-expression matrix [24] (see Methods section). The first eigenvectors of (according to the norm of the corresponding eigenvalues) are then used to identify an approximate decomposition of the input co-expression matrices and form the decomposition basis . Specifically, each selected column vector of is used to reorder the input data matrices such that candidate “common” (or “differential”) clusters can be identified. In the third step (cluster nodes selection and validation), we employ a mixture model approach to classify genes and assign them to each cluster based on a misclassification error rate (MER). Finally, we implemented an empirical cluster validation procedure to identify the conditions where clusters are present and assess the level of significance for clusters within each condition.
To demonstrate the increased power and benefits of our HO GSVD-based algorithm, we carried out an extensive simulation study and benchmarked C3D against commonly used methods that were designed to detect either common (WGCNA [16], [17]) or differential network structures (DiffCoEx [13]) across multiple conditions. We show that our approach has higher power and stability in detecting both common and differential co-expression clusters across all simulated conditions, while being two to seven fold less computationally intensive than alternative methods. In contrast with alternative approaches that require specification of ad-hoc input parameters, the proposed method has the distinctive advantage of being parameter-free, which makes it a powerful tool for real data exploration and analysis. To substantiate this claim, we applied C3D to publicly available transcriptomic datasets in rats and humans and identified several multi-tissue gene co-expression networks that were associated with specific functional processes relevant to phenotypic variation and disease.
We carried out a simulation study to compare our method with commonly used approaches for identification of “common” or “differential” clusters across multiple networks: (1) WGCNA and (2) DiffCoEx. The WGCNA method for detection of common clusters across co-expression networks employs a “soft” threshold to assign a connection weight to each gene pair and extract densely connected gene clusters that are present in all conditions. The DiffCoEx method follows a strategy similar to WGCNA but, instead, it focuses on detecting the differences in co-expression patterns (“differential” clusters) between multiple conditions. Additional details on the specific parameterizations used in for WGCNA and DiffCoEx analyzes are reported in Text S1.
To simulate a realistic example of gene expression data from multiple conditions that represent a typical “small large ” scenario, we draw inspiration from a publicly available multi-tissue microarray dataset consisting of genome-wide expression profiles from recombinant inbred rat strains in seven tissues [27]. We simulated different types of clusters that are either detected in all conditions (“common” clusters) or are specific to a subset of conditions (“differential” clusters), Figure 2. We considered dense clusters of variable sizes (100–500 nodes) where each node is connected with all other nodes in the cluster with a given weight (), which is defined as the Pearson correlation between expression profiles of genes and . We simulated clusters with varying cluster densities (0.1, 0.3, 0.5, 0.7), which were defined as the average Pearson correlation between any pair of nodes within a cluster. In addition to the simple case of a cluster common to all conditions and with the same size (Cluster pattern 1), we set out to evaluate the sensitivity of our and alternative approaches to detect clusters which are present only in a subset of conditions and that overlap partially across conditions. This is more likely to be relevant for analysis of pathways and gene networks across tissues or during development, where varying gene-sets can exert their function only at specific developmental times or in specific cell-types. To account for these more complex scenarios, we simulated “nested” (Cluster pattern 2) and partially “overlapping” (Cluster pattern 3) cluster structures (Figure 2). Cluster pattern 2 and Cluster pattern 3 have an intersection part, defined by the nodes in common to all conditions, and a union part, defined by the nodes in common to all conditions plus the nodes present in individual conditions. In summary, for each of the four cluster densities considered one dataset consisted of a and matrix in conditions, where each cluster type (Clusters patterns 1–3) was simultaneously present in the data matrix. To assess reliability of the results, for each of these data we generated 20 independent replicates, yielding a total of 560 simulated datasets. Similarly, to evaluate how the number of available observations affects the methods' performance we simulated datasets consisting of a and matrix in conditions (20 replicates, 560 datasets in total). See Text S1 for additional details.
The True Positive Rate (TPR) and the False Positive Rate (FPR) are widely used as evaluation metrics for a classification model and can be used to quantitatively assess (and compare) methods performance [28]. The TPR defines how many correct positive results (simulated clusters genes within the called cluster) occur among all results called positive in the analysis by a given method. FPR, on the other hand, defines how many incorrect positive results occur among all results called positives. Typically, a and the corresponding indicate a perfect classifier (or a perfect method). In our simulation study, the best cluster detection method would yield both high TPR and low FPR levels for different cluster types, sizes and densities.
For each simulated cluster type, Figure 3 shows the TP/FP rates for C3D, WGCNA and DiffCoEx methods as a function of the simulated cluster densities. For C3D we controlled the (local) misclassification error (i.e., the probability to assign wrongly a gene to a cluster) to be less than 0.05 or less than 0.2, and required that each cluster is detected with , whereas for WGCNA and DiffCoEx we used two (default) parameterizations chosen according to the software guidelines (see Methods section). The C3D method outperformed WGCNA in the identification of clusters present in all conditions (Cluster pattern 1, Figure 3), and showed to have consistently high TPR (and very low FPR, ) irrespective of the simulated cluster density. WGCNA performance varied considerably as a function of the simulated cluster density and, depending on the adopted parameterization, FPR levels were (reaching 20% in one case), Figure 3. Furthermore, we observed large variations in WGCNA performance (mostly in the TPR), which are indicated by the large standard deviations in TPRs calculated from the 20 replicated datasets. For more complicated patterns (“nested” and “overlapping” clusters), we compared C3D with WGCNA to detect the intersection part (100 nodes) of common clusters. Since WGCNA is designed to detect only those clusters shared across all conditions, for clusters present in a subset of conditions, we run WGCNA only in the set of conditions where the simulated clusters were present. For Cluster patterns 2–3, C3D and WGCNA performances were similar, reaching high TPR for detection of the intersection part of clusters with simulated (Figure 3). However, C3D showed higher TPRs than WGCNA to detect clusters with low densities (0.1–0.3), while controlling the FPR at low levels (, Cluster pattern 2 intersection).
In the case of partially overlapping clusters present in a subset of conditions (Cluster patterns 2–3) we compared C3D with DiffCoEx in respect of detecting the union part (500 nodes) of “differential” clusters, and calculated TPR and FPR for detection of this cluster (indicated with a black square at the top of Figure 3). We found that C3D outperformed DiffCoEx across the simulated scenarios. In the case of the “nested” cluster structures that are present in 5 out of 7 conditions, C3D had consistently higher TPR levels than DiffCoEx, which showed comparable TPR levels only for detection of highly-dense clusters (i.e., , Cluster pattern 2 union, Figure 3). However, similarly to what observed for WGCNA method, in this case DiffCoEx showed large variability in its performance across the 20 replicated datasets. The difference in performance between C3D and DiffCoEx was observed also in the more complicated case of partially overlapping cluster structures (Cluster pattern 3). In this case, C3D showed consistently higher TPR than DiffCoEx that reached a maximum as compared with of C3D. Both methods showed comparably low FPR () for detection of the union part of Cluster patterns 2–3 (Figure 3). Similarly to what observed for the simulated data with observations, C3D performed better than (or as good as) both WGCNA and DiffCoEx when benchmarked on simulated datasets with only observations (Figure S1). As expected, all methods had lower TPRs associated with the detection of low-density clusters, however also with a small number of observations, C3D showed significantly better (and more stable) results than WGCNA and similar performance as compared with DiffCoEx. Notably, for detection of “common” clusters present in all conditions (Cluster pattern 1), CD3 held high TPR levels (and FPR) whereas WGCNA's performance dropped significantly, reaching a maximum TPR (Figure S1).
These data show that C3D on balance performed better than WGCNA and DiffCoEx across all simulated scenarios. We underline that while WGCNA and DiffCoEx methods are specifically designed to detect either common or differential clusters, respectively, here we showed that C3D was equally or more accurate than both methods in the detection of common and differential cluster structures. We also highlight how C3D ability to detect correctly the simulated clusters was highly consistent across all runs on the replicated datasets, as shown by the small standard deviations of the mean TP and FP estimates (Figure 3). In contrast, we observed that both WGCNA and DiffCoEx performances varied appreciably across the replicated simulations, often resulting in large standard deviations of the mean TP and FP estimates. To better assess the reliability of the different methods we calculated the relative standard deviation of the TPR measured in all analyzed datasets. In 560 simulated datasets of size , the C3D method had a median RSD of (range 113.36) whereas WGCNA and DiffCoEx have median (range 447.2) and median (range 133.39), respectively. Similarly, in 560 datasets of size we estimated the following RSDs of TPR: 12.43 (range 113.38) for C3D, 57.52 (range 161.89) for WGCNA and 87.96 (range 120.59) for DiffCoEx. The large RSDs of TPR calculated from the WGCNA and DiffCoEx analyzes originated because these methods often detected the simulated cluster(s) only in small number of replicates (e.g., 2 out of 20).
Besides, in a few cases the TP/FP rates of WGCNA and DiffCoEx were influenced by the adopted parameterization (for instance, FPR in the WGCNA analysis of Cluster pattern 1, Figure 3), suggesting that different choices of the input parameters can affect the detection of clusters (see Text S1 for additional details). The C3D algorithm is built on the HO-GSVD framework and as such does not require the user to specify ad-hoc parameters to detect common or differential clusters. In our implementation of the C3D algorithm the user can control the MER at a specified level before the cluster genes are empirically validated using a permutation-based procedure (see Methods section). In these simulation studies, we have used two different MERs (5% and 20%) to inform a suitable choice of MER that maximizes true positive without inflating false positive rates. On average, we observed a increase in the TPR when was adopted as compared with . However, we found no significantly higher FPR, which were always across all simulated datasets, this suggesting that using the less stringent in real data analyzes is likely to increase the detection of true gene clusters, without increasing significantly false positives.
Finally, we used a standard desktop computer (Mac Pro, GHz Quad-core Intel Xeon with 20 Gb RAM) to evaluate the computational time required by C3D and compare it with WGCNA and DiffCoEx to analyze the simulated datasets. While the run time of C3D scales exponentially with the number of genes in the input matrices or the number of conditions, our Matlab implementation of C3D is relatively fast and requires only 1,200s to analyze a gene co-expression matrix in conditions and 10s to analyze a gene co-expression matrix in conditions (Figure S2). When compared with competing approaches, we assessed that to process simulated datasets of 1,000 and 10,000 genes (with observations and conditions) C3D requires significantly smaller CPU time than DiffCoEx (up to 2.3 fold more CPU time) and WGCNA (up to 8.2 fold more CPU time), respectively (Figure S2).
To show how C3D provides a powerful, practical framework for real genome-scale analyzes and yields new biological insights into pathways and molecular networks, we report an application to two large multi-tissue gene expression datasets in rats and humans. Transcriptional profiling was carried out by Affymetrix microarray in the rat and mRNA sequencing (RNA-seq) in humans, respectively. The microarray dataset consisted of genome-wide expression profiles ( probe sets) that were measured in seven tissues (adrenal, aorta, fat, kidney, left ventricle, liver and skeletal muscle) in a panel of recombinant inbred rat strains [29], which is a well characterized model of hypertension, metabolic syndrome and cardiovascular disease [27], [30], [31]. The RNA-seq datasets consisted of genome-wide transcriptomic data of human fetal neocortex, which have been generated to investigate the molecular mechanisms underlying differences in germinal zones of the developing human brain. The human dataset consisted of expressed genes which were analyzed in four regions of the fetal neocortex (ventricular zone (VZ), inner subventricular zone (ISVZ), outer subventricular zone (OSVZ) and cortical plate (CP)) from six 13–16 weeks postconception human fetuses [32]. In both rat and human analyzes, to identify common and differential clusters we extracted the top ten eigenvectors (based on the modulus of the eigenvalues of the decomposition of ) as candidates which are then used as input for the cluster nodes selection and validation step of the C3D algorithm (see Methods).
Building on the HO GSVD framework, we have developed a new algorithm (C3D) for efficient, parameter-free and automatic detection of co-expression clusters and networks in multiple conditions. Our method is designed for analysis of weighted (and unweighted) networks (input matrices) across conditions, enabling applications to diverse data types and structures. Although the original HO GSVD algorithm assumes the non-singularity of the co-expression matrix , by using the Moore-Penrose pseudo-inverse, our C3D algorithm can be applied to the non-invertible case. We show that when an exact HO-GSVD of the input matrices exists (as defined in (4), see Methods), our HO GSVD is able to extract the right decomposition basis through the eigen-decomposition of , whereas it finds an approximate decomposition of the data in the absence of an exact solution (Figure S4). In particular, our empirical simulations and real-case applications reveal that our approximate decomposition is able to capture both common and differential co-expression structures for a wide range of noise levels, suggesting that our algorithm can be useful for practical applications to genomic data.
Here, through the HO GSVD of large-scale genomic datasets we aimed to uncover the complex interactions between genes (networks) that can occur within or across multiple conditions. One distinctive feature of our computational method is in the flexible and simultaneous identification of both “common” and “differential” sub-network structures across several conditions. Selecting informative vectors of , we provide different orderings of to reveal candidate clusters that are important to all conditions or specific to a sub-set of conditions; then, we can distinguish the specific conditions where the clusters are present using a permutation-based approach. This procedure allows to pinpoint automatically the specific conditions where the sub-network structures are present and, at the same time, to provide an empirical estimate of the statistical significance (empirical P-value) for each cluster identified.
In simulation studies, we demonstrated how C3D outperforms competing approaches in accuracy and reliability while being computationally less demanding. We highlight how our method allowed accurate detection of clusters within complex structures (i.e., “common”, “nested” and “overlapping” networks) by specifying only the desired level of statistical significance: misclassification error rate to assign genes to clusters and empirical P-value for cluster detection. In contrast with other approaches, C3D does not need the user to specify ad-hoc parameters related to the expected number of clusters or cluster density [15] or necessary to determine the optimal height cut-off in the gene clustering tree [13], [16], [17]. Typically, these unknown parameters need to be “finely tuned” on each dataset in order to obtain the best compromise between TP and FP for each cluster (see Text S1 for additional details). We also showed that the results obtained by two competing and widely-used methods (WGCNA and DiffCoEx) were less stable than those provided by C3D. This was apparent in the significantly smaller relative standard deviations in TPR calculated across simulated datasets in the C3D analyzes as compared with WGCNA and DiffCoEx. Since C3D utilised raw gene expression data matrices as input, the higher stability of C3D might be due to the reduced influence of the small number of observations on the stability of co-expression estimates, which can result in extreme patterns of correlation changes, corresponding to stable and fragile co-expression, as previously shown [62].
The high stability in the results and the parameter-free “nature” of the HO GSVD approach make the C3D algorithm a powerful computational tool for real genomic data exploration and analysis. To demonstrate this point, we reported an application of C3D to two large transcriptional datasets: (i) microarray-based gene expression profiles in seven rat tissues and (ii) RNA-seq-based gene expression analysis of germinal zones from human fetal neocortex. In the rat analysis, we reported several functionally enriched co-expression clusters, including a previously identified inflammatory gene network driven by the IRF7 transcription factor that represents a gene expression signature of macrophages within complex tissues. While this co-expression network was experimentally validated [27] it was not recovered by WGCNA, that surprisingly placed the IRF7 transcription factor and many regulated target genes in the group of “non-clustered” genes. In addition, our C3D analyzes revealed novel gene co-expression networks in sub-sets of tissues. For instance, we identified a network comprising Hsp and known cardiomyopathy genes, which suggested coordinated regulation of heat shock proteins genes in multiple tissues, and their potential functional role in cardiovascular disease [50]. While this network was not recovered by either WGCNA or DiffCoEx analyzes, we were able to replicate this new finding using separate cardiac and liver gene expression datasets in humans (Figure 4). In the study of human fetal neocortex we demonstrated previously undescribed co-expression between cell cycle and ECM-receptor interaction pathways and support their role in the proliferation and self-renewal of neural progenitors. In addition, our analyzes highlighted that pathways central to later cognitive functions (e.g., calcium signaling, long-term potentiation, axon guidance) are present at an early stage in the developing human brain [61], which was not previously appreciated. These studies illustrated how our method can be effectively applied to leverage the vast stream of genome-scale transcriptional data that has risen exponentially over the last years, promising to aid the fine-scale characterization of both context-specific and systems-level networks and pathways.
We describe a new computational method (Cross-Conditions Cluster Detection or C3D) to detect both similarity and dissimilarity clustering patterns in weighted networks across multiple conditions (). After a data initialization step, C3D employs HO GSVD-based algorithm and cluster nodes selection and validation procedures to identify clusters, the specific conditions where the clusters are detected and the statistical significance of the clusters, as summarized in Figure 1 and detailed below.
In this step we assume the input data are non-square matrices , where the rows represent the observations and the columns indicate genes. The number of genes must be the same across datasets while the number of observations can differ. We first log transform the data and subtract for each gene its average gene expression to avoid capturing differences in average gene expression across conditions. We then calculate the co-expression matrices corresponding to each condition . Each represents the covariance matrix of the data in condition . As in classic principal component analysis, the columns of can be scaled to unit variance to work on the correlation matrices rather than the covariance. Alternatively, our algorithm can directly take any co-expression matrix as input. This feature of our algorithm allows to extract common and differential clusters from matrices based on different co-expression measures, including robust correlation (e.g. Spearman, Kendall) and non linear metrics such as mutual information [63].
Similarly to classic SVD, each observation from the input data can be characterized by its expression profile and represented by a data point in a dimensional space. The observations from all datasets are contained in a subspace of dimension , which thereafter is referred to as the HO GSVD subspace. Here, we aim at finding directions in the HO GSVD subspace that either capture the variability in gene expression that is common to all conditions (common factors) or that is specific to a subset of conditions (differential factors). Inspired by [26] we developed a general algorithm that allows computation of an approximate solution to the HO GSVD problem in the non full column rank case. In the HO GSVD, are decomposed into where , is a diagonal matrix with elements for and contain the right basis vectors of the HO GSVD subspace where . The right basis vectors allow to identify set of genes (clusters) with similar co-expression patterns, that are either specific to a subset of conditions or common to all conditions. Here we explain the derivation of our HO GSVD-based algorithm in the general case of non-square matrices. The derivation and discussion of the special cases ( square, symmetric matrices with full rank and square, symmetric matrices with full rank) is reported in Text S1. In the most general case, we define the right basis vectors as the solution of the eigen-decomposition problem of the matrix(3)where is the arithmetic mean of all the pairwise quotients and denotes the Moore-Penrose inverse of the co-expression matrix [24]. Here the Moore-Penrose inverse is used as a substitute of since the invertibility of is not guaranteed when , which is the typical scenario in genomics. We now assume there is an approximate HO GSVD where is composed of orthonormal left basis vectors and . In this case, for all we have(4)and its Moore-Penrose inverse is given by(5)Therefore we have(6)since is full row rank. Hence we can rewrite as follows(7)When there exists a common subspace of dimension , with basis vectors , for which the decomposition of the co-expression matrices (4) is exact, equation (7) becomes an equality and the eigenvectors of will lead to the exact basis of the common subspace. In HO GSVD applications to genomics data, can be as large as the total number of observations (i.e., ), and an exact common decomposition of the co-expression matrices might not be possible. In this case the eigenvectors of do not provide an exact decomposition of the subspace. Moreover, is not guaranteed to be non-defective and have a full set of real eigenvalues and eigenvectors. However, even in the absence of an exact common decomposition, the real part of the complex eigenvectors can be used to derive a low rank approximation of the common subspace and extract common and differential covariance structures from the data. To test the ability of our HO GSVD based algorithm to capture these covariance structures in the data in the presence of a “noisy” HO GSVD decomposition we performed an empirical simulation study (see Text S1 for details). Our simulations suggest that if a common subspace of dimension with basis vectors explains a significant fraction of the variance in the original datasets , the approximation (4) holds and the first eigenvectors of the matrix (corresponding to the largest eigenvalues of ) will provide a good approximation of the basis vectors of the HO GSVD subspace (Figure S4).
We selected two large gene expression datasets from rats and humans, where genome-wide expression profiles were assessed in the same subject/animal across multiple tissues. The rat datasets consisted of microarray-based expression profiles for probe sets that were measured in adrenal, aorta, fat, kidney, left ventricle, liver and skeletal muscle tissues in a panel of recombinant inbred rat strains [29]. Microarray expression data were retrieved from ArrayExpress, http://www.ebi.ac.uk/arrayexpress/, (skeletal muscle, E-TABM-458; aorta, E-MTAB-322; liver, E-MTAB-323, fat and kidney, E-AFMX-7; heart, MIMR-222; adrenal, E-TABM-457); gene expression summaries were derived using robust multichip average (RMA) algorithm [66] and normalized using Z-score transformation before analysis with C3D. The human data were retrieved from the Gene Expression Omnibus (GEO) database (www.ncbi.nlm.nih.gov/geo) under accession number GSE38805. Briefly, total RNA from the VZ, ISVZ, OSVZ, and CP of six 13–16 wk postconception human fetuses was isolated from laser-capture microdissected Nissl-stained cryosections of dorsolateral telencephalon (see [32] for additional details on experimental procedures). RNA-seq data were expressed as fragments per kilobase of exon per million fragments mapped (FPKM) values and normalized on log2 scale, yielding an expression matrix of in neocortex regions, which were analyzed by C3D.
The Matlab implementation of the C3D algorithm, detailed instructions to run the code and an example of the simulated datasets used in these studies can be downloaded from http://www.csc.mrc.ac.uk/Research/Groups/IB/IntegrativeGenomicsMedicine/ contact information: enrico.petrettocsc.mrc.ac.uk or xiaolin.xiaocsc.mrc.ac.uk
|
10.1371/journal.ppat.1006696 | Natural killer cell-intrinsic type I IFN signaling controls Klebsiella pneumoniae growth during lung infection | Klebsiella pneumoniae is a significant cause of nosocomial pneumonia and an alarming pathogen owing to the recent isolation of multidrug resistant strains. Understanding of immune responses orchestrating K. pneumoniae clearance by the host is of utmost importance. Here we show that type I interferon (IFN) signaling protects against lung infection with K. pneumoniae by launching bacterial growth-controlling interactions between alveolar macrophages and natural killer (NK) cells. Type I IFNs are important but disparate and incompletely understood regulators of defense against bacterial infections. Type I IFN receptor 1 (Ifnar1)-deficient mice infected with K. pneumoniae failed to activate NK cell-derived IFN-γ production. IFN-γ was required for bactericidal action and the production of the NK cell response-amplifying IL-12 and CXCL10 by alveolar macrophages. Bacterial clearance and NK cell IFN-γ were rescued in Ifnar1-deficient hosts by Ifnar1-proficient NK cells. Consistently, type I IFN signaling in myeloid cells including alveolar macrophages, monocytes and neutrophils was dispensable for host defense and IFN-γ activation. The failure of Ifnar1-deficient hosts to initiate a defense-promoting crosstalk between alveolar macrophages and NK cell was circumvented by administration of exogenous IFN-γ which restored endogenous IFN-γ production and restricted bacterial growth. These data identify NK cell-intrinsic type I IFN signaling as essential driver of K. pneumoniae clearance, and reveal specific targets for future therapeutic exploitations.
| The isolation of multidrug-resistant Klebsiella pneumoniae strains has significantly narrowed, or in some settings completely removed, the therapeutic options for the treatment of Klebsiella infections. Therapies targeting the immune system rather than the pathogen represent important alternatives. Despite the clinical relevance, there are still major gaps in our understanding of immune responses which drive the clearance of this pathogen. Type I interferons (IFNs) are known as powerful immune system regulators yet their effects on bacterial infections are disparate and remain elusive. In this study we show that type I IFN signaling is indispensable for mounting a protective and bacterial clearance-promoting immune response against K. pneumoniae. K. pneumoniae-induced type I IFNs launch a crosstalk between alveolar macrophages and NK cells by enabling NK cell IFN-γ production which in turn activates the macrophage anti-microbial armament. Type I IFN-responsive NK cells or IFN-γ administration rescue K. pneumoniae clearance in type I IFN-unresponsive hosts. Our study suggests that manipulation of type I IFN or IFN-γ levels might represent a valid strategy for treatment of drug-resistant K. pneumoniae infections.
| Klebsiella pneumoniae is a capsulated Gram negative pathogen which causes a wide range of infectious diseases, from urinary tract infections to pneumonia, the latter being particularly devastating among immunocompromised patients [1]. Of particular concern is the increasing isolation of multidrug resistant strains that narrows the therapeutic options for the treatment of K. pneumoniae infections [2–4]. To further complicate this scenario, recent population genomic studies have shown that virulent and multidrug resistant clones have access to a diverse mobile pool of virulence and antimicrobial resistance genes [2, 5] hence making possible the emergence of an extremely drug-resistant hypervirulent K. pneumoniae strain capable of causing untreatable infections in healthy individuals. It is then not surprising that multidrug resistant K. pneumoniae has been singled out as a significant threat to global public health by the World Health Organization, Centers of Diseases Control and Prevention, European Union and other organizations [4]. In light of this growing health problem, it is essential to better understand the immune pathways that are critical to host defense towards K. pneumoniae.
Successful defense against infections requires a coordinated action of multiple immune cell subsets. In this context, it is widely appreciated that type I interferons (IFNs) decisively coordinate immune responses by modulating cell-autonomous immunity and inflammatory responses, and by dictating immune cell-to-cell communications [6, 7]. While type I IFNs are the major effector cytokines of the host defense response against viral infections, a body of mainly recent data indicate that type I IFNs are also produced in response to bacteria [8–10]. However, depending on the bacterial infection, type I IFNs exert seemingly opposing functions [8, 10]. Type I IFNs protect against the progression of a localized Streptococcus pneumoniae lung infection to invasive disease [11, 12], prevent IL-1β-mediated tissue-damaging hyperinflammation in invasive soft tissue infection with Streptococcus pyogenes [13, 14] and restrict the growth of Legionella pneumophila in macrophages [15]. By contrast, type I IFNs impair the clearance of the intracellular pathogen Mycobacterium tuberculosis [16–18] and they are detrimental to host survival after Francisella tularensis infection [19, 20]. Overall, it is currently impossible to predict from the pathogen tropism or biology whether type I IFNs will be beneficial or detrimental for the host. This knowledge gap calls for additional mechanistic studies dissecting the type I IFN functions in other bacterial infections in order to define common and unique principles of the role of type I IFNs in host defense against bacterial infections.
Research over the last twenty years demonstrates that activation of early inflammatory responses, including production of TNF, IL-12, IL-23, IL-17 and IFN-γ, is essential to clear K. pneumoniae infections [21–23]. The role of type I IFNs in K. pneumoniae has so far not been addressed. Here we show that type I IFN signaling coordinates the communication between macrophages and NK cells to launch a protective immune response during lung infection with K. pneumoniae. Type I IFNs produced by macrophages upon K. pneumoniae challenge promote IFN-γ production by NK cells. IFN-γ in turn feeds back to prime macrophages for enhanced IL-12 production and bacterial killing. These results establish that NK cell-restricted type I IFN signaling entails resistance against severe lung infection caused by K. pneumoniae.
The importance of type I IFN signaling during K. pneumoniae-induced pneumonia was assessed by infecting type I IFN 1 receptor-deficient (Ifnar1-/-) mice. Ifnar1 is one of the subunits of the type I IFN receptor which mediates type I IFN responses in innate and acquired immunity to infection. Mice were intranasally infected with K. pneumoniae strain 52.145. This infection model recapitulates Klebsiella-triggered human pneumonia [24, 25]. Strain 52.145 belongs to the K. pneumoniae KpI group which includes the vast majority of strains most frequently associated with human infection, including numerous multidrug-resistant or hypervirulent clones [2]. This strain encodes all virulence functions significantly associated with invasive community-acquired disease in humans [2, 5]. Ifnar1-/- mice infected with 5 x 104 CFU K. pneumoniae 52.145 exhibited a markedly decreased survival (Fig 1A) and increased weight loss (Fig 1B) as compared to wild-type (WT) controls, demonstrating essential contribution of type I IFN signaling to host defense against this pathogen.
To elucidate possible mechanisms by which the absence of type I IFN signaling resulted in increased lethality, we examined the ability of Ifnar1-/- mice to control bacterial growth. Analysis of bacterial loads in lungs 12, 24 and 48 h following K. pneumoniae infection revealed a significant defect of Ifnar1-/- mice to restrict bacterial replication (Fig 1C). Both Ifnar1-/- and WT mice exhibited similar bacterial burdens 12 h post infection (p.i.) (Fig 1C, left panel). However, in contrast to WT mice, Ifnar1-/- mice failed to control bacterial growth at later time points resulting in significantly higher bacterial burdens in lungs at 24 and 48 h post infection (p.i.) (Fig 1C, middle and right panel). Consistently, the difference in CFU between Ifnar1-/- and WT mice was increasing with time of infection. Higher bacterial loads were also found in the spleen and liver of Ifnar1-/- mice than in organs of WT animals at 48 h p.i. (Fig 1D). The bacterial burden in the spleen and liver was negligible and similar in both genotypes at 12 h p.i. Analysis of lung sections 48 and 72 h p.i. showed more severe bronchopneumonia in Ifnar1-/- mice compared to WT controls, as revealed by a higher overall histopathology score comprising airway inflammation, intralesional bacterial burden and neutrophil infiltration (Fig 1E and 1F). The bronchopneumonia involved bronchi, bronchioles and to a lesser extent alveolar ducts and alveoli. The inflammation was predominantly neutrophilic, and aggregates of bacteria were evident within the foci of inflammation. No appreciable differences between Ifnar1-/- and WT mice were noted in lung histology of animals which were given intranasal PBS (S1A Fig).
The lung pathology analysis suggested that severe lung destruction was the cause of increased mortality of Ifnar1-/- animals. To test this, we infected a cohort of Ifnar1-/- mice and monitored the progress of bronchopneumonia toward a lethal disease by comparing lungs of mice which were reaching behavioral and/or patho-physiological humane endpoints with lungs of mice which were showing mild symptoms at the same time of infection. In addition, we sampled lungs of mice which survived longer than 10 days and appeared recovered. Lung section analyses revealed that animals showing the most severe symptoms, i.e. animals approaching human endpoints, exhibited large areas of lung inflammation (more than 60% of the lung tissue affected) including increased intralesional bacteria and neutrophil infiltration (S1B Fig, left panel). The high degree of pathological changes in lungs of these mice was not compatible with life since animals with similarly severe symptoms ultimately developed a lethal disease in survival experiments (Fig 1). In contrast, animals showing only mild symptoms or animals surviving the 10 day observation period exhibited low or no lung inflammation at all, respectively (S1B Fig, middle and right panels).
Having established the importance of type I IFN signaling for host defense against K. pneumoniae, we sought to determine the signaling pathway(s) activated by the pathogen to induce type I IFN production and signaling. Myeloid cells such as alveolar macrophages produce type I IFNs in response to many bacterial pathogens [8–10]. To test whether alveolar macrophages are type I IFN producers during K. pneumoniae infection, alveolar macrophages (CD45+CD11chighSiglecF+) were isolated from infected C57Bl/6 mice 24 h p.i. These cells displayed induction of Ifnb, the key type I IFN gene, when compared to uninfected controls (Fig 2A). Consistently, the type I IFN-stimulated gene (ISG) Isg15 was also induced (Fig 2A). Induction of Tnf confirmed that alveolar macrophages were activated by the infection (Fig 2A). Type I IFNs were also induced in the mouse alveolar macrophage cell line MH-S and bone marrow derived macrophages (BMDMs) infected with K. pneumoniae (S2A Fig). In agreement, K. pneumoniae infection induced the expression of the ISGs Mx1, Ifit1 and Isg15 in BMDMs (Fig 2B). Tnf expression was similar in WT and ifnar1-/- BMDMs (Fig 2B). Further, we observed ISG15 modification (ISGylation) of proteins in K. pneumoniae-infected cells (S2B Fig). The induction of ISGs and ISGylation by K. pneumoniae was abrogated in Ifnar1-/- BMDMs demonstrating the requirement for type I IFN signaling (S2B Fig).
The central tenet of type I IFN production is that the initial wave of type I IFN production relies on the activation of the IFN-regulatory factor (IRF) 3 [26]. K. pneumoniae-induced expressions of Ifnb and of the ISGs Mx1, Ifit1 and Isg15 were ablated in infected Irf3-/- BMDMs Fig 2B). In contrast, Tnf expression was not affected by the Irf3 deletion (Fig 2B). These data reveal a fundamental role of IRF3 in induction of K. pneumoniae-triggered type I IFNs and ISGs.
Bacteria trigger type I IFN production by activation of various pattern recognition receptors (PRRs) [8, 10]. In vivo and in vitro evidence has demonstrated that TLR4 governs host defenses against K. pneumoniae [27–30]. To test the involvement of TLR4 in K. pneumoniae-induced IFN-β production, we employed BMDMs derived from Tlr4-/- mice. Supporting the key role of TLR4 in Ifnb induction, K. pneumoniae-triggered phosphorylation of Ifnb gene drivers IRF3 and the IRF3 kinase TBK1 [26] were abrogated in Tlr4-/- BMDMs (Fig 2C). The lack of TLR4 abolished induction of Ifnb, Mx1, Ifit1 and Isg15 and protein modification by ISG15 (ISGylation) (Fig 2D and S2C Fig). Consistent with the involvement of the canonical TLR4-activated pathway [26], macrophages lacking the adaptors TRIF and TRAM did not induce protein ISGylation and the expression of Ifnb, Mx1, Ifit1 and Isg15 in response to K. pneumoniae infection (S2D and S2E Fig). Phosphorylation of IRF3 and TBK1 was not stimulated in macrophages lacking the adaptors TRIF and TRAM, but proceeded normally in the absence of MyD88 (S2F Fig). Together, these results demonstrate that K. pneumoniae-induced type I IFN production and signaling is dependent on the TLR4-TRIF-TRAM-IRF3 pathway.
We and others have shown that K. pneumoniae capsule polysaccharide (CPS) and lipopolysaccharide (LPS) are recognized by TLR4 to launch inflammatory responses [31, 32]. Therefore, we hypothesized that these polysaccharides might be involved in TLR4-mediated type I IFN induction. To test this, macrophages were infected with a cps, O-polysaccharide, and double cps-O-polysaccharide K. pneumoniae mutants, and type I IFN was quantified in the supernatants of infected cells. The three mutants induced less type I IFN than the wild-type strain (Fig 2E) although the cps mutant induced more type I IFN than the two O-polysaccharide mutants. The lack of induction of type I IFN production was in agreement with the reduced phosphorylation of IRF3 triggered by the mutants (S2G Fig) indicating that the CPS and LPS O-polysaccharides are the K. pneumoniae factors activating TLR4 to induce type I IFN.
To confirm that type I IFN signaling is activated by K. pneumoniae also in vivo, we examined the lung tissue 12 h p.i. Expression of the ISGs Mx1, Ifit1 and Isg15 was induced in WT but not Ifnar1-/-mice (Fig 3A) corroborating the results obtained using infection of macrophages. To assess whether type I IFN signaling influenced the inflammatory response in lungs of K. pneumoniae-infected animals, we analyzed several inflammation-associated cytokines and chemokines 12 h p.i. K. pneumoniae–induced expression of IFN-γ, a critical cytokine for defense against lung infection with K. pneumoniae [22], was virtually absent in Ifnar1-/- mice at both mRNA and protein levels (Fig 3B and S3A Fig). Similarly, the induction of IL-12 mRNA (Il12b) and protein (IL-12p70), a key IFN-γ inducer, was impaired in Ifnar1-/- mice (Fig 3B and S3A Fig). The induction of CXCL10, a chemokine required for NK cell recruitment and host defense against K. pneumoniae [33], was abolished in Ifnar1-/- mice (Fig 3B). The mRNA of the immediate early cytokine Tnf was induced upon infection in both WT and Ifnar1-/- mice although the levels were lower in Ifnar1-/- when compared to WT animals (Fig 3B). At the protein level, TNF was comparable in both genotypes (S3A Fig). Expression of the anti-inflammatory Il10, the neutrophil chemoattractant Cxcl1 and the pro-inflammatory Il1b, were not affected by type I IFN signaling (Fig 3B and S3B Fig). The defect of Ifnar1-/- mice in induction of Ifng, Il12b and Cxcl10 was persistent and clearly detectable at 48 p.i. (Fig 3C). However, the expression of Tnf was no longer different between Ifnar1-/- and WT mice at this later time point (Fig 3C). The expression of Il10, Cxcl1 and Il1b was comparable in Ifnar1-/- and WT mice at 48 h p.i. (Fig 3C and S3C Fig), as observed already at the 12 h time point (Fig 3B). Together, the lack of type I IFN signaling results in defect in the production of IFN-γ, IL-12 and CXCL10 in K. pneumoniae-infected lungs.
The failure of Ifnar1-deficient mice to induce IL-12 and IFN-γ upon K. pneumoniae infection suggested a defect in immune cells producing these cytokines. Flow cytometry analysis revealed comparable numbers of alveolar macrophages (4% CD11chighSiglecF+ alveolar macrophages of CD45+ leukocytes) in lungs of Ifnar1-/- and WT animals 12 h after infection with K. pneumoniae (Fig 4A and S4A Fig). Infection did not increase the alveolar macrophage population over PBS-treated controls (Fig 4A). The numbers of neutrophils (Cd11b+Ly6G+Ly6Cmed) and inflammatory monocytes (CD11b+Ly6G-Ly6Chigh) in K. pneumoniae-infected lungs were also similar in both genotypes (Fig 4A, S4B Fig). Both neutrophils and inflammatory monocytes increased upon infection as compared to PBS-treated controls (Fig 4A), consistent with previous studies [21]. In contrast, the population of CD3-NK1.1+ NK cells, which have been implicated in defense against K. pneumoniae [34], was decreased in lungs of infected Ifnar1-/- mice when compared to WT controls both in terms of percentage of total leukocytes as well as absolute cell numbers (Fig 4B). The NK cell numbers in infected Ifnar1-/- mice were comparable to those in PBS-treated mice (Fig 4B). Importantly, the NK cell population from infected lungs of Ifnar1-/- animals contained significantly lower percentage of IFN-γ-producing cells than that from WT mice (Fig 4C). The populations of CD4 and CD8 T cells were also lower in lungs of infected Ifnar1-/- mice compared to WT controls (Fig 4D and 4F, S4C Fig) but these cells did not produce IFN-γ regardless of the genotype (Fig 4E and 4G, S4C Fig), as revealed by comparison with PBS-treated controls. The lower numbers of CD4 and CD8 T cells in Ifnar1-/- mice can be explained be impaired expression of the T cell chemokine Cxcl10. In sum, the absence of type I IFN signaling results in a defect in NK cell accumulation and NK cell-derived IFN-γ production in the lung of K. pneumoniae-infected mice.
Macrophages require IFN-γ and/or IFN-γ priming for a complete anti-microbial response and for expression of the NK- and T cell-activating cytokine IL-12 and the NK and T cell chemoattractant CXCL10 [35, 36]. These responses were insufficiently activated in Ifnar1-/- mice (Figs 3 and 4) suggesting that the impaired IFN-γ production in Ifnar1-/- mice was causally involved in the low IL-12 and CXCL10 expression, and in the failure to control bacterial growth. To test this hypothesis, we isolated alveolar macrophages from WT and Ifnar1-/- mice, primed them with IFN-γ and infected subsequently with K. pneumoniae. Both WT and Ifnar1-/- alveolar macrophages primed for 5 h with IFN-γ induced Il12b upon K. pneumoniae infection (Fig 5A). Without priming, no Il12b was induced (Fig 5A). Cxcl10 was induced by IFN-γ alone in both WT and Ifnar1-/- alveolar macrophages (Fig 5A), consistent with the known direct activation of Cxcl10 by IFN-γ [36]. In contrast, Tnf and Il1b were induced regardless of priming (Fig 5A). To assess the effect of priming on the anti-bacterial activity, IFN-γ–primed or mock-treated macrophages were infected, treated with gentamicin 1 h p.i. and incubated for additional 2 h. As anticipated, priming of macrophages with IFN-γ for 2 h resulted in a significant decrease in the number of intracellular bacteria at 3 h p.i. (Fig 5B) confirming the activating effect of IFN-γ on anti-bacterial macrophage activity [37]. Adhesion and uptake of bacteria were not affected by IFN-γ treatment (Fig 5C). Together, these results suggest that the key function of type I IFN production and signaling in defense against K. pneumoniae infection is the induction of IFN-γ. Since we found that the TLR4-IRF3 axis drives type I IFN production and, consequently, type I IFN signaling in K. pneumoniae-infected BMDMs (Fig 2) we asked whether IRF3 is required for IFN-γ induction and host defense in vivo. Infection of Irf3-/- mice confirmed an impairment in Mx1 and Ifit1 but not Tnf induction (Fig 5D) suggesting that IRF3 is critically involved in K. pneumoniae-elicited type I IFN production in vivo. Moreover, Irf3-/- mice exhibited impaired expression of Ifng, Il12b and Cxcl10 and reduced ability to control bacterial growth (Fig 5E and 5F) demonstrating that deficiency in type I IFN induction has similar consequences for IFN-γ production as the lack of Ifnar1. Interestingly, however, direct triggering of type I IFN signaling by intranasally administered IFN-β into WT mice in the absence of K. pneumoniae did not activate Ifng gene expression in the lung despite a strong induction of ISGs as well as Cxcl10 (Fig 5G). Thus, IFN-γ, which is required for efficient killing of K. pneumoniae by macrophages, is activated by type I IFN signaling only in the context of K. pneumoniae infection.
Since alveolar macrophages produce and respond to type I IFNs during K. pneumoniae infection (Fig 2 and S2 Fig) we asked whether alveolar macrophage-intrinsic type I IFN signaling contributes to the immune response in vivo by using Ifnar1fl/fl-CD11cCre mice. CD11c promoter-driven of Cre recombinase expression results in the deletion of a loxP-flanked allele in conventional DCs (CD11chigh) and alveolar macrophages [38, 39]. Consistently, Ifnar1 was deleted in alveolar macrophages isolated from Ifnar1fl/fl-CD11cCre mice (S5A Fig). Ifnar1fl/fl-CD11cCre mice were similarly resistant against K. pneumoniae infection as Ifnar1fl/fl controls (Fig 6B). Expression of the type I IFN target gene Mx1 was significantly reduced in lungs of Ifnar1fl/fl-CD11cCre mice (Fig 6C) demonstrating that the CD11c+ cells represent a significant population of type I IFN-responding cells in the lung during K. pneumoniae infection. However, in agreement with the efficient defense, expression of Ifng, Il12b, and Cxcl10 was not impaired in Ifnar1fl/fl-CD11cCre mice (Fig 6C). Expression of Tnf, Il10 and Cxcl1 was similar in both Ifnar1fl/fl-CD11cCre and Ifnarfl/fl mice (Fig 6C). Flow cytometry analysis revealed comparable numbers of alveolar macrophages, neutrophils, inflammatory monocytes, NK cells as well as CD4 and CD8 T cells in lungs of Ifnar1fl/fl-CD11cCre and Ifnar1fl/fl mice 12 h after K. pneumoniae infection (Fig 5D and S5B–S5D Fig). Furthermore, mice lacking Ifnar1 in macrophages and neutrophils (Ifnar1fl/fl-LysMCre) or neutrophils only (Ifnar1fl/fl-MRP8Cre) exhibited similar clearance of K. pneumoniae as Ifnar1fl/fl controls (S6A and S7A Figs). The expression of type I IFN responsive genes Mx1 and Ifit1 was reduced in Ifnar1fl/fl-LysMCre and to a lesser extent also in Ifnar1fl/fl-MRP8Cre mice (S6B and S7B Figs). Importantly, expression of Ifng, Il12b, and Cxcl10 and the numbers of key immune cell subsets were not impaired in Ifnar1fl/fl-LysMCre and Ifnar1fl/fl-MRP8Cre mice (S6C–S6F and S6C–S6F Fig).
Together, type I IFN signaling in alveolar macrophages and in myeloid cells in general does not contribute to protective immune responses against K. pneumoniae although these cells generate a substantial part of the type I IFN signature in the infected lung.
To find out whether the impaired activation and accumulation of NK cells in Ifnar1-/- animals were causatively involved in the increased susceptibility to K. pneumoniae infection, we carried out NK cell transfer experiments. NK cells isolated from WT mice were adoptively transferred into Ifnar1-/- mice using 1 x 106 NK cells per recipient animal. Following K. pneumoniae infection, lungs of recipient mice were examined for NK cell accumulation, IFN-γ production by NK cells and bacterial burden. The numbers of donor WT NK cells in lungs of recipient Ifnar1-/- mice were higher than the numbers of recipients’ own NK cells (i.e., Ifnar1-/- NK cells) (S8 Fig), consistent with the observation that NK cell accumulation was higher in WT mice than in Ifnar1-/- animals (Fig 4B). Importantly, the recipient Ifnar1-/- mice showed similar percentages of IFN-γ-producing exogenous WT and endogenous Ifnar1-/- NK cells, and both of these percentages were significantly higher than the percentage of IFN-γ-producing NK cells in Ifnar1-/- mice without adoptive transfer (Fig 7A). Thus, the transfer of WT NK cells raised the percentage of IFN-γ-producing endogenous Ifnar1-/- NK cells. Finally, Ifnar1-/- mice that received WT NK cells exhibited approximately 10 times lower bacterial loads in lungs and almost lacked dissemination to the spleen when compared to mice which did not receive WT NK cells (Fig 7C and 7D). WT NK cells transferred into WT animals exhibited substantial IFN-γ production (Fig 7B) but they did not significantly raise bacterial clearance (Fig 7C and 7D), which is in agreement with the efficient defense of WT mice against K. pneumoniae infection.
These results suggested that the diminished NK cell-derived IFN-γ production was causative of the impaired bacterial clearance in Ifnar1-/- mice. To test this, Ifnar1-/- mice were intranasally administered IFN-γ at the time of infection. The IFN-γ treatment boosted the expression of Ifng and Cxcl10 (Fig 7E) and reduced bacterial loads in lungs (Fig 7F).
In sum, the data show that donor WT NK cells display similar properties in recipient Ifnar1-/- mice as in WT mice. Thus, the deficient NK cell accumulation and lower percentage of IFN-γ-producing NK cells observed in Ifnar1-/- mice (Fig 4B and 4C) result from a cell-autonomous defect. Moreover, Ifnar1-deficient NK cells regain the ability to produce IFN-γ if macrophage priming is accomplished by exogenous IFN-γ or by IFN-γ derived from transferred WT NK cells.
In this study, we examined the role and mechanisms of action of type I IFN signaling in the context of lung infection with K. pneumoniae, a pathogen with one of the highest emergence of antibiotic resistant strains [2–4]. Our findings identify type I IFN signaling as a key driver of the mutually activating crosstalk between NK cells and alveolar macrophages which ultimately results in bacterial clearance and successful host defense. This discovery reveals a previously unrecognized mechanism of type I IFN-mediated anti-bacterial immunity and sheds light into regulation of immune responses against an utmost challenging human pathogen.
The protective effect of type I IFNs against K. pneumoniae-triggered pneumonia is caused by different mechanisms than those reported for other bacterial pathogens with the same tropism. Defense against S. pneumoniae, a gram-positive pathogen, is dependent on type I IFN signaling in alveolar epithelial type II cells. These cells require type I IFN signals to resist the destructive and death-promoting environment elicited in the course of S. pneumoniae infection [12]. The absence of type I IFN signaling causes an excessive destruction of the lung epithelial barrier and subsequent massive systemic dissemination of the pathogen. The lung barrier function is supported by type I IFNs by their protective effects on epithelial tight junctions and by inhibition of bacterial transmigration [11]. The protective effects of type I IFNs against infection with the gram negative intracellular pathogen L. pneumophila result from inhibition of the intracellular replication of the pathogen in infected macrophages [15]. The key type I IFN effector in this context appears to be the bactericidal itaconic acid which is produced by an enzyme encoded by the type I IFN target gene Irg1 [15]. In contrast, the protective effects of type I IFNs against K. pneumoniae described in our study result from NK cell-dependent IFN-γ-mediated restriction of bacterial growth. IFN-γ controls K. pneumoniae growth in the lung but the precise mechanism has not been elucidated [22]. Lung failure resulting from exacerbated infection-elicited tissue destruction is a critical aspect of pneumonia since mitigation of lung injury and/or lessening of inflammation is associated with disease amelioration [40, 41]. Thus, the higher and progressively increasing bacterial burden and tissue injury in the lung of K. pneumoniae-infected Ifnar1-deficient mice suggest that these mice ultimately suffer respiratory failure.
It is well established that type I IFNs are typically induced following intracellular sensing of invading and/or phagocytosed bacteria thereby resembling induction by viruses, i.e. obligatory intracellular pathogens [9, 10]. This intracellular signaling principle applies also to the cell wall component LPS which activates the IFN-β-inducing TBK1-IRF3 pathway upon signaling emanating from the LPS-TLR4 complex localized in the endosomal membrane [42]. Our study reveals that this intracellular signaling is also the driver of type I IFN induction by K. pneumoniae which triggers the TBK/IRF3 pathway upon TLR4-dependent recognition of LPS and the capsule polysaccharide (CPS). Most common bacterial inducers of type IFNs are cell wall components (e.g. LPS) and nucleic acids (both RNA and DNA) [10]. The capacity of individual inducers to activate type I IFN production appears to be different in different innate immune cell types [13] leaving open the possibility that K. pneumoniae employs other inducers in addition to LPS and CPS in vivo. Interestingly, the K. pneumoniae-induced type I IFNs have no autocrine functions since mice lacking Ifnar1 in alveolar macrophages, the key sentinel cells in the lung [43], are similarly resistant to infection as Ifnar1-proficient mice. This is in marked contrast to skin infection with S. pyogenes in which myeloid cells are both the key type I IFN producer and effector cells [14]. Our data demonstrate that type I IFNs promote production of IFN-γ, a critical activator of antimicrobial effector functions of macrophages, by NK cells in K. pneumoniae infection. The mechanism of antimicrobial macrophage activation by IFN-γ involves primarily the stimulation of oxidative burst but other effector functions are emerging [37, 44]. For example, the IFN target genes GBPs (guanylate-binding proteins), which are linked to phagosomal processes, and IFN-γ-stimulated metabolic reprogramming have been implicated in microbial killing by macrophages [44, 45]. Interestingly, recent evidence demonstrates the ability of K. pneumoniae to manipulate phagosome maturation and survive antimicrobial attacks by macrophages [46] suggesting that IFN-γ might counteract such phagosome evasion mechanisms of K. pneumoniae. Future studies should investigate molecular mechanisms of K. pneumoniae eradication by macrophages in detail. Importantly, our data show that such mechanisms are activated by IFN-γ but not type I IFNs since autocrine type I IFN signaling in alveolar macrophages is dispensable for both macrophage priming and protective immune response. Type I IFN and type II IFN (i.e. IFN-γ) signaling pathways share the transcription factor STAT1, raising the therapeutically important question of STAT1 target genes acting as specific effectors of IFN-γ signaling during host defense against K. pneumoniae infection.
We provide complementary evidence for the fundamental importance of NK cell-autonomous type I IFN signaling in defense against K. pneumoniae infection. First, the use of Ifnar1fl/fl-Cd11cCre, Ifnar1fl/fl-LysMCre and Ifnar1fl/fl-MRP8Cre mouse strains allows the conclusion that type I IFN signaling in none of the major myeloid cell subsets present in infected lungs contributes to IFN-γ and IL-12 production, and to host defense. Second, WT NK cells introduced into Ifnar1-/- mice produce IFN-γ in the Ifnar1-deficient environment and promote bacterial clearance. The NK cell-intrinsic requirement for type I IFN signaling in IFN-γ production in the course of K. pneumoniae infection is unprecedented in the context of bacterial infections studied to date. Interestingly, Ifnar1-/- NK cells retain the ability to induce IFN-γ in K. pneumoniae-infected Ifnar1-/- mice, as evident from IFN-γ production by Ifnar1-/- NK cells after adoptive transfer of WT NK cells. Classically, NK cell production of IFN-γ is triggered by IL-12 originating from various subsets of myeloid cells [47]. This implies that in context of the NK cell transfer experiment, the WT NK-derived IFN-γ restores macrophage priming and IL-12 production which in turn activate NK cells regardless of their type I IFN signaling capacity.
Type I IFNs can directly stimulate NK cells to produce IFN-γ in response to certain viral infections [48, 49]. It is speculated that this mode of IFN-γ production is relevant during infection with viruses which induce little or no IL-12 such as the lymphocytic choriomeningitis virus [48]. Type I IFN-mediated NK cell stimulation involves direct activation of the IFN-γ gene driver STAT4 by the type I IFN receptor-associated JAK kinases [50, 51]. Such activation has so far not been reported for bacterial infections. Our experiment showing that administration of IFN-β alone does not result in IFN-γ induction indicates that type I IFNs act on NK cells in concert with accessory signals generated during K. pneumoniae infection. This mechanism would resemble IL-12 production by myeloid cells which requires signals derived from the pathogen recognition and from IFN-γ. NK cells express several pattern recognition receptors, e.g. TLR2 and TLR4, which potentially provide means for activation by bacterial products [52, 53]. However, the relevance of TLRs for NK cell activation during bacterial infections is unclear as NK cell-autonomous MyD88-dependent TLR signaling does not contribute to NK cell stimulation [54]. Of emerging interest are NK cytotoxic receptors since the activating receptor NKp46 and the mouse ortholog NCR1 have recently been demonstrated to act as pattern recognition receptors for Fusobacterium nucleatum and Candida glabrata [55, 56]. Thus, activating NK cell receptors might provide accessory signals for induction of IFN-γ by type I IFNs in the course of bacterial diseases such as the K. pneumoniae lung infection described in this study. The prominent role of type I IFN signaling in IFN-γ induction is in line with increased susceptibility of TRIF-deficient mice to K. pneumoniae infection [57]. TRIF-deficient mice exhibit impaired IFN-γ induction during K. pneumoniae infection and display improved resistance if treated with exogenous IFN-γ [57]. Our data showing the requirement for TRIF in IFN-β induction suggest that the impairment of IFN-γ production in TRIF-deleted mice is secondary to the deficient type I IFN (e.g. IFN-β) production in these mice. The exceptionally tight link between type I IFN signaling and NK cell IFN-γ production might represent a specific feature of defense against K. pneumoniae.
By virtue of their inhibition of viral replication type I IFNs became the first cytokines to be used in therapy of human diseases, most notably in infections caused by the hepatitis C virus [58, 59]. In contrast, the use of type I IFNs or their effectors for therapy of infectious diseases caused by bacteria is currently precluded by their incompletely understood and disparate effects during bacterial infections. Our study reveals an unexpected dependence of NK cell IFN-γ production on NK cell-intrinsic type I IFN signaling during bacterial pneumonia. This mechanism of IFN-γ induction and the resulting anti-bacterial activation of macrophages are indispensable for successful defense against K. pneumoniae lung infection. The increasing isolation of multidrug-resistant K. pneumoniae strains makes an urgent priority to develop effective therapeutics based on new targets and concepts. K. pneumoniae is exemplary of the mismatch between unmet medical needs and the current antimicrobial research and development pipeline. Arguably, therapies targeting the immune responses which thereby circumvent antibiotic resistance are highly desirable. Our study reveals an important and therapeutically exploitable aspect of immune defense against K. pneumoniae. Importantly, clinical records reporting frequent K. pneumoniae infections in patients deficient in IL-12 [60] suggest that the type I IFN-driven communication network between IFN-γ-producing NK cells and IL-12-producing myeloid cells is relevant also for humans.
Animal experiments were carried out at the Queen’s University Belfast and Max F. Perutz Laboratories of the University of Vienna. Experiments involving mice at Queen’s University Belfast were approved by the Queen’s University Belfast’s Ethics Committee and conducted in accordance with regulations described in the UK government Animals Act 1986 under the Project License PPL2700 issued by the UK Home Office. Animals were randomized for interventions but researches processing the samples and analyzing the data were aware which intervention group corresponded to which cohort of animals. Experiments involving mice at the Max F. Perutz Laboratories of the University of Vienna were discussed with the institutional ethics committee and performed in accordance with the Austrian law for animal experiments (BGBl. I Nr. 114/2012) under the permissions BMWF-66.006/0006-II/3b/2013 and BMWFW-66.006/0019-WF/V/3b/2016 issued by the Austrian Ministry of Science to PK.
Mice were bred and kept under specific pathogen free (SPF) conditions according to recommendations of the Federation of European Laboratory Animal Science Association (FELASA). Ifnar1-/- mice have been previously described [61]. Ifnar1fl/fl-CD11c-Cre, Ifnar1fl/fl-LysMCre and Ifnar1fl/fl-MRP8Cre mice were obtained by crossing Ifnar1fl/fl mice [62] with CD11c-Cre mice, LysMCre and MRP8Cre [38, 63, 64], respectively, and littermate Cre+ and Cre- control mice were used. All mice were on C57BL/6 background. C57BL/6N wild type (WT) mice were purchased from Charles River (Vienna) and Harlan (Belfast). Experiments were carried out using 7–12 weeks old mice with age and gender being matched between genotypes. For experiments requiring anesthesia, a solution of 10 mg/ml ketamine and 1 mg/ml xylazine (aniMedica) in isotonic saline (Sigma) was injected intraperitoneally (i.p.).
K. pneumoniae 52.145 is a clinical isolate (serotype O1:K2) belonging to the CC65K2 virulent clonal group [5, 65]. The isogenic capsule mutant, strain 52145-ΔwcaK2, has been described [66]. The LPS O-polysaccharide mutants, targeting the glf glycolstransfrase essential for the LPS O-polysaccharide biosynthesis [67] were constructed by insertion mutagenesis using the pir replication dependent plasmid pSF100 (to be described elsewhere). Bacteria were grown in LB medium at 37°C and carbenicillin 50 μg/ml was added to the growth medium to grow O-polysaccharide mutants.
Immortalized BMDM (iBMDM) cells (BEI Resources, NIAID, NIH, wild-type, NR-9456; Trif-/-, NR-9566; Tram-/-, NR-9567; Myd88-/-, NR-15633; and Trif-/-Tram-/- mice, NR-9568) were grown in Dulbecco’s Modified Eagle Medium (DMEM; Gibco 41965) supplemented with 10% heat-inactivated fetal calf serum (FCS), 100 U/ml penicillin, and 0.1 mg/ml streptomycin (Gibco) at 37°C in a humidified 5% CO2 incubator. Murine alveolar macrophages MH-S (ATCC, CRL-2019) were grown in RPMI 1640 tissue culture medium supplemented with 10% heat-inactivated fetal calf serum (FCS), 100 U/ml penicillin, and 0.1 mg/ml streptomycin (Gibco) and 10 mM HEPES (Sigma-Aldrich). Cells were routinely tested for Mycoplasma contamination. To isolate BMDMs, tibias and femurs from wild-type, irf3-/- and tlr4-/- knock-out mice were removed using a sterile technique and the bone marrow was flushed with fresh medium. To obtain macrophages, cells were plated in L929-conditioned medium and cultivated for 4–6 days. Medium was replaced with fresh supplemented media every 3 days. For infections, bacteria were adjusted to an OD600 of 1.0 in PBS and infections were performed using a multiplicity of infection (MOI) of 70 bacteria per cell. To synchronize infection, plates were centrifuged at 200 x g for 5 min. For incubation times longer than 60 min, bacteria were killed by addition of gentamicin (100 μg/ml) which was not removed until the end of the experiment.
To isolate alveolar macrophages (AMs), mice were euthanized by anesthetic overdose, trachea was surgically exposed and cannulated (BD Angiocath) followed by flushing of lungs with 5 x 1 ml of cold PBS + 0.5 mM EDTA. Obtained bronchoalveolar lavage fluids (BALF) from mice of the same genotype (n ≥ 6) were pooled on ice. Cells were collected from BALF by centrifugation (10 min, room temperature, 400 x g), resuspended and cultured in RPMI + 10% FCS + 1% penicillin and streptomycin. After 2 h, adherent cells were > 95% AMs as determined by trypan blue staining and were seeded for stimulation at the final concentration of 2 x 105 cells/ml in RPMI + 10% FCS. For priming, cells were pretreated with 5 ng/ml mouse IFN-γ (eBioscience) for 5 h, followed by infection (MOI = 70) for 1 h. To synchronize infection, plates were centrifuged at 200 x g for 5 min. One hour after treatment with gentamicin (150 μg/ml) cells were lysed using Isol-RNA Lysis Reagent (5 Prime) and RNA was isolated.
Bacteria in the stationary phase (overnight culture) were sub-cultured and grown at 37°C with agitation to reach mid log phase. Subsequently, bacteria were harvested by centrifugation (20 min, 2500 x g, 24°C), resuspended in 1x PBS and adjusted to 5 x 104–1 x 105 colony forming units (CFU) per 30 μl as determined by plating 10-fold serial dilutions on LB plates. Mice were anesthetized and 30 μl of bacterial suspension were inoculated intranasally. Infected animals were monitored every 4 to 8 hours and were euthanized when reaching behavioral and/or pathophysiological humane endpoints. Survival was monitored for 10 days. For experiments other than survival, animals were euthanized at indicated time points by either cervical dislocation or anesthetic overdose.
To determine bacterial loads, mice were euthanized at indicated time points. Lungs, spleens and livers were aseptically removed, weighted and placed in cold 1 x PBS. After mechanical homogenization, 10-fold dilution series were prepared from organ homogenates and plated on LB plates. The following day, colonies were counted and bacterial load was calculated as colony forming units (CFUs) per gram of tissue.
Mice were euthanized by anesthetic overdose. To collect lungs for histology trachea was surgically exposed and cannulated (BD Angiocath). 4% paraformaldehyde (PFA) was injected through trachea to inflate lungs, followed by their aseptic dissection from the thoracic cage. Inflated lungs were fixed overnight in excessive volume of 4% PFA, dehydrated, embedded in paraffin and 3 μm thick sections were prepared. Hematoxylin and eosin (H&E) staining of the sections was performed according to the standard protocol. H&E stained slides were evaluated by a board certified pathologist using an Axioskop 2 MOT microscope (Carl Zeiss). For additional review and image acquisition, representative slides were scanned using a Pannoramic Scan II slide scanner (3D Histech). Digital images were acquired with the Pannoramic Slide Viewer software (3D Histech). Sections were examined for airway inflammation (inflammatory cell infiltration of the intrapulmonary airways, alveolar ducts and alveoli), neutrophilic infiltration and intralesional bacterial burden (qualitative/semi-quantitative extent of bacterial presence within the regions of inflammation). Standard pathology criteria were used to score histomorphologic features of the lesion—degree of the lesion and the extent of involvement: score 0 –none/insignificant; score 1 –minimal; less than 10%, score 2 –mild; 10 to 30%, score 3 –moderate; 30–60% and score 4 –severe, more than 60%.
Lung homogenates of infected and uninfected mice were prepared as for determination of bacterial load in 1x PBS containing protease inhibitors (Complete protease inhibitor, Roche) and frozen at -80°C. Homogenates were subjected to two cycles of thawing and freezing, centrifuged (5 min, 16000 x g, 4°C) and supernatants were collected for further measurements. Cytokine concentrations of TNF, IL-1β, IFN-γ and IL-12p70 were measured in supernatants using DuoSet ELISA kits (R&D Systems). Kits were used according to the manufacturer’s instructions. Total protein in lungs was determined by Pierce BCA Protein Assay Kit (Thermo Scientific) and used for normalization purposes.
Lysates were prepared in lysis buffer (1x SDS Sample Buffer, 62.5 mM Tris-HCl pH 6.8, 2% w/v SDS, 10% glycerol, 50 mM DTT, 0.01% w/v bromophenol blue). Proteins were resolved by standard 10% SDS-PAGE and electroblotted onto nitrocellulose membranes. Membranes were blocked with 4% bovine serum albumin (w/v) in TBST and protein bands were detected with specific antibodies using chemiluminescence reagents and a G:BOX Chemi XRQ chemiluminescence imager (Syngene). The following rabbit antibodies were used: anti-phospho IRF3 (Ser 396) (1:1000; Cell Signaling #4947), anti-phospho-TBK1 (Ser 172) (1:1000; Cell Signaling #5483), and anti-ISG15 (1:1000; Cell Signaling #9636). Immunoreactive bands were visualized by incubation with horseradish peroxidase-conjugated goat anti-rabbit immunoglobulins (1:5000) or goat anti-mouse immunoglobulins (1:1000; Bio-Rad). To ensure that equal amounts of proteins were loaded, blots were re-probed with α-tubulin (1:3000; Sigma-Aldrich). To detect multiple proteins, membranes were re-probed after stripping of previously used antibodies using a pH 2.2 glycine-HCl/SDS buffer.
Cells were seeded in (6-well plate; 1 x 106 cells per well) and grown for 24 h. Cells were infected (MOI = 100) for the indicated time points, and supernatants were collected. Murine type I IFNs were detected using B16-Blue IFN-α/β cells (Invivogen) which carry a SEAP reporter gene under the control of the type I IFN-inducible ISG54 promoter enhanced by a multimeric ISRE. Levels of SEAP in the supernatants were determined as per the manufacturer’s instructions.
K. pneumoniae intracellular survival was assessed as previously described with minor modifications [46]. Briefly, macrophages were seeded in 12-well tissue culture plates at a density of 2.5 x 105 cells per well 15 h before the experiment. Bacteria were grown in 5-ml LB, harvested in the exponential phase (2500 x g, 20 min, 24°C), washed once with PBS and a suspension containing approximately 1 x 108 CFU/ml was prepared in 10 mM PBS (pH 6.5). Cells were infected with 175 μl of this suspension to obtain MOI of 70 bacteria per cell in a final volume of 1 ml DMEM tissue culture medium supplemented with 10% heat-inactivated FCS and 10 mM Hepes. To synchronize infection, plates were centrifuged at 200 x g during 5 min. After 30 min of contact, cells were washed twice with PBS and incubated for additional 60 min with 1 ml tissue culture medium supplemented with gentamicin (100 μg/ml) to eliminate extracellular bacteria. Initial attachment of bacteria was assessed after 30 min contact as previously described [46]. To determine intracellular bacterial load, cells were washed three times with PBS and lysed with 300 μl of 0.05% saponin in PBS for 10 min at room temperature. Serial dilutions were plated on LB to quantify the number of intracellular bacteria. Intracellular bacterial load is represented as CFU per ml. All experiments were done on at least three independent occasions.
Total RNA was isolated from lung homogenates and cells using Isol-RNA Lysis Reagent (5 Prime) or Trizol according to the manufacturer’s protocol. DNase digestion was performed using 10 U of recombinant DNase I (Roche). RNA was PolyA primed with Oligo (dT)18 primers (Eurofins Genomics) and reverse-transcribed using Mu-MLV reverse transcriptase (Fermentas). qPCRs were run on a Realplex Mastercycler (Eppendorf) or Stratagene Mx3005P qPCR System and cDNA was quantified by SYBR Green method using HOT FIREPol EvaGreen qPCR supermix (Medibena) or KAPA SYBR FAST qPCR Kit. mRNA expression of the housekeeping gene Hprt was used for normalization purposes. Primers for qPCR were (in 5’-3’ orientation):
Hprt fwd-GCAGTCCCAGCGTCGTGAT, rev-CAGGCAAGTCTTTCAGTCCTGTC
Il12b (p40) fwd-ACAGCACCAGCTTCTTCATCAG, rev-TCTTCAAAGGCTTCATCTGCAA
Ifng fwd-CGGCACAGTCATTGAAAGCC, rev-TGTCACCATCCTTTTGCCAGT
Il1b fwd-AGATGAAGGGCTGCTTCCAAA, rev-AATGGGAACGTCACACACCA
Ifnb fwd- TCAGAATGAGTGGTGGTTGC, rev- GACCTTTCAAATGCAGTAGATTCA
Tnf fwd-GATCGGTCCCCAAAGGGATG, rev-CACTTGGTGGTTTGCTACGAC
Cxcl1 fwd-TGCACCCAAACCGAAGTCATAG, rev-TTGTATAGTGTTGTCAGAAGCCAGC
Il10 fwd-GGACTTTAAGGGTTACTTGGGTTGCC, rev-CATGTATGCTTCTATGCAGTTGATGA
Mx1 fwd-GACTACCACTGAGATGACCCAGC, rev- ATTTCCTCCCCAAATGTTTTCA
Ifit1 fwd-CAGGTTTCTGAGGAGTTCTG, rev-TGAAGCAGATTCTCCATGAC
Isg15 fwd-GGGGCCACAGCAACATCTAT, rev-CGCTGGGACACCTTCTTCTT
Cxcl10 fwd-TGCGAGCCTATCCTGCCCACGTG, rev-CCGGGGTGTGTGCGTGGCTTCA
Ifnar1 fl/fl fwd-GCCCTGCTGAATAAGACCAG, rev-ACTGGCCTCAAACTCACTGC
To prepare whole lung cell suspensions, lungs were cut to pieces with scissors and digested in RPMI containing 10% FCS, 1 mg/ml collagenase I (Roche) and 0.25 mg/ml DNase I (Roche) for 1 h at 37°C with agitation. Single-cell suspensions were obtained by flushing the samples through 70 μm strainer. Red blood cells were lysed using hypotonic shock and washed twice with PBS. To exclude dead cells, samples were stained with FVD eFluor 506 (eBioscience), prior to Fc blocking with anti-CD16/CD32 (2.4G2, BD). Suspensions were stained for cell surface proteins and intracellular IFN-γ in appropriate combinations of following monoclonal antibodies conjugated to allophycocyanin-eFluor 780, allophycocyanin, brilliant violet 711, phycoerythrin, brilliant violet 421, phycoerythtrin-cyanine7, peridinin chlorophyll protein-cyanine 5.5 and fluorescein isothiocyanate: anti-CD45 (30-F11), anti-CD11c (HL3), anti-SiglecF (E50-2440), anti-Ly6G (1A8), anti-Ly6C (HK1.4), anti-CD11b (M1/70), anti-CD3 (11-26c(11–26)), anti-CD8 (53–6.7), anti-NK1.1 (PK136) and anti IFN-γ (XMG1.2), purchased from BD, eBioscience and Biolegend. To prepare cells for intracellular staining, suspensions were incubated in fixation/permeabilization buffer (eBioscience), followed by Fc blocking, washing and staining in permeabilization buffer (eBioscience). Dead cells were excluded based on their light-scattering characteristics and FVD staining. Cell doublets were excluded based on FSC-H/FSC-A and SSC-H/SSC-A. All data acquisitions were performed using LSR Fortessa II (BD) cytometer interfaced with FACSDiva. FlowJo X (Tree Star) software was used for data analysis and graphical representation.
Mice were either infected or given PBS and euthanized 24 h post-treatment. Lungs were harvested, and dissociated using VDI 12 tissue homogeniser (VWR) in sterile PBS. Single-cell suspensions were obtained by flushing the samples through 70 μm strainer. To exclude dead cells, samples were stained with FVD eFluor 506 (eBioscience), prior to Fc blocking with anti-CD16/CD32 (2.4G2, BD). To sort out CD45+CD11chighSiglecF+ alveolar macrophages, cell suspensions were stained for cell surface proteins using monoclonal antibodies conjugated to allophycocyanin-eFluor 780, allophycocyanin and brilliant violet 421: anti-CD45 (30-F11), anti-CD11c (N418) and anti-SiglecF (E50-2440), purchased from BD Bioscience, eBioscience and Biolegend. RNA from sorted alveolar macrophages was extracted using Power SYBR Green Cells-to-CT Kit (4402954, Ambion) according to manufacturer’s instructions.
Splenic NK cells from WT mice were isolated by magnetic bead labeling (CD49b+ DX5 microbeads) following manufacturer’s instructions (Miltenyi Biotec). Before magnetic separation, 1 x 107 cells were labeled with CellTrace CFSE kit (Thermo) according to the manufacturer’s protocol. Labeled NK cells (1 x 106 cells/mouse) were adoptively transferred intravenously into ifnar1-/- mice, which were infected intranasally with 3 x 105 K. pneumoniae 52.145 CFU. 24 h post infection mice were euthanized and bacterial loads in lungs determined. Intracellular IFN-γ was detected using anti IFN-γ (XMG1.2). Cells were analyzed using a FACSCantoII flow cytometer and FlowJo software (Tree Star).
For IFN-γ rescue analysis, Ifnar1-/- mice were infected with standard inoculum, and immediately post-infection intranasally given 20 μl of PBS, or 100 ng rIFN-γ in 20 μl of PBS (recombinant Mouse IFN-gamma Protein, R&D, Carrier Free, cat. number 485-MI-100, LOT#CFP2516032). Mice were euthanized 24 h p.i. and bacterial loads and mRNA expression in lungs were determined.
Mice were anesthetized and inoculated intranasally with 30 μl PBS containing 30,000 U carrier-free recombinant mouse IFN-β (PBL Interferon Source, LOT#6450). Control animals received 30 μl of PBS. 6 h post-treatment animals were euthanized and total RNA was isolated from the lungs.
Data analysis, statistical testing and visualization was performed with Prism 6 (GraphPad Software) using Log-rank Mantel-Cox test, Mann-Whitney test, unpaired two-tailed Student’s t test and One-way ANOVA with multiple comparisons, as indicted in the figure legends. Medians are depicted as horizontal bars, and means are depicted as horizontal bars ± SEM. Statistical significance is indicated as follows: ns (not significant), P > 0.05, *, P < 0.05; **, P < 0.01; ***, P < 0.001.
|
10.1371/journal.pgen.1007822 | De novo variants in congenital diaphragmatic hernia identify MYRF as a new syndrome and reveal genetic overlaps with other developmental disorders | Congenital diaphragmatic hernia (CDH) is a severe birth defect that is often accompanied by other congenital anomalies. Previous exome sequencing studies for CDH have supported a role of de novo damaging variants but did not identify any recurrently mutated genes. To investigate further the genetics of CDH, we analyzed de novo coding variants in 362 proband-parent trios including 271 new trios reported in this study. We identified four unrelated individuals with damaging de novo variants in MYRF (P = 5.3x10-8), including one likely gene-disrupting (LGD) and three deleterious missense (D-mis) variants. Eight additional individuals with de novo LGD or missense variants were identified from our other genetic studies or from the literature. Common phenotypes of MYRF de novo variant carriers include CDH, congenital heart disease and genitourinary abnormalities, suggesting that it represents a novel syndrome. MYRF is a membrane associated transcriptional factor highly expressed in developing diaphragm and is depleted of LGD variants in the general population. All de novo missense variants aggregated in two functional protein domains. Analyzing the transcriptome of patient-derived diaphragm fibroblast cells suggest that disease associated variants abolish the transcription factor activity. Furthermore, we showed that the remaining genes with damaging variants in CDH significantly overlap with genes implicated in other developmental disorders. Gene expression patterns and patient phenotypes support pleiotropic effects of damaging variants in these genes on CDH and other developmental disorders. Finally, functional enrichment analysis implicates the disruption of regulation of gene expression, kinase activities, intra-cellular signaling, and cytoskeleton organization as pathogenic mechanisms in CDH.
| Congenital diaphragmatic hernia (CDH) is a life-threatening condition affecting about 1 every 3000 newborns. Although the role of genetics in the pathogenesis of CDH has been well established, only a handful of disease genes have been identified so far. We and other have previously shown that de novo variants, those carried by the cases but not inherited from parents, are enriched in sporadic CDH cases consistent with their negative effects on reproductive fitness. To further investigate the genetics of CDH, we analyzed de novo variants in 362 proband-father-mother trios from whole exome or genome sequencing data and identified four patients carrying damaging variants in MYRF, a membrane associated transcription factor that is highly expressed in developing diaphragm and heart. We then ascertained a total of 12 patients with MYRF de novo variants, and found they shared common phenotype characteristics including congenital abnormalities in diaphragm, heart and reproductive organs. The high rate of recurrence and similar phenotypic manifestations suggest that de novo variants of MYRF have pleiotropic effects and cause a novel syndrome. The identified new gene is reminiscent of previously identified CDH genes (e.g., GATA4, GATA6, NR2F2, ZFPM2, and WT1) that are also associated with other developmental disorders. Indeed, we found in our cohort more than 20 damaging de novo variants in genes implicated in other developmental disorders but not previously linked to CDH. The overlap was unlikely to occur by chance and can be best explained by their pleiotropic effects. We also showed that, despite the shared genetic basis with other disorders, damaging de novo variants in CDH as a whole were enriched in specific functional pathways that recapitulated our current knowledge about diaphragm development. So additional candidate genes can be prioritized based on the genetic pleiotropy and functional specificity. The findings have general implications in design and analysis in genetic studies of rare birth defects.
| Congenital diaphragmatic hernia (CDH) is a severe developmental disorder affecting 1 in 3000 live births [1, 2]. It is characterized by defects in diaphragm that allow the abdominal viscera to move into the thoracic cavity and is associated with pulmonary hypoplasia and in some cases pulmonary hypertension. CDH can be isolated (50–60%) or associated with anomalies in other organs including the heart, brain, kidneys and genitalia [3, 4]. Despite advances in treatment, mortality rate remains high [5, 6]. A better understanding of the causative factors for CDH may inform disease prevention and treatment.
The genetic contribution to CDH has been established by familial aggregation [7], rare monogenic disorders associated with CDH in humans [8], chromosome abnormalities [9], copy number variations [10–12], and mouse models [13]. However, our understanding of the genetic basis of CDH is still rudimentary. The historically low reproductive fitness of individuals with CDH led to the hypothesis that de novo variants with large effect sizes may explain a fraction of CDH patients as in other developmental disorders [14, 15]. We and others have previously reported an enrichment of damaging variants in sporadic CDH patients [16, 17]. However, no recurrently mutated gene was identified in our genome wide analyses due to the limited sample size.
To continue the search for new CDH genes, we performed whole exome (WES) or whole genome sequencing (WGS) of 271 new trios. Combined with previously published WES data [16, 17], we analyzed all 362 trios. We confirmed the overall burden of damaging de novo variants and identified a new disease gene recurrently mutated in cases with similar syndromic features. To prioritize additional risk genes, we analyzed cross-disorder overlap and pathway enrichment. The results provide insights into the genetic architecture of CDH and suggest additional candidate genes.
Patients were recruited from the multicenter, longitudinal DHREAMS (Diaphragmatic Hernia Research & Exploration; Advancing Molecular Science) study [11]. We excluded patients with known genetic causes from clinical karyotype or chromosome microarray or with a family history of CDH. WES was performed on 118 proband-parents trios, a subset (39) of whom were published previously [17]. WGS was performed on 192 trios including 27 without damaging variants from the previous study [17]. On average, 91% of coding regions in WES samples and 98% in WGS samples were covered by 10 or more unique reads (S1 Fig). WGS showed more uniform distribution of sequencing depth that contributes to higher power in detecting coding variants [18, 19]. For the 27 overlapping samples, 12 additional de novo coding variants were identified in WGS including 10 not included in the exome targets or with low depth of coverage and two that failed stringent QC filters in our previous study.
Combined with trios collected by Boston Children’s Hospital/Massachusetts General Hospital (BCH/MGH) [16], we analyzed a total 362 unique trios (S1 Table). Clinical and demographic information of patients are given in S1 Data. In the combined cohort, there were 212 (58.6%) male and 150 (41.4%) female patients. The male-to-female ratio (1.4:1) was consistent with published retrospective and prospective cohorts [20, 21]. The most common type of CDH was left-sided Bochdalek; rare forms of CDH or atypical lesion sides were also included (Table 1).
A total 149 (41.2%) cases had additional congenital anomalies or neurodevelopmental disorders (NDD) at the time of last follow up and were classified as complex cases; and 209 (57.7%) patients had no additional anomalies at last contact were classified as isolated cases. The most frequent comorbidity among complex cases was cardiovascular anomalies (44.3%). NDD, skeletal malformations, and genitourinary defects were also observed in complex cases (Table 1).
We identified 471 coding de novo variants in 264 (72.9%) cases including 430 single nucleotide variants (SNV) and 41 indels. Transition-to-transversion ratio of de novo SNVs was 2.64. The number of de novo coding variants per proband closely followed a Poisson distribution, with an average of 1.32 in WGS trios and 1.28 in combined WES trios (S2 Fig). Variants that were likely gene disrupting (LGD) or predicted deleterious missense (“D-mis” defined by CADD score [22] ≥25) were considered as damaging. A total of 193 damaging variants (57 LGD and 138 D-mis) were identified in 150 (41.4%) cases, including 38 (10.5%) cases harboring two or more such variants. Compared with the baseline expectations (Material and methods) [23], both de novo LGD variants (0.16 per case) and D-mis variants (0.38 per case) were significantly enriched in cases (fold enrichment (FE) = 1.73, P = 8.6x10-5 by one-sided Poisson test for LGD; FE = 1.5, P = 1.1x10-6 for D-mis) while the frequency of silent variants closely matched the expectation (0.30 per case, FE = 1.01, P = 0.48 by one-sided Poisson test).
Consistent with the previous study [16], damaging variants showed a higher enrichment in complex cases than isolated cases (FE = 1.70 vs 1.64 for LGD, 1.61 vs 1.38 for D-mis; S2 Table); and the proportion of complex cases who carried damaging variants was higher than isolated cases (43.6% vs. 39.4%). Burden of damaging variants was also higher in female than male cases (FE = 2.09 vs 1.47 for LGD, 1.63 vs 1.36 for D-mis; S2 Table), supporting a “female protective model” similar to autism and other NDD with male bias [24, 25].
Recent studies highlighting the use of large population reference sequencing data in interpreting LGD variants has demonstrated that genes depleted of LGD variants in the general population were more likely associated with disorders with reduced reproductive fitness[26]. We defined constrained genes by the estimated probability of loss-of-function intolerance (pLI) [27] ≥0.5 and found the burden of LGD variants was largely explained constrained genes (Table 2). D-mis also showed a higher enrichment in constrained genes (Table 2).
We identified eight genes affected by more than one de novo LGD or missense variant (S3 Table). The top ranked gene, MYRF, has one frameshift insertion and three damaging missense variants, all of which were validated by Sanger sequencing. It is the only constrained gene in the list. By comparing with baseline expectations, only MYRF reaches genome-wide significance after Bonferroni correction of ~20000 coding genes (P = 5.3x10-8 <0.01/20000, by one-sided Poisson test).
Notably, all four patients with MYRF variants also had congenital heart disease (CHD), and three of them had genital anomalies including blind-ending vagina in a female and ambiguous genitalia or undescended testes in two male cases (Table 3). By screening another 220 CDH trios collected by the DHREAMS study, we identified another patient harboring a de novo splice acceptor site variant. The female patient had a diagnosis of Scimitar syndrome (a complex form CHD). She also had a monozygotic twin sister with hypoplastic left heart syndrome who also carried the same variant but no known CDH.
Given the strong association of MYRF variants with CHD, we then searched for de novo variants from a recently published study of CHD conducted by Pediatric Cardiac Genomics Consortium (PCGC) [29] and identified three additional de novo missense variants in MYRF from 2645 trios. All CHD patients also had genitourinary anomalies, including a patient with Swyer syndrome (46XY karyotype with female reproductive organs). One CHD patient with the Q403H variant had hemidiaphragm eventration. Recently, Pinz et al. [30] and Chitayat et al [31] reported three additional cases with complex CHD who carried de novo LGD variants in MYRF. All cases had genital defects, and one had CDH and the other two had pulmonary hypoplasia. Furthermore, from clinical WES, we also identified a Swyer syndrome patient with a stop-gain variant in MYRF who had dextrocardia and pulmonary hypoplasia.
In total, we identified 13 patients harboring 12 different de novo functional variants in MYRF (6 LGD and 6 missense variants; Fig 1A). All patients had CHD; and excluding those who died in infancy and had incomplete phenotypic information, all patients also had genitourinary anomalies. CDH was present in 7 out of 12 patients, and diaphragm defects were not systematically evaluated in cases without reported CDH. There was no clear phenotypic difference between patients with LGD variants and those with missense variants (Table 3). Taken together, the unique association of CDH and similar non-diaphragm defects including CHD, Scimitar syndrome, genitourinal anomalies and sex reversal in 46XY patients with de novo variants in MYRF establish it as a new syndromic CDH gene.
MYRF is a highly constrained gene in the population (pLI = 1). By examining both public databases (ExAC and gnomAD) and our own cohort, we only identified two rare LGD variants that affect all functional isoforms, yet their functional consequences were not clear (S5 Table). We also searched for inherited variants in 362 CDH trios and 2645 CHD trios from PCGC but did not find any inherited LGD variants in probands. Enrichment for de novo LGD variants associated with CDH and near complete absence of loss-of-function variants in the general population suggest that variants causing loss of MYRF function are likely fully penetrant for one or more aspects of this syndrome. All six de novo missense variants identified patients were also absent from the public databases, consistent with their high penetrance as LGD variants in this gene.
MYRF is a membrane-associated transcription factor that plays a pivotal role in oligodendrocyte differentiation and myelination [32, 33]. Although it has not previously been implicated in diaphragm or cardiac development, its expression level was ranked at the top 21% of genes expressed in mouse developing diaphragm at E11.5 [34] and top 14% in developing heart at E14.5 [35].
The MYRF protein has two functional isoforms. Both isoforms contain a N-terminal proline-rich region followed by a DNA binding domain (DBD), which can be cleaved from the membrane by a region called intramolecular chaperon auto-processing (ICA) domain. All frameshift and stop gained variants resulted in truncated protein products in both functional isoforms and may trigger non-sense mediated decay. The precise functional effects of splice site variants were not evaluated, but are predicted to cause exon skipping, intron retention or activation of cryptic splice site and also result in a truncated protein. All six missense variants aggregated in the two DBD and ICA functional domains (Fig 1a). The missense variants were predicted as deleterious by a majority of bioinformatics tools (S4 Table). Most of the affected amino acid residues are highly conserved across species (S3 Fig).
MYRF DBD is homologous to yeast transcriptional factor Ndt80 but MYRF can only function as a trimer [36]. All missense variants in this domain are located in a region depleted of missense variants in the population (observed/expected = 0.31; Fig 1A) and have high MPC scores [37] (S4 Table). Protein structure modeling predicted that those variants may affect DNA binding affinity (F387S), change surface charge distribution (Q403H), or destabilize the protein structure (G435R and L479R) (S4 Fig).
Previous studies also showed that full length MYRF forms a trimer before cleavage, and trimerization is required for auto-cleavage and subsequent activation [38]. The ICA domain which is distantly related to bacteriophage’s tailspike protein was believed to play an essential role in MYRF trimerization. Two missense variants (V679A, R695H) are located at the C-terminal end of the ICA domain where the triplet helix bundle is formed [39]. V679 is one of the critical residues in ICA that is fully conserved from human to bacteriophage (S3 Fig). Structure modeling predicted that the variant R695H may destabilize the trimer structure (S4 Fig) and would fail to produce functional MYRF DBD trimers by trimerization-dependent auto-proteolysis.
To evaluate the effect of MYRF variants on gene expression, we performed RNA-seq on diaphragm fibroblast cell cultures from neonatal patients. After removing outlier samples (S5 Fig), we obtained transcriptome data of 31 patients including three with a de novo MYRF variant (one frameshift insertion and two missense variants in the ICA domain). Most patients (27/31, 87%) included in the RNA-seq analysis were self-reported non-Hispanic White. Additionally, we identified 74 putative MYRF target genes from a previous study of rat oligodendrocyte progenitor cells (S3 Data) [40]. Gene expression levels were quantified as TPM (transcripts per million mapped reads). The z-scores of expression levels of putative MYRF target genes were systematically shifted down in MYRF mutant cells (P = 2.4E-7 by Kolmogorov-Smirnov test; Fig 1B), consistent with the reduced transcription factor activities caused by the damaging variants. We quantified differential expression (DE) of genes between samples with and without de novo MYRF variants by a shrinkage estimator of fold change [41]. Selected DE genes were validated by quantitative polymerase chain reaction (qPCR) on the same cell cultures (S7 Fig). Using gene set enrichment analysis [42] of genes ranked by the fold changes, putative MYRF target genes are significantly enriched among the down-regulated genes (normalized enrichment score (NES) = -2.10, P<5.0E-4; Fig 1C). Since all MYRF mutation carriers were males, we repeated the analysis using only males and found the results are similar as using all samples (NES = -1.95, P<5.0E-4), suggesting that sex is not a confounding factor. The patient with the MYRF frameshift variant was the only MYRF mutation carriers whose ethnicity was not self-reported White. The enrichment of MYRF target genes is also observed in genes down-regulated in the two samples with missense variants (S6 Fig), suggesting that the result was not driven by the LGD variant or ethnicity.
Manual inspection of top DE genes (S4 Data) revealed that GATA4, a known CHD gene that has also been implicated in familial and sporadic CDH [43], was significantly down-regulated in cases with de novo MYRF variants (estimated fold change = 0.54, q-value = 0.03). Interestingly, we observed that expression trajectories of MYRF and GATA4 were similar in mouse developing diaphragm and lung (S8 Fig) suggesting that they play similar functional roles during diaphragm and pulmonary development.
Besides MYRF, we estimated there were 64 (95% CI: 38–93) genes with de novo variants implicated in CDH based on the overall burden analysis. Most of those genes have only one damaging variant in the cohort. To prioritize among all the genes with de novo damaging variants, we took two approaches.
We noted that CHD was the most common non-diaphragm defect in complex cases (Table 1). Damaging mutations in MYRF have been identified in a previous CHD study but the gene did not reach genome-wide significance [29]. The identification of the MYRF syndrome suggested that the comorbidity of CHD and CDH in some cases can be explained by the same genetic factors, many of which remain to be discovered. CDH is also part of the phenotype spectrum of several rare Mendelian disorders [8]. Recently discovered genes for developmental disorders are often pleiotropic and implicated in multiple diseases [15, 29, 44]. Thus, the finding of MYRF motivated us to assess the genetic overlap between CDH and other developmental disorders, especially CHD, to help us prioritize additional CDH genes with pleiotropic effects. To this end, we curated genes that were known or implicated in CHD and other developmental disorders (S5 Data; Materials and Methods). Hereafter we refer to these known or candidate genes as CHD or DD genes.
In addition to MYRF, we identified a total of 26 DD/CHD genes with damaging de novo variants in 25 CDH patients (Fig 2A). Using a simulation approach that accounted for the number of variants, gene size, and sequence context (Materials and Methods), we found that damaging variants in CDH were significantly enriched in the DD and CHD genes (Fig 2B). For example, we observed 6 CHD genes with de novo LGD variants in CDH which was 4.7-fold higher than expected (P = 1.7x10-3); the number of DD genes with de novo LGD variants (8) was 3.4 folder higher than expected (P = 2.3x10-3). Among CHD genes with at least one damaging variant in CDH, haploinsufficiency of WT1 is a known cause of several syndromic forms of CDH [8]; ZFPM2 and GATA6 have already been established as CDH genes by previous studies [45, 46]. However, the enrichment of damaging variants and especially LGD variants remained significant after excluding known or candidate CDH genes [47] (S9 Fig). Furthermore, the enrichment cannot fully be explained by the over-representation of constrained genes, because the enrichment persisted after conditioning on all constrained genes and remained significant for LGD variants (S9 Fig).
The cross-disease overlap suggests that pleiotropic effects of variants in the genes associated with other developmental disorders are also associated with CDH in a fraction of cases. Since CHD genes were curated based on the damaging mutations in CHD patients and DD genes were mostly implicated in other developmental disorders, the genes that appear in both sets were more likely to participate in a broader range of developmental process. Accordingly, the enrichment in genes found exclusively in one set was significantly reduced (Fig 2B, S9 Fig).
We reviewed the most recent medical records of those patients (S7 Table) and identified six complex cases with CHD and/or NDD compatible with the initial reported phenotypes for these genes. Two additional cases were found to have non-CHD cardiovascular defects like two-vessel cord or dilated aortic root; and another four had mild-to-moderate developmental delay/intellectual disability at latest evaluation. Four patients who carried LGD variants in known DD genes (POGZ, ARID1B, FOXP1, and SIN3A) and one patient who carried a known activating variant in the Noonan syndrome gene PTPN11 were considered pathogenic variants by the American College of Medical Genetics and Genomics guidelines [48].
Pleiotropy was further supported by the gene expression data. The majority of the 26 DD/CHD genes with damaging de novo variants in CDH were not only highly expressed in mouse developing diaphragm but also in developing heart or brain (Fig 2C). Indeed, over all coding genes, expression ranks in the three developing organs were highly correlated (Spearman rank correlation r = 0.74 between diaphragm and heart, 0.74 between diaphragm and brain). Therefore, high diaphragm expression can be a proxy for a pleiotropic effect. Consistent with this, we found that all damaging de novo variants in complex cases, presumed to enrich causative variants affecting multiple organs, were greatly enriched in genes at the top quartile of expression in developing diaphragm (FE = 4.6, P = 7.9x10-7 by one-sided Poisson test for LGD; FE = 2.4, P = 1.8x10-4 for D-mis). By contrast, in isolated cases, the enrichment of damaging variants was distributed in genes across a broad range of expression (Fig 3).
As a second approach to prioritize CDH genes, we hypothesized that different CDH genes converge onto a small number of pathways, and novel genes in the enriched pathways could be candidates for new disease genes. We evaluated functional enrichment of genes affected by damaging de novo variants to identify biological processes involved in CDH. To boost the signal, only constrained genes or known haploinsufficient genes were included in the pathway analysis (Materials and Methods). A total of 63 Gene Ontology Biological Process gene sets were enriched at a false discovery rate (FDR) of 0.1 (S6 Data). To remove the redundancies between gene sets, we used a similarity score to organize functionally related gene sets into a network. The resulting network was annotated and visualized as a functional enrichment map (Fig 4A). Eleven functional groups were identified that recapitulated our current knowledge about the molecular genetic basis of CDH [49]. They were supported by 48 genes including 27 novel genes (Fig 4B).
Transcription factor haploinsufficiency is an established cause of CDH [50] and other birth defects [51]. Recently, disruption of epigenetic machinery was also found to underlie many developmental disorders [35, 44, 52]. The majority of DD/CHD genes directly or indirectly regulate gene expression which formed a highly connected cluster of enriched gene sets, some of the transcription factors are involved in the development of heart, lung and reproductive organs. We identified nine novel genes encoding transcription factors or histone modifiers.
Proper cell migration is critical during diaphragm development. Initially, mesenchymal precursor cells migrate from mesoderm to form the primordial diaphragm. After that, pleuroperitoneal folds of the primordial diaphragm become the targets of migration of muscle progenitors, where they undergo myogenesis and morphogenesis [53]. Several related pathways were implicated including cellular response to growth factors or stress events that initiate directional migration [54], actin cytoskeletal organization and cell-cell junction assembly that drive and fine tune cell movement [55, 56]. Gene sets in protein phosphorylation and JUN-MAPK (mitogen-activated protein kinase) cascades were also enriched but not entirely due to three Noonan syndrome genes (PTPN11, BRAF, RAF1). The enrichment in kinase activity related pathways was supported by six novel kinase genes that overlapped with intracellular signaling functions. One kinase gene, MAPK8IP3, has been implicated in lung development in a mouse model [57].
In this study, by analyzing de novo coding variants in CDH, we confirmed the overall enrichment of damaging de novo variants and identified MYRF as a new syndromic CDH gene. All our CDH cases with MYRF mutations also had CHD and most of them had genitourinary defects. The striking phenotypic similarities among the cases suggest that damaging de novo variants of MYRF disrupt the function of progenitor cells of developing diaphragm, heart and reproductive organs. In this novel MYRF syndrome, all cases with disease associated variants had CHD including three with Scimitar syndrome, whereas penetrance CDH was incomplete. It suggests that the manifestation of CDH in this syndrome depends on other genetic, environmental, or stochastic factors. The monozygotic twin case discordant for CDH supports that stochastic developmental events are involved.
MYRF is well known for its function in regulating myelination of the central nervous system [32]. A mouse model with conditional deletion of MYRF in oligodendrocyte precursors has abnormal motor skill [58]. Recently, an inherited missense variant in MYRF (Q403R) has been reported as the cause of encephalopathy with reversible myelin vacuolization in a Japanese pedigree [59]. This variant is located at the same residue as the de novo missense variant in one of the PCGC cases but with a different substutition (Q403H). No other congenital defects were reported for the variant carriers in that family. The Q403R variant has been experimentally shown to diminish the transcription activity of a target gene [59], similar to our finding in two other missense variants (S6 Fig). Why the two different substitutions at the same amino acid position result in different phenotypes remains to be elucidated in the future. Among patients with de novo damaging variants in MYRF, one individual with the R695H variant also had intellectual disability and delayed motor skills (Table 3).
We identified 25 other individuals harboring damaging de novo variants in known or candidate DD/CHD genes, most of which have not been reported to be associated with CDH before. The significant enrichment of damaging variants among DD/CHD genes strongly suggest their causative role for majority of these cases. Similar to the case of MYRF, many DD/CHD genes have yet to be established as known disease genes. The enrichment of CDH damaging variants support their possible involvement in a broader range of developmental abnormalities which should be further evaluated in additional case cohorts with other congenital anomalies. Some recent studies of other congenital anomalies and developmental disorders have already provided further evidence for a few putative DD/CHD genes. For example, a damaging missense variant in LAMA5, a gene that plays a role in the maintenance and function of the extracellular matrix critical for pattern formation during development [60], was associated with multi-system syndrome in an Italian family [61]. Duplication of STAG2, which encodes a subunit of cohesin complex, was associated with intellectual disability and behavioral problems [62]. MEIS2 was previously nominated as a potential CDH candidate by transcriptome analysis [34] and encodes an interaction partner of transcription factor gene PBX1, haploinsufficieny of which has recently been associated with multiple developmental defects including CDH [63].
Since our knowledge of DD/CHD genes is incomplete, it is possible that this observed genetic overlap represents only the tip of an iceberg. Our pathway analysis not only captured general biological process during developmental, but also identified pathways that are closely related to diaphragm development. Some novel genes prioritized by the pathway analysis have also been supported by new genetic data in other disorders. For example, de novo copy number loss or missense variants in TAOK2, one of the kinase gene implicated by the enriched gene sets of the kinase activity and MAPK signaling, has been demonstrated to cause autism and other NDD [64]. Because CDH is a relatively uncommon and lethal condition as are many other rare congenital anomalies, it is difficult to recruit large numbers of patients for genetic studies. The findings from this and other studies [15] suggest that cross-disorder analysis can be a powerful strategy for future gene discovery.
The genetic overlap between CDH and other disorders is consistent with pleiotropy among developmental disorder genes and is further supported by the highly correlated gene expression levels in multiple developing organs. We also showed that different enrichment patterns of de novo damaging variants between complex and isolated CDH cases is consistent with the hypothesis that variants in complex cases affect genes with more pleiotropic effects.
The pleiotropic effects of genes during development also suggest that our current classification of “isolated” cases may understate their non-diaphram abnormalities. A limitation of our study is the lack of long term clinical outcome data on many of the patients since our cohort is still relatively young. Examining the most recent medical records of patients with variants in DD/CHD genes revealed mild-to-moderate cadiovascular or NDD symptoms in several cases initially classified as isolated at birth (S7 Table). The medical records were often incomplete for patients who died at early infancy or were lost to follow-up (Table 1), and it is likely that NDD outcome in many isolated patients were underestimated [65, 66]. Furthermore, almost all isolated cases also had pulmonary hypoplasia. Traditionally it was assumed that lung defects were caused by the mechanical compression by the herniated visceral, but it is clear now that development of lung and diaphargm are two intricatelly connected developmental processes [67], and lung defects may share common etiologies with CDH [68]. Among MYRF variant carriers, four patients who did not have diaphragm defects developed pulmonary hypoplasia (Table 3), further supporting common genetic control of these two processes. Larger cohorts with more detailed neurodevelopmental and long term outcomes will enhance our ability to identify additional CDH genes and provide accurate prognostic information to families to allow for future clinical diagnosis of these conditions.
In summary, our analysis of de novo coding variants in 362 CDH trios identified a new disease gene MYRF, revealed genetic overlap with other developmental disorders, and identified biological processes important for diaphragm development. Future studies will beneifit from larger sample sizes, analyzing different types of genetic variants, leveraging the information from other developmental disorders, and integrating functional genomic data.
Study subjects were enrolled by the DHREAMS study (http://www.cdhgenetics.com/). Neonates, children and fetal cases with a diagnosis of diaphragm defects were eligible for DHREAMS. Clinical data were collected from the medical records by study personnel at each of 16 clinical sites. A complete family history of diaphragm defects and major malformations was collected on all patients by a genetic counsellor. A blood, saliva, and/or skin/diaphragm tissue sample was collected from the patient and both parents. All studies were approved by local institutional review boards, and all participants or their parents provided signed informed consent.
Cases without known pathogenic chromosome abnormalities or copy number variations [11] were selected for exome or whole-genome sequencing. A total of 283 trios with no family history of CDH with three generation and not born to consanguineous marriages were included in the current study. De novo coding variants on a subset trios (n = 39) have been described in our previous study [17]. In Neonates cohort, longitudinal follow-up data including Bayley III and Vineland II developmental assessments since discharge at 2 years and/or 5 years of age were gathered. Patients were evaluated to have developmental delay if at least one of the composite scores was 2 standard deviations below population average.
Patients with additional birth defects or developmental delay or other neuropsychatric phenotypes at last contact were classified as complex, and otherwise as isolated. Pulmonary hypoplasia, cardiac displacement and intestinal herniation were considered to be part of the diaphragm defect sequence and were not considered to be additional birth defects.
Subjects of BCH/MGH cohort were enrolled in “Gene Mutation and Rescue in Human Diaphragmatic Hernia” study as described previously [16]. Among 87 trios from BCH/MGH cohort, 8 trios were found to be duplicates with DHREAMS trios and were excluded from the analysis.
Exome sequencing was performed in 79 trios that were not published before. Eleven trios were processed at the New York Genome Center. The DNA libraries were prepared using the Illumina TruSeq Sample Prep Kit (Illumina). The coding exons were captured using Agilent SureSelect Human All Exon Kit v2 (Agilent Technologies). Samples were multiplexed and sequenced with paired-end 75bp reads on Illumina HiSeq 2500 platform according to the manufacturer’s instructions. Sixty-eight trios processed at University of Washington Northwest Genome Center were captured using NimbleGen SeqCap EZ Human Exome V2 kit (Roche NimbleGen), and sequeced on HiSeq 4000 in 75 bp paired-end reads.
Another 192 trios were processed at Baylor College of Medicine Human Genome Sequencing Center using whole genome sequencing as part of the Gabriella Miller Kids First Pediatric Research Program. Among these, 27 trios were included in the previous exome study [17] but had no damaging de novo variants. Genomic libraries were prepared by the Illumina TruSeq DNA PCR-Free Library Prep Kit (Illumina) with average fragment length about 350 bp, and sequenced as paired-end reads of 150-bp on Illumina HiSeq X platform.
Exome and whole-genome sequencing data were processed using an inhouse pipeline implementing GATK Best Practice (version 3). Briefly, reads were mapped to human genome reference (GRCh37) using BWA-mem (version 0.7.10); duplicated reads were marked using Picard (version 1.67); variants were called using GATK (version 3.3–0) HaplotypCaller to generate gVCF files for joint genotyping. All samples within the same batch were jointly genotyped and variant quality score recalibration (VQSR) was performed using GATK. Common SNP genotypes within exome regions were used to valid parent-offspring relationships using KING (version 2.0) [69].
A variant that was presented in the offspring and had homozygous reference genotypes in both parents was considered to be a potential de novo variant. We used a series of stringent filters to identify de novo variants as described previously[70]. Briefly, we first kept variants that passed VQSR filter (tranche≤99.8 for SNVs and ≤99.0 for indels) and had GATK’s Fisher Strand≤25, quality by depth≥2. Then we required the candidate de novo variants in proband to have ≥5 reads supporting alternative allele, ≥20% alternative allele fraction, Phread-scaled genotype likelihood ≥60 (GQ), and population allele frequency ≤0.1% in ExAC; and required both parents to have > = 10 reference reads, <5% alternative allele fraction, and GQ≥30.
We used ANNOVAR [71] to annotate functional consequence of de novo variants on GENCODE (v19) protein coding genes. All coding de novo variants were manually inspected in the Integrated Genomics Viewer (http://software.broadinstitute.org/software/igv). A total of 169 variants were selected for validation using Sanger sequencing; all of them were confirmed as de novo variant. The number of coding de novo variants per proband was compared with expectations under Possion distribution.
All coding variants were classified as silent, missense, inframe, and likely-gene-disrupting (LGD, which includes frameshift indels, canonical splice site, or nonsense variants). The most severe functional effect was assigned to each variant. We defined deleterious missense variants (D-mis) by phred-scaled CADD (version 1.3) [22] score≥25.
Baseline rate for different classes of de novo variants in each GENCODE coding gene were using a previously described mutation model [23, 70]. Briefly, the tri-nucleotide sequence context was used to determine the probability of each base in mutating to each other possible base (precomputed rates are available at: https://github.com/jeremymcrae/denovonear/blob/master/denovonear/data/rates.txt). Then, the mutation rate of each functional class of point mutations in gene was calculated by adding up point mutation rates in the longest transcript. The rate of frameshift indels was presumed to be 1.1 times the nonsense mutation rate. The expected number of variants in different gene sets were calculated by summing up the class-specific variant rate in each gene in the gene set mutiplied by twice the number of patients (and if the gene is located on the non-pseudoautosomal region of chromsome X, further adjusted for female-to-male ratio [14]). The observed number of variants in each gene set and case group was then compared with the baseline expectation using Poisson test.
In burden analysis, constrained genes were defined by pLI metrics [27] ≥0.5 which include a total of 5451 GENCODE genes, and all remaining genes were treated as other genes. We used a less stringent pLI threshold than previously suggested [27] for defining constrained genes, because it captured more known haploinsufficient genes important for heart and diaphragm development. Genes were also grouped by their expression levels in mouse developing diaphragm. Microarray expression profile of mouse pleuroperitoneal folds at E11.5 was taken from a previous study [34]. Normalized gene expression levels were converted to rank percentiles with smaller values corresponding to higher expression. Human orthologs of mouse genes were identified using annotations from MGI database (http://www.informatics.jax.org/). When a human gene mapped to multiple mouse genes, the highest expression level was assigned to the human gene.
Fibroblasts were obtained from diaphragm biopies at the time of diaphragm repair from 36 CDH neonatal cases most of whom carried damaging de novo variants, including three cases carrying MYRF variants (p.G81Wfs*45, V679A, and R695H). Cells were cultured in Dulbecco's Modified Eagle's Medium supplemented with 10% heat-inactivated fetal bovine serum and 1x Antibiotic/antimycotic (Gibco; Life Technologies), following standard conditions. Cells were cultured in parallel in successive passes until optimal confluence was reached, and were collected with 2.5% Trypsin (Gibco; Life Technologies) and harvested by centrifugation 5 minutes at 1200rpm. Total RNA was extracted from the cell pellet of each subject using RNeasy LipidTissue mini Kit (QIAGEN) according to manufacturer's protocol. The quality and quantity of RNA were assayed using a Qubit RNA Assay Kit in a Qubit 2.0 Fluorometer (Life Technologies) and RNA Nano 6000 Assays on a Bioanalyzer 2100 system (Agilent Technologies). cDNA libraries were prepared with the TruSeq Stranded Total RNA Sample Preparation kit (Illumina), following the manufacturer instructions. And the purified products were evaluated with an Agilent Bioanalyzer (Agilent Technologies). The library was sequenced on Illumina HiSeq 2000 platform in 100-bp paired-end reads.
RNA-seq reads were mapped to the human reference genome (GRCh37) using STAR (version 2.5.2b) [72]. Gene expression levels were quantified as TPM from the output of FeatureCounts (2015–05 version) [73]. Only protein coding genes were kept for analysis and genes with no mapped reads in at least half of the samples were filter out. All sequenced samples had >20 million mapped read pairs with >90% mapping rate. Principle component (PC) analysis of gene expression profile showed that five samples were separated from others on the first two PC axes (S5 Fig). The outlier samples were likely due to different number of passages in cell culture, and were removed from analysis.
Differential expressed genes (DEG) between cases with MYRF variants and others were identified using DESeq2 package [41]. DEG were selected using following criteria: adjusted p-value < 0.5 and adjusted fold change > 0.5 or < -0.5. We noted that all three MYRF de novo variant carriers were male. To avoid confounding effect of gender, DEG analysis was also performed by comparing male samples with or without MYRF variants. The full DEG list is given in S4 Data.
To evaluate the consequence of MYRF damaging variants on patients’ transcriptome, we tested if putative MYRF target genes were systematically down-regulated in the fibroblast cells with MYRF variants using gene set enrichment analysis (GSEA). The MYRF target genes as oligodendrocyte-specific genes that had at least one MYRF ChIP-seq binding peaks with 100kb of transcription start site [40]. We then identified corresponding human orthologs using biomaRt package [74]. A total of 74 human genes were defined as putative target genes for GSEA.
We selected six genes from differentially expressed genes between MYRF mutation carriers and other cases, including four down-regulated (GATA4, DBNDD2, MYO1D and NFASC) and two up-regulated (H3F3C and SEMA3A) in MYRF mutant cells. First-strand cDNA was synthesized from the total RNA (500ng~1 µg) using the RNA to cDNA EcoDry Premix (Random Hexamers) kit (TaKaRa) according to manufacturer's instructions. Primers for the selected genes (S6 Table) were synthesized by IdtDNA. All qPCR reactions were performed in a total of 10 µl volume, comprising 5 µl 2x SYBR Green I Master Mix (Promega), 1 µl 10nM of each primer and 2 µl of 1:20 diluted cDNA in 96-well plates using CFX Connect Real-Time PCR Detection System (Bio-Rad). All reactions were performed in triplicate and the conditions were 5 minutes at 95°C, then 40 cycles of 95°C at 15 seconds and 60°C at 30 seconds. The relative expression levels were calculated using the standard curve method relative to the β-actin housekeeping gene. Five-serial 4-fold dilutions of cDNA samples were used to construct the standard curves for each primer.
To assess the genetic overlap with other developmental disorders and especially CHD, we tested if the de novo damaging variants in CDH cases were enriched in known and putative CHD and DD genes. DD genes were extracted from DDG2P database [75] (accessed on Jan 11, 2018) and filtered to keep “allelic requirement” as monoallelic, X-linked dominant or hemizygous, and required “organ specificity list” to include brain, heart or not specific to any organ. A total 508 DD genes were identified, including 460 confirmed DD genes. CHD genes were collected based on a recent exome study of 2645 trios [29]. CHD genes included high heart expressed genes (HHE; ranked at top 25%) or known human CHD genes that were affected by more than one damaging de novo variants (LGD or D-mis defined by meta-SVM [76] as the original publication on CHD [29]) or constrained (pLI≥0.5) HHE genes affected by only one damaing variants from the same study. A total 200 CHD genes were identified, 57 of which overlapped with DD genes.
To assess if the exome-wide de novo damaging variants in CDH were enriched in CHD and DD genes, simulations were done to randomly place variants to the coding regions in a way that keeps the number of variants, tri-nucleotide context, functional effect, and deleteriouness prediction the same as that of the observed data [77]. Here the coding region was defined as coding sequences and canonical splice sites of all GENCODE v19 coding genes. For damaging mutations identified from WES data, the coding regions were restricted to the regions that have at >10X coverage in least 80% samples. Empirical p-value was calculated as the chance when there were more simulated damaging variants than observed in the given gene set. We ran 50,000 simulations to evaluate the significance. And the expected number of variants in a gene set was the average number of randomly generated variants in a gene set over all simulations.
To evaluate the functional convergence of genes affected by damaging variants, we extracted 89 genes that included 86 constrained genes (pLI≥0.5), two known candidates for CDH (GATA6, WT1), and a known haploinsufficient gene (KDM5B). Gene sets were derived from Gene Ontology Biological Process (GO-BP, accessed Feb 1st, 2018). The GO-BO categories that were statistically over-represented in the gene list (FDR<0.1) were identified using hyper-geometric test implemented by BINGO [78]. Terms annotating more than 750 or less than 25 genes were discarded, because large gene-sets usually represent broad categories without specific biological meaning. Small gene sets on the other hand are not likely to produce statistically significant results.
Enriched gene sets were graphically visualized as a network, in which each gene set is a node and edges represent overlap between sets. The Cytoscape software [79] and EnrichmentMap plugin [80] were used to construct the network. The color gradient of nodes reflects the enrichment p-values. Node size is proportional to the number of genes in the gene set. Edge thickness is proportional to the similarity score between gene sets which is defined by the average of Jaccard coefficient and overlap coefficient [80]. Enriched gene sets with highly overlapping genes (S6 Data) were grouped together and annotated manually.
|
10.1371/journal.pcbi.1000855 | Organization of Cellular Receptors into a Nanoscale Junction during HIV-1 Adhesion | The fusion of the human immunodeficiency virus type 1 (HIV-1) with its host cell is the target for new antiretroviral therapies. Viral particles interact with the flexible plasma membrane via viral surface protein gp120 which binds its primary cellular receptor CD4 and subsequently the coreceptor CCR5. However, whether and how these receptors become organized at the adhesive junction between cell and virion are unknown. Here, stochastic modeling predicts that, regarding binding to gp120, cellular receptors CD4 and CCR5 form an organized, ring-like, nanoscale structure beneath the virion, which locally deforms the plasma membrane. This organized adhesive junction between cell and virion, which we name the viral junction, is reminiscent of the well-characterized immunological synapse, albeit at much smaller length scales. The formation of an organized viral junction under multiple physiopathologically relevant conditions may represent a novel intermediate step in productive infection.
| The entry of human immunodeficiency virus (HIV) into cells is the target for new therapies preventing HIV infection. While intermediate steps of viral entry have been characterized, the progression between these steps and how they result in productive infection are not well understood. By using stochastic modeling, we examine the initial interaction of a single viral particle with a flexible plasma membrane populated with viral receptors. The model predicts the formation of an organized receptor ultrastructure beneath the viral particle, which we name viral junction and which may contribute to productive viral infection. The organization of the viral junction depends on receptor density, CD4 bond stability, membrane mechanical flexibility, as well as viral protein organization and density.
| Strategies for antiretroviral therapy have recently focused on inhibiting human immunodeficiency virus (HIV) adhesion, fusion, and entry. The biochemical properties of the dynamic binding interactions between host cell and viral receptors have been well characterized [1], [2]. However, how these interactions may work together for viral adhesion to progress toward an effective fusion event is not well understood. In particular, whether a single viral particle has the ability to spontaneously organize receptors at the cellular plasma membrane is unknown. Viral adhesion occurs on length and time scales that are difficult to monitor in real time because of the limited spatial and temporal resolution of current light and electron microscopes and the small size of virions (100nm in diameter, with entry into the cell taking place after only a few minutes [3], [4]). Here we use stochastic modeling to test the fundamental hypothesis that the cell-virus interfacial area forms an organized ultrastructure during viral adhesion, similar to the well-characterized immunological synapse [5] but at much smaller length scales (i.e. 0.1µm versus 10µm, respectively [3], [6]).
Virus-cell adhesion is primarily governed by bimolecular bonds formed between the viral surface protein gp120, and its cellular receptor CD4 [7]. gp120 molecules are arranged on the viral surface in trimers [8], [9]. For type-1 HIV (HIV-1), productive infection is also dependent on the subsequent binding of a cellular co-receptor, most commonly CCR5 or CXCR4 [10]. gp120-coreceptor binding induces a dynamic refolding of viral surface proteins which provides the driving force for fusion of the viral and cellular membranes [11], [12]. The formation of lipid microdomains on length scales similar to the viral diameter have been shown to result in protein colocalization [13]. Here we hypothesize that viral protein adhesion to cellular receptors, coupled with plasma membrane rigidity can produce highly organized protein structures at the cell surface on the same length scales as the virion and lipid rafts [13].
We used stochastic modeling based on recent single-molecule force spectroscopy measurements [14] to assess the spatial and temporal organization of cellular receptor CD4 and co-receptor CCR5 at the plasma membrane, as they dynamically interacted with gp120 on the viral surface (Fig. 1). We will discuss how the energy of bond formation between gp120 on the virion and receptors on the cell surface acts as an organizing force amidst disordering thermal energy. Specifically, thermal energy drives stochastic movement of the laterally diffusing receptors on the plasma membrane, the formation and destruction of bonds between viral proteins and cellular receptors, and the deformation of the plasma membrane. The conditions that we explore here are designed to determine how each type of stochastic movement contributes to the organization of cellular receptors at the plasma membrane. In addition, we study whether viral particles with gp120 trimers that are capable of diffusing on the viral surface result in distinct receptor organization between the virus and the cell.
The combined system of virion, plasma membrane, viral receptors and cellular receptors studied here, is modeled as a succession of discrete states. The transition to new states is assumed to be a Markov process [15] and computed using the local-steady state approximation to the Fokker-Planck equation [16]. The organizations of virus-cell bonds discussed here were produced using the 3-D location of receptors on the plasma membrane actively bound to gp120 molecules relative to the center of the virion itself.
We performed simulations considering the dynamic progress of the junction between the virion and the cell surface to be a stochastic, Markov process [15]. The system itself consists of a rigid sphere (the 100nm-diameter virion) interacting with a deformable surface (the plasma membrane, 200×200nm in dimension). The virion is populated with gp120 trimers, which can bind receptors (CD4 and CCR5) located on the plasma membrane (Fig. 1). All entities within the system including the plasma membrane, the configurations of viral proteins, the position of the virion itself, the dynamics of the bonds between the virus and the cell, as well as the positions of the cellular receptors (which are either bound or unbound to gp120), are specified by discrete states in Markov dynamics. The probabilities of generating a particular sequence of states, or a trajectory, are governed by the transition rates between these states. The transition rates between states that involve changes in the physical configuration of the system (i.e. the movement of the virion, the positions of the proteins and the plasma membrane from their current position to each possible new position) were calculated according to the local-steady state approximation of the Fokker-Planck equation [16],Here, is the forward association rate for the possible state adjacent to the current state (), is the diffusion coefficient of each physical entity of the system (membrane proteins, viral proteins, etc.) whose position has changed between the current and future state, is the change in position of that physical parameter, is the difference in system energy which results from the movement of that physical parameter between states and is where is Boltzmann constant and is the absolute temperature. This approximation calculates the rate of transition between the current state and subsequent adjacent state using the total change in energy that accompanies the progression from one discrete state to another. Specifically, a favorable change in energy (e.g. the relaxation of a bond between viral proteins and cellular receptors) results in an increased transition rate and an unfavorable change in energy (e.g. the deformation of the plasma membrane) results in a decreased transition rate from one state to the next.
Each possible state differs from the current state by the discrete position of any physically real object within the system (e.g. the x, y, z position of the virion, the position of a CD4 protein or a CCR5 protein, a discrete point along the plasma membrane, etc.) or by the creation or destruction of a bond between a receptor and gp120. For example, a single CD4 protein with no neighboring proteins on the plasma membrane would offer four possible new states based on its physical location because of the Cartesian coordinates used to define the flexible plasma membrane. In this example, is an experimentally determined diffusion coefficient for CD4 on a cellular membrane, is the three-dimensional distance between the current position of CD4 and that of each available discrete position within the plasma membrane. For an unbound CD4 protein, a new state dictating the movement of the protein would use . However, an unbound CD4 protein may also offer additional states which do not correspond to the physical movement of the CD4 if it is capable of binding an unbound gp120 protein. Those additional states will vary from the current state by the creation of a previously nonexistent bond; the forward rates toward these states are discussed below. However, should the CD4 protein in question already be participating in an existing bond, the forward rate corresponding to the movement of the bound CD4 protein would have a nonzero between states. This nonzero is the change in energy of the existing bond and determines the probability that this CD4 protein moves in an energetically favorable (relaxing the existing bond) or an energetically unfavorable (applying tension or compression to the bond) manner. For the case of receptors located on the plasma membrane, the forward rate constant is calculated using a determined by the distance between discrete points along the plasma membrane, which vary during the simulation according to the membrane deformation. For system parameters such as the virion position, is a fixed step size which dictates that the forward rates will vary only according to the change in energy of the existing bonds. Similarly, the transition rates involving a change in z-position of discrete plasma membrane points are calculated using a fixed . Here the z-position corresponds to the height of each plasma membrane point in the z-axis while the plasma membrane itself is oriented in the x-y plane. A fixed , dictates that the transition rates of plasma membrane points will vary only according to the local membrane free energy, as discussed below, and if it be the location of a bound cellular receptor, the change in energy of that particular bond.
For the systems examined here, the elapsed time between states and the distance over which physical objects move are of such a small order of magnitude that two assumptions can be made. First, the local energy landscape is approximated to be linear. Second, the probability density is assumed to be at a local steady state. Therefore, at small length scales, the system of the virion and plasma membrane is well described by the high-friction limit of the Fokker-Plank equation.
The on and off rates ( and ) for CD4 and CCR5 bond formation with gp120 were calculated using experimentally measured rates [14]. Initial values, , were computed using a model described by Hummer and Szabo [17],Here, is the effective diffusion coefficient, is the molecular spring constant, is the distance along the free energy well from the minimum to bond rupture, and is the minimum bond potential energy. This value of was then used to calculate for all bonds using the energy relationFor already existing bonds between cell and virion, values were calculated using the relationwhere the bond potential energy, , was calculated using a parabolic approximation of the Lennard-Jones potential,Here, is the distance along the energy potential calculated by subtracting the length of the proteins involved in the bond from the shortest distance between the location of the proteins on the plasma membrane and the virion.
The total probability of transitioning out of the current state was equal to the sum of all forward rates for each possible destination state:The specific destination state of the system was determined by a pseudo random number generator (PRNG). Briefly, once all possible states available to the current state are determined and their forward rates are calculated, the probability (or rate constant) describing the likelihood that the system transitions away from the current state is the total sum of all forward rates, . To determine which of these possible states is the next destination state, the PRNG yields a random number, , uniformly distributed between 0 and 1. is multiplied by resulting in a random position between 0 and , which corresponds to a particular adjacent state. The system is subsequently updated, newly available states are determined, and their new forward rate constants are calculated according to the imposed changes (e.g. if a gp120-CD4 bond breaks, new rates are calculated for the newly free gp120 and CD4 molecules). This process was repeated, updating each new state sequentially. Energy changes that governed the evolution of the system included those of individual gp120-CD4 bonds, gp120-CCR5 bonds, and the deformation of the plasma membrane with a specified elastic modulus and surface tension.
The fluctuations of the membrane, diffusion of the receptors and the diffusion of the virus are described by Fokker-Planck equations. The simulation methodology is described in Atilgan et al. [18]. The total free energy of the system, , is given bywhere is the free energy of the plasma membrane calculated using the Canham-Helfrich form [19], [20],Here, is the mean curvature of the membrane, is the local area, is the elastic modulus and is the surface tension of the membrane. It should be noted that electrostatic interactions between protein pairs not bound together and between the viral and cellular membranes were not included in the computation of the energy for a given state. In addition, we simplified our model by assuming that the concentration of the local actin filament network beneath the cellular membrane is sufficiently low so as to not dictate plasma membrane deformation [21].
The time elapsed as the system stepped from one state to another was also calculated and used to determine the total time elapsed during the simulation, starting at = 0s. The duration of each time step was calculated using the equationHere, is the time step from the to the +1 state and is the uniformly distributed random number between 0 and 1 provided by the PRNG.
The plasma membrane was initially defined as a completely flat surface. To simulate a more realistic interaction between cell and virion, the plasma membrane was allowed to evolve during the initialization of the system without the ability to form productive bonds with the virion above it. After this brief initialization (2×106 sequential iterations), the simulation of receptor-mediated viral adhesion to the cell surface was allowed to begin. Surface proteins on the plasma membrane were randomly distributed during the initialization according to the PRNG and concentrations of diffusing, unbound proteins were kept constant throughout the simulation. Proteins on the viral surface were either evenly spaced over the entire particle as previously described [22], or randomly distributed according to the PRNG using a random zenith, θ, between 0 and π according to the probability density distribution , and random azimuth, , uniformly distributed between 0 and 2π.
Throughout the simulation, proteins had a finite volume, so that other proteins were not allowed to diffuse through one another on either the viral surface or the plasma membrane. In addition, the plasma membrane and virion could not occupy the same space. If a physical obstacle was encountered, the forward rate for that adjacent state was set equal to 0. The actual lengths of gp120, CD4 and CCR5 molecules were also used when calculating bond interaction distances and free energies (Table 1). Lastly, gp120 trimers located on the viral surface were capable of binding up to three CD4 molecules and three CCR5 molecules at a time. As stated earlier, CCR5 adhesion to gp120 in contingent on a previously existing gp120-CD4 bond. In our system, gp120 trimers were not allowed to bind CCR5 unless that trimer was already involved in a gp120-CD4 bond. However, the number of CCR5 bonds was not allowed to exceed the number of CD4 bonds formed with a single gp120 spike, i.e. no synergistic effect was imposed throughout a gp120 trimer that would allow a single CD4 adhesion to promote multiple CCR5 bonds. For a more detailed explanation of the modeling algorithm see Text S1.
First, we studied the development of viral-cell adhesion with a viral gp120 organization in which protein trimers were evenly distributed over the viral surface and were not allowed to move. We observed that as the system progressed towards steady state, the gp120-CD4 bond probability distribution displayed three distinct phases of organization (Fig. 2A).
The system quickly transitioned from single gp120 trimers bound to the plasma membrane, which we call phase I, to a second state through rotation and translocation, which allowed multiple gp120 trimers to bind to cellular receptors and produced a single broad node of bond probability formation, which we call phase II (Fig. 2B). Accordingly, upon initial physical contact between the cell and virion, the gp120-CD4 bond probability distribution displayed a single maximum near the viral center (∼0nm). This is a result of initial gp120-CD4 bonds occurring most preferably at the closest point on the virion to the plasma membrane (Fig. 2B). The shift of the bond probability maximum from the viral center is the result of averaging the initial bond locations of multiple simulations ( = 8) where the viral particle is binding to a non-uniform plasma membrane surface. At the initial time of contact the closest point on the cellular membrane is not always presented to the virion directly at = 0nm. Phase II resulted in a broadening of the CD4 bond probability distribution, during which its maximum shifted to a distance = 37nm from the center (radius of the virion, 50 nm), i.e. CD4 receptors participating in the viral junction organized into a ring-like structure or corona (Fig. 2B). CCR5 bonds also organized into a corona, with a spatial distribution similar to that of CD4 but containing much fewer bonds (Fig. 2B).
Through continued rotation and translocation of the virion, the interfacial region between cell and virion further evolved to develop a central, “anchor” gp120 spike surrounded by trimers bound to adjacent cellular receptors. This configuration produced a bimodal bond probability distribution of bound cellular receptors, which we call phase III (Fig. 2, A–D). While the organization of the adhesive junction between the cell and virion developed, the flexible plasma membrane spontaneously deformed and engulfed the virion. The deformation of the plasma membrane increased the cell surface area which was close enough to the virion to allow further receptor binding to gp120 trimers, thereby increasing the total bond number (Fig. 2E). The time of formation of an organized viral junction (Fig. 2, C and D) remained fairly constant over multiple simulations with different (random) initial states indicating that the final organization and number of bonds in the viral junction were relatively independent of the initial positions of the cellular membrane, virion, and receptors.
The virion formed a stable adhesion interface and progressively increased the number of bonds. Simultaneously, the position of the virion above the membrane decreased as the membrane deformed to engulf the virion with time. Here, the change in the vertical height of the virion from its position at initial cell contact (i.e. the depth of virion engulfment) is referred to by , and the evolution of as adhesion progresses from Phase I to Phase III is illustrated in Fig 2F. Interestingly, after CD4 and CCR5 bond probabilities had reached steady states, the depth of engulfment of the virion continued to increase until the plasma membrane reached an equilibrium deformation (Fig. 2, G and H). Again, there are two counteracting “forces” that dictate the direction in which the organization of the virion-cell interface progresses. The first force is the energetically favorable formation of bimolecular bonds between proteins on the viral and cell surfaces; the second force is the energetically unfavorable deformation of the plasma membrane. The plasma membrane may be maintained in an unfavorable, deformed position if it is permissive of an increase in bond number (Fig. 2H). Fixed gp120 units on the viral surface forced the virion to maximize bond formation by rotating and laterally moving the virion so as to minimally deform the plasma membrane, while recruiting new receptors to bind. Eventually, the depth of engulfment of the virion and the radial profile of the plasma membrane stabilized at heights that no longer exposed new gp120 units to cellular receptors on the plasma membrane. The organization of spatially fixed gp120 molecules on the viral surface regulates the organization of bonds between cell and virion (Fig. 3).
Together these results suggest two important findings: (i) CD4-gp120 bimolecular bonds can be highly organized in the interfacial region between the cell and virus; (ii) the receptor organization is dynamic: initially a peak forms at the center, followed by a corona or bull's eye pattern of virus-cell bonds develops, and finally a peak and a corona co-exist while the plasma membrane deforms and engulfs the virion (Fig. 2, C and D).
In the following, we conduct simulations over a wide range of parameters to investigate the effects of viral protein organization, receptor concentration, plasma membrane rigidity, and overall bond stability between the cell and virion, on the organization of cellular receptors.
Recent studies suggest that the increased density of gp120 on the surface of viral particles could increase infection [3], [23], [24]. Therefore, we studied the effect of gp120 density (at fixed positions on the viral surface) on the organization of the viral junction over the physiological range of 7–20 gp120 trimers per virion [3]. We found that, when viral particles contained few gp120 (7–9 trimers), organized viral junctions did not form and little plasma membrane deformation occurred (Fig. 4). For an increased number of gp120 trimers per virion (14–20 trimers), the distance between adjacent gp120 trimers was sufficiently reduced that spontaneous deformations in the plasma membrane could result in an increase in the number of bonds (Fig. 4). The number of bound CD4 and CCR5 bonds correlated directly with the number of gp120 trimers on the viral surface (data not shown), indicating that reduced gp120 density result in fewer virus-cell bonds, a less dynamic protein organization within the viral junction, and greatly reduced plasma membrane deformation.
Recent work suggests that gp120 units may diffuse on the viral surface [25]. To determine whether diffusing gp120 trimers on the viral surface could also produce an organized viral junction, we conducted simulations allowing gp120 trimers to diffuse freely on the viral membrane (Fig. 5). The presence of freely diffusing gp120 trimers allowed for the formation of three times as many CD4 bonds than in the fixed gp120 case (Figs. 2E and 5E). These bonds also formed 10 times faster than for virions with fixed gp120 positions. While the number of CD4 bonds increased beyond the steady state value of the fixed case, the number of CCR5 bonds did not change significantly from the fixed case (Figs. 2E and 5F).
Initially the CD4 bond probability profile in the organized viral junction was similar to that of the case of spatially fixed gp120 units. Then, the maximum of the bond probability grew outward resulting in a local maximum at a stable distance ∼27nm from the center (Fig. 5A). After a slight delay, the CCR5 bond probability grew in a similar manner, but was dwarfed by the probability of forming CD4 bonds (Fig. 5A). The probability distribution of bound cellular receptors did not feature a single peak directly beneath the center of the virion because of the deformation of the plasma membrane. The resulting curvature of the plasma membrane while induced by the virion, did not exactly match the virion curvature. These mismatched curvatures resulted in only a fraction of the area directly under the virion to be close enough to mediate the formation of bonds between viral and cellular receptors (Fig. 5, G and H, Stable CD4 and Rigid PM). These mismatched curvatures did not result from specific random starting conditions, distributions shown here result from eight independent simulations. As in the fixed gp120 case, virions with freely diffusing gp120 induced membrane deformation and resulted in viron engulfment (Fig. 5, G and H). Ultimately, membrane deformation resulted in a depth of virion engulfment of = 22nm. However, while the number of bonds increased much faster than in the fixed gp120 case, plasma membrane deformation and virion engulfment occurred at rates similar to those in the fixed gp120 case (Figs. 2F and 5G).
The location of bound cellular receptors corresponds to the shortest distance between the virion and the plasma membrane. When the plasma membrane is completely flat (κ = ∞), the virion is brought in close proximity to the cell to maximize the adhesion competent viral surface area (Fig. 6A), cellular receptors at = 0 nm cannot bind the virion due to lack of space, resulting in a corona (Fig. 5, B, G, and H, hyper rigid PM). When the rigidity of the plasma membrane is relatively low (κ = 20 kbT/nm), the maximum in the probability of bound receptors away from = 0nm, i.e. a corona of bound receptors is formed. When the rigidity of the plasma membrane is increased (κ = 100 kbT/nm; Fig. 5, D and H, Rigid PM), the steady state deformation of the plasma membrane is reduced and the maximum probability of bound receptors is shifted toward the radial center. Simple geometric analogies for the soft and rigid plasma membrane case are a sphere sitting in the bottom of a cone compared to a sphere sitting at the bottom of a larger sphere, respectively (Fig. 6, B and C).
The slight decrease in bond number over time for the case of diffusive gp120 (Fig. 5E, Stable CD4 and Rigid PM) also results from the membrane deformation. Initially, bond number increases rapidly because diffusive gp120 molecules can concentrate between the virion and the plasma membrane. Bond number decreases slowly when the plasma membrane deforms, limiting the adhesion-competent area for cellular receptors to occupy (Fig. 6). Taken together these results suggest that the potential ability of gp120 molecules to diffuse in the viral membrane has a qualitative effect on the type of organization of the virus-cell interface and a quantitative effect on the number of bound receptors in that interface.
We note that a virion with diffusing gp120 trimers produces patterns of bound receptors similar to those recently reported using electron tomography [25], especially those produced by CD4 receptors with biphasic stability (Fig. 5C, Unstable CD4; see more details below and in the Discussion section).
Previous work suggests that gp120 trimers are unevenly distributed on the viral surface [3], [25] and that these trimers may form fixed clusters [3]. We examined whether the formation of an organized viral junction underneath the virion containing a corona of bonds at an intermediate radial distance required uniformly distributed gp120 trimers by examining viral adhesion governed by randomly distributed gp120 trimers on the viral surface. Frequently, interfaces developed by these virions contained randomly clustered gp120 units which resulted in complex bond distributions similar to that produced with evenly placed gp120 (Fig. S1). However, the distance between the concentric rings of maximum bond probability depended on the spacing of gp120 on the viral surface and varied from one simulation to the next.
This double-corona viral junction was a direct result of the non-uniform gp120 spike distribution on the viral surface. The interaction between local viral gp120 organization and receptors on the plasma membrane resulted in the progression of bond organization through distinct phases similar to the case of evenly distributed gp120 (Fig. 2B and Fig. S1). The virion first made contact with the plasma membrane, rotated, and as the deformation of the plasma membrane continued, a bimodal bond probability was established as virion rotation presented gp120 trimers previously unattainable to the receptors. The distance between gp120 trimers ultimately determined the distance between the nodes of the bond probability distribution (Fig. S1). Indeed, other simulations resulted in distributions that resembled phase II organization, i.e. a single-maximum bond probabilities single (Fig. 2A, Fig. S1). Unbound gp120 trimers that were spaced too far away on the viral surface for the virion to successfully rotate and expose to the cellular receptors resulted in halting the progression of bond organization at a single, off center probability maxima (Fig. S1). Together these results suggest that the final organization of CD4 receptors bound to fixed gp120 at the virus-cell interface before viral entry depends on the spatial organization of gp120 molecules on the surface of virions.
The depletion of cholesterol from cellular membranes may significantly inhibit HIV-1 infection [26] and patients treated with cholesterol lowering statins seem to present decreased viral loads [27]. Aside from affecting subcellular pathways activated by statins, the depletion of cholesterol can dramatically affect the mechanical stiffness of the plasma membrane [28]. Previous work has demonstrated that varying cholesterol levels have a direct effect on the deformability of lipid vesicles [28]. We found that the plasma membrane rigidity critically influenced the spatial organization within the viral junction (Figs. 5D and 7, A and B).
Cells with a completely rigid plasma membrane (κ = ∞) coupled with evenly distributed gp120 units never progressed beyond the second phase of adhesion/bond formation described above (Fig. 7B). This completely rigid plasma membrane resulted in slightly fewer CD4 bonds than in viral junctions involving cells with a flexible membrane and virions with fixed gp120 trimers (Fig. 7E). A fivefold increase in membrane rigidity (κ = 100 kbT/nm vs. 20 kbT/nm) did not significantly change the steady state number of bonds in the viral junction (Fig. 7E, Rigid PM) and viral junctions typically took approximately twice as long to progress beyond the second phase of adhesion. Moreover, a fivefold increase in membrane rigidity resulted in a little more than double the depth of virion engulfment (Fig. 7G, Rigid PM).
For systems containing diffusing gp120 trimers adhering to receptors on a more rigid membrane, the virion depth of engulfment quickly stabilized at almost half the depth observed with a more flexible plasma membrane (Fig. 5G). This was accomplished without a significant change in the numbers CCR5 bonds (Fig. 5H). However, a more rigid membrane in conjunction with diffusive gp120 resulted in increased CD4 bond numbers and a higher probability of bond towards the center, = 0 nm (Fig. 5, D and E, Rigid PM). Together these results suggest that the mechanical properties of the plasma membrane can affect the organization and distribution of bonds within the viral junction. If organized virion-cell bonds are indeed important for successful infection, then the effect that cholesterol concentration has on the mechanical properties of the plasma membrane may contribute to its effect on HIV-1 infection.
The concentration of CD4+ T cells harvested from the blood and lymph node tissue of infected patients correlates with infection and the ratio of [CCR5]∶[CD4] [29]; infection will correlate more to CD4 expression when CCR5 is expressed in limiting amounts and vice versa [30]. Our results indicate that the steady state number of CCR5 bonds in the viral junction depended most critically on the concentration of CCR5, and did not strongly depend on the mechanical stiffness of the plasma membrane or the number, mobility, and the organization of gp120 on the virions (Fig. 5F and 7F).
For fixed gp120 trimers on the virion, a tenfold increase in CCR5 concentration produced a twofold increase in steady state number of CCR5 bonds in the viral junction (Figs. 2E and 7F, for 10∶1 and 1∶1 [CD4]∶[CCR5] molar ratios, respectively). When CCR5 concentration was increased to the same level as CD4, the physical hindrance of additional bound proteins located in an area of comparable size became more of a determining factor for protein organization than when CCR5 was present at a lower concentration. For example, increasing the CCR5 concentration resulted in a CD4 bond probability distribution containing three local maxima (Fig. 7C). The increased number of CCR5 molecules, coupled with the increased stiffness and decreased length of the CCR5 bond quickly resulted in a virion with comparably little rotational freedom. The virion was bound by so many more CCR5 proteins than in any other simulation described here, that it had a much greater resistance to rotation and translocation so to minimize bond free energy (because of the higher stiffness of CCR5 bonds).
Figures 7, G and H, compare the depths of virion engulfment and the plasma membrane profiles for increased CCR5 concentration, [CD4]∶[CCR5] = 1∶1. While the plasma membrane has a similar z-position for a more rigid membrane (∼10nm away from viral center), the virion itself is sitting almost 10 nm higher than with a rigid membrane. This increased resistance ultimately results in phase III organization of bonds, however the increased CCR5 bonds at larger radii produced a CD4 bond probability distribution with three local maxima (Fig. 7C). The two outward CD4 bond probability maxima (∼45nm and 65nm) could have actually formed a single maximum, were it not for the steric hindrance of the shorter CCR5 bond coupled with the local plasma membrane deformation. The shorter CCR5 bonds concentrate in a small area of the plasma membrane and surrounded by the longer CD4 bonds which are able to diffuse over a larger area. Unable to occupy the same space, two local CD4 probability maxima flanking the outward CCR5 bond probability maxima were formed.
We recently demonstrated that the presence of CCR5 could result in a decrease of CD4-gp120 bond stability [14]. Therefore, we examined the effect of CD4-gp120 bond instability on viral junction formation and organization. For fixed gp120, CD4-gp120 bond instability resulted in fewer CD4 bonds between the cell and virus, averaging 8 at long time scales, as well as a decrease in average CCR5 bond number to practically zero (Fig. 7, E and F). The CD4 bond probability distribution was noticeably bare at the center of the virion-cell interface region compared to the stable CD4 bond case, indicating a slight difference in the third phase of organization (Fig. 7D).
Initial CD4 bonds were formed and increased in number similarly to the previous cases. However, when CCR5 bonds began to form and the CD4 bonds became unstable. CD4-gp120 bonds broke and the CD4 molecules diffused away, leaving only the CCR5 bond between the cell and virus. Ultimately, the last CCR5 bond broke and could not reform as no CD4 bonds were sufficiently close to initiate the gp120 conformation change. This production of CCR5 bonds and destruction of local CD4 bonds continued while the viral-cell interface as a whole maintained a constant number of bonds (Fig. 7, E and G), forming a globally stable adhesion interface (Fig. 7D). Interestingly, imposing a biphasic gp120-CD4 instability while also allowing gp120 trimers to diffuse on the viral surface most accurately recreated the bond organization previously observed using electron tomography (Fig. 5C).
Our computational results suggest that a viral particle can induce the formation of a highly organized ring-like ultrastructure of cell receptors bound to viral proteins, which we termed the viral junction. Our model involves biochemical (e.g. binding constants) and biophysical parameters (e.g. membrane stiffness) that have previously been measured. The diffusion rate constants of the plasma membrane and gp120 on the viral surface were assumed. However, these two unknown constants only set the rate of formation of a viral junction, not its steady state organization. Results from the model suggest that the formation of an organized viral junction is robust against relatively large variations in receptor concentrations, virion properties, physical properties of the plasma membrane, and dynamic properties of virus-cell bimolecular bonds.
The simulations revealed that several factors contribute to the organization of bonds on the flexible membrane. The ability of CD4 and CCR5 molecules to diffuse while bound to gp120 contributes to the organization of the viral junction by increasing bond formation while decreasing plasma membrane deformation. For fixed gp120 (i.e. unable to diffuse on the viral surface), the organization of the viral junction primarily depends on the gp120 distribution on the viral surface. While for the diffusive gp120 case, the organization of the viral junction primarily depends on the deformability of the plasma membrane. The longer and more flexible CD4 bond compared to the CCR5 bond, coupled with the local deformation of the plasma membrane, often result in different bond organization for these two receptors.
Our computational model suggests that the mechanical properties of the plasma membrane work in concert with viral gp120 organization to organize cellular receptors at the virion-cell interface. Changes in the stiffness of the plasma membrane, which could be mediated by changes in cholesterol content [26], affect the properties of the viral junction, including the total number of bonds between cell and virion. The virion-cell bonds work with the plasma membrane to reduce the overall potential energy of the system by two mechanisms. First, an energetically unstable bond forms where the plasma membrane is not locally deformed and the membrane is subsequently deformed to stabilize the bond. Second, the membrane spontaneously deforms to an unfavorable configuration and before it can relax a receptor forms a bond that maintains the deformation of the plasma membrane. The finite deformability of the plasma membrane limits how much the membrane can spontaneously deform without forming new bonds. By spacing gp120 trimers too far away on the viral surface for the plasma membrane receptors to spontaneously encounter them, the distribution of gp120 dictates the extent of plasma membrane deformation.
Recently it has been suggested that endocytosis plays an important role in HIV-1 infection [4]. The formation of an organized viral junction, which is computationally described here within a small 200×200nm area, could occur anywhere from the plasma membrane to within an endocytic vesicle.
Previous theoretical work using thermodynamic steady states predicted that an increase in membrane rigidity would decrease the number of bonds between cell and virion [21]. However, this conclusion was reached assuming a uniformly binding viral surface. Here, discrete locations of adherent gp120 trimers reveal that the point at which the plasma membrane is unable to bind and continue deformation is dependent on the distance at which unbound gp120 units are spaced.
A computational model of how CD4 receptor organization on a planer surface responds to binding gp120 trimers has previously been introduced [31]. This earlier work focused on the rate at which the gp120 molecules of a trimeric spike become bound to CD4 as a function of gp120 density and CD4 diffusion coefficient. While this work predicts an increase in local CD4 concentration under the virion due to bond formation, it does not appear to display receptor organizations similar to those predicted by our work. This difference may stem from this earlier works use of a rigid plane to simulate the cellular membrane and the absence of co-receptor adhesion. Here, we employ a flexible membrane, which we demonstrate plays a critical role in the organization of the viral junction. The addition of coreceptor adhesion also offered insight into the roles that CD4 and CCR5 bond micromechanics play in the formation of the viral junction. Lastly, we had the advantage of using experimentally measured kinetic and micromechanical values for gp120-CD4 bonds.
Recent reports suggest that gp120 could be partially disorganized or diffuse on the viral surface [25] and that viral infection correlates with gp120 concentration [3], [23]. Therefore, we studied the effects of diffusing vs. non-diffusing gp120, as well as gp120 density on the formation, organization, and dynamics of the viral junction. Viral particles with diffusive gp120 organized cellular receptors into a distinct corona to maximize the number of bonds between the cell and virus (Fig. 5, A–D). Sougrat et al. suggest that a virion adhering to a cell induces a similar formation of proteins between the cell and virus, called the ‘entry claw’, with comparatively little gp120 elsewhere on the virion. It was suggested that this concentration gradient in gp120 along the viral surface is best explained by the ability for gp120 to diffuse on the surface. The conditions that best reproduced the bond organization observed using electron tomography is a combination of diffusive gp120 trimers with an imposed biphasic stability on the gp120-CD4 bond, suggesting that gp120 are not permanently organized on the viral surface. However, the gp120 gradient observed by Sougrat et al. can also be explained by the possibility that virions with clustered gp120 trimers have preferential binding as well as the possibility that gp120 trimers could be cleaved from the virion during sample preparation for electron tomography. Therefore, in addition to the diffuse case, we also considered the effect of gp120 configurations on viral junction organization.
An experimental test of our computational predictions is challenging given the small sizes of the virion and associated viral junction. However, the advent of super resolution microscopy approaches, such as photoactivated localization microscopy (PALM) [32], could help determine the organization of the viral junction through co-labeling of gp120 and receptors CD4 and CCR5. Combined immunolabeling and cryo-electron microscopy (EM) or tomography could also help assess the organization of the viral junction; however EM often creates artifacts especially when visualizing the plasma membrane.
If we assume that certain bond configurations result in enhanced infection, then the subpopulation of infectious viral particles not only depends on gp120 density, but also gp120 organization on the viral surface. A viral population containing particles equipped with a functional gp120 organization (i.e. with relatively clustered gp120 on the viral surface) could lead to the formation of an organized viral junction, while those with dysfunctional organizations (with relatively distant gp120), may be unable to infect cells. Since the formation of a viral junction depends on receptor density and the mechanical properties of plasma membrane, viral junction formation could be cell type-specific. One might speculate from the spectrum of mechanical properties exhibited by different cell types as well as the distribution of viral particle size and gp120 concentration that a subset of viral particles might preferentially infect one cell type while another subset of particles could preferentially infect another cell type.
Although the formation of the viral junction would occur at the much smaller length scales than that of a whole cell, the organization of the viral junction is reminiscent of the immunological synapse. It is therefore tempting to speculate on the possible signaling function of the viral junction. For instance, gp120 binding induces assembly of a local actin network and coreceptor binding induces disassembly of actin filaments [33], [34]. By analogy to the better-characterized immunological synapse, we speculate that part of this signaling function could be related to the complex evolution of an organized viral junction by relying on one organizational phase between receptor and coreceptor for an initial signaling event and a subsequent organizational phase for a secondary signaling event.
|
10.1371/journal.pntd.0001012 | Collagenolytic Activities of the Major Secreted Cathepsin L Peptidases Involved in the Virulence of the Helminth Pathogen, Fasciola hepatica | The temporal expression and secretion of distinct members of a family of virulence-associated cathepsin L cysteine peptidases (FhCL) correlates with the entry and migration of the helminth pathogen Fasciola hepatica in the host. Thus, infective larvae traversing the gut wall secrete cathepsin L3 (FhCL3), liver migrating juvenile parasites secrete both FhCL1 and FhCL2 while the mature bile duct parasites, which are obligate blood feeders, secrete predominantly FhCL1 but also FhCL2.
Here we show that FhCL1, FhCL2 and FhCL3 exhibit differences in their kinetic parameters towards a range of peptide substrates. Uniquely, FhCL2 and FhCL3 readily cleave substrates with Pro in the P2 position and peptide substrates mimicking the repeating Gly-Pro-Xaa motifs that occur within the primary sequence of collagen. FhCL1, FhCL2 and FhCL3 hydrolysed native type I and II collagen at neutral pH but while FhCL1 cleaved only non-collagenous (NC, non-Gly-X-Y) domains FhCL2 and FhCL3 exhibited collagenase activity by cleaving at multiple sites within the α1 and α2 triple helix regions (Col domains). Molecular simulations created for FhCL1, FhCL2 and FhCL3 complexed to various seven-residue peptides supports the idea that Trp67 and Tyr67 in the S2 subsite of the active sites of FhCL3 and FhCL2, respectively, are critical to conferring the unique collagenase-like activity to these enzymes by accommodating either Gly or Pro residues at P2 in the substrate. The data also suggests that FhCL3 accommodates hydroxyproline (Hyp)-Gly at P3-P2 better than FhCL2 explaining the observed greater ability of FhCL3 to digest type I and II collagens compared to FhCL2 and why these enzymes cleave at different positions within the Col domains.
These studies further our understanding of how this helminth parasite regulates peptidase expression to ensure infection, migration and establishment in host tissues.
| Fasciola hepatica is a helminth parasite that causes liver fluke disease (fasciolosis) in domestic animals (sheep and cattle) and humans worldwide. In order to infect their mammalian hosts, F. hepatica larvae must penetrate and traverse the intestinal wall of the duodenum, move through the peritoneum and penetrate the liver. After migrating through the liver, causing extensive tissue damage, the parasites move to their final niche in the bile ducts where they mature and feed on host haemoglobin to support the production of eggs. To achieve these tasks, F. hepatica secretes a number of distinct cathepsin L cysteine peptidases (FhCL). Thus, the infective larvae that penetrate the host gut secrete cathepsin L3 (FhCL3), the migrating liver-stage juvenile parasites secrete both FhCL1 and FhCL2 while mature bile duct parasites that feed on host blood secrete predominantly FhCL1 but also FhCL2. Here we show that the major cathepsin L peptidases secreted by F. hepatica (FhCL1, FhCL2 and FhCL3) display differential ability to degrade host collagen (an important component of host tissues) and investigate this phenomenon at the molecular level.
| Papain-like cysteine peptidases, including cathepsins B and L, are ubiquitously secreted extracorporeally by helminth parasites of human and veterinary importance where they perform many important roles that are critical to the development and survival of the parasite within the mammalian host [1]. These roles include penetration and migration through host tissues [2], catabolism of host proteins to peptides and amino acids [3], [4], and modulation of the host immune response by cleaving immunoglobulin [5], [6] or by altering the activity of immune effector cells [7]. Accordingly, cathepsin peptidases are leading targets for novel anti-parasitic drugs and vaccines that block their function [8], [9].
Fasciola hepatica is the causative agent of liver fluke disease (fasciolosis) of domestic animals in regions with temperate climates. Although traditionally regarded as a disease of livestock, fasciolosis is now recognised as an important emerging foodborne zoonotic disease in rural areas of South America (particularly Bolivia, Peru and Equador), Egypt and Iran [10]. It is estimated that over 2.4 million people are infected with F. hepatica worldwide and around 91 million are at risk of infection [11]. To infect their mammalian hosts, F. hepatica larvae, which are ingested with vegetation contaminated with dormant cysts (metacercariae), penetrate the intestinal wall, enter the liver capsule and migrate through the parenchyma before invading into the bile ducts [12]. To facilitate this tissue migration, Fasciola secrete various members of a multigenic family of cathepsin L peptidases that exhibit overlapping but complementary substrate specificities and together cleave host macromolecules very efficiently [13], [14]. In fact, the ability of Fasciola to infect and adapt to a wide range of host species has been attributed to the effectiveness of this proteolytic machinery [14], [15].
Phylogenetic analyses have shown that the Fasciola cathepsin L gene family expanded by a series of gene duplications followed by divergence which gave rise to three clades expressed by tissue-migrating and adult worms (Clades 1, 2, and 5) and two clades specific to the early infective juvenile stage (Clades 3 and 4) [13], [14]. Consistent with these observations, our proteomics analysis identified representative enzymes from Clades 1, 2 and 5, but not from Clades 3 and 4, in the secretory products of adult F. hepatica [14]. More recently, we showed that the temporal expression and secretion of the specific cathepsin L clades correlated with the migration of the parasite through host issues; members of cathepsin L clade 3 (FhCL3) are secreted by Fasciola infective larvae and effected penetration of the host intestinal wall while clades 1, 2 and 5 (FhCL1, FhCL2 and FhCL5) peptidases are secreted by the immature liver-stage flukes and adult worms and function in preparing a migratory path through the liver and in the acquisition of nutrient by degrading host blood and tissue components. While clade 4 (FhCL4) peptidases are expressed by infective larvae they do not seem to be secreted and, therefore, may play an intracellular house-keeping function [13], [16]. Recent transcriptomic analyses of juvenile and adult stages have confirmed these observations [17], [18].
The secreted Fasciola cathepsins are produced in specialised gastrodermal cells which line the parasites's gut and are packaged in secretory vesicles before being extruded into the gut lumen from where they are released into host tissues [19]. The peptidases can efficiently degrade a range of host macromolecules including haemoglobin, immunoglobulin and interstitial matrix proteins such as fibronectin and laminin [3], [4], [20]–[22]. Notably, however, studies in our laboratory using functionally-active recombinant enzymes have shown that FhCL2 and FhCL3 exhibit an unusual ability to cleave native collagen [22], [23]. This is of relevance because collagenase-like activity is restricted to very few enzymes (e.g. bacterial collagenases, matrix metalloproteinases and human cathepsin K) and, therefore, the evolution and maintenance of such an activity in Fasciola suggests that it is essential to the parasite's ability to degrade the connective tissue matrix of the organs through which it migrates.
The active site of papain-like cysteine peptidases is relatively short, and while consisting of four subsites (S2-S1-S1′-S2′) with additional binding areas (S4-S3 and S3′) the specificity of substrate binding is principally governed by the residues that make up the S2 subsite [24], [25]. This S2 site forms a deep pocket capable of holding the P2 amino acid of the substrate and positioning the scissile bond into the S1 subsite for cleavage. In Carica papaya papain (PDB ID: 9PAP), the S2 subsite is composed of residues occupying positions 67, 68, 133, 157, 160 and 205. An analysis of these residues in the various cathepsin L clades clearly demonstrates divergence within the S2 subsite, in particular at the three positions that have the greatest influence on P2 binding i.e. at residues 67, 157 and 205 [14], [15], [23]. For FhCL2, the collagenolytic activity has been attributed to the presence of a particular residue, Tyr69, within the enzyme's S2 substrate binding site which is also found in human cathepsin K, the only mammalian cathepsin with the ability to cleave within the covalently-linked triple helices, of Col domains of native collagen [26], [27]. The S2 Tyr69 is also suggested to allow both enzymes to cleave macromolecular and dipeptide substrates with a Pro residue in the P2 position [22], [28]. More recently we showed that in FhCL3, this position is occupied by a larger Trp residue, a feature shared only with a ginger rhizome peptidase [29] which is also capable of cleaving collagen. Notably the S3 subsite of both FhCL3 and the plant enzyme are quite shallow, an observation that led us to advance the idea that the specificity of these enzymes might be restricted [23].
Our laboratory recently determined the three-dimensional structure of one of the major cathepsin L peptidases of adult F. hepatica, FhCL1 [22]. Similar to other cathepsins, the enzyme is composed of two domains (R and L) at the juncture of which is a cleft that forms the substrate-binding site and contains the enzyme catalytic machinery. Super-imposition of the alpha carbons of FhCL1 with cysteine peptidases from plants (e.g. papain, PDB ID 9PAP) and mammals (e.g. human cathepsin L PDB ID 1CJL) yields an r.m.s. deviation in the range 0.78 Å to 1.085 Å which is indicative of the very high conservation of the overall fold and shape that exists amongst all members of the papain family of cysteine peptidases [30]. Since the FhCL1 structure and fold can be described as practically identical to all other cathepsin L-like peptidases its scaffold can be exploited as a ‘prototype’ to investigate the role of critical amino acids within the S2 subsite in substrate binding (Table 1), particularly those of FhCL2 and FhCL3, whose primary structures are 78% and 70% identical to FhCL1, respectively. In the present study, we investigated and compared the substrate specificity of active recombinant forms of FhCL1, FhCL2 and FhCL3 with specific emphasis on their ability to degrade native collagen. This activity was interpreted by obtaining the enzymatic kinetic parameters (Km, kcat, and kcat/Km) of these enzymes on a range of peptide substrates and binding kinetics for specific inhibitory compounds. Furthermore, using mass spectrometry we mapped the cleavage sites of native collagen I and derived peptides, adding evidence for differential activities between FhCL2 and FhCL3. Finally, using our FhCL1 crystal structure as a template we created molecular dynamics simulations to explain how the active sites of these enzymes accommodate collagen-like substrates and endow them with this unusual collagenolytic activity. Our study provides biochemical and structural insights into the molecular mechanism of tissue invasion by these important parasitic helminths.
Z-Phe-Arg-NHMec, Z-Leu-Arg-NHMec, Z-Val-Val-Arg-NHMec, Tos-Gly-Pro-Arg-NMec, Tos-Gly-Pro-Lys-NMec, Boc-Ala-Gly-Pro-Arg-NMec, Boc-Val-Leu-Lys-NMec, Boc-Val-Pro-Arg-NMec, Z-Phe-Ala-CHN2, Z-Gly-Pro-Gly-Gly-Pro-Ala and Z-Gly-Pro-Leu-Gly-Pro were obtained from Bachem (St. Helens, UK). Cathepsin K inhibitor II was purchased from BD Biosciences (Sydney, Australia). E-64, DTT, EDTA and bovine nasal septum collagen type II were obtained from Sigma-Aldrich (Sydney, Australia). Calf skin collagen type I was purchased from Calbiochem. Pichia pastoris strain X33 was obtained from Invitrogen (San Diego, CA, USA). Ni-NTA agarose and columns were obtained from Qiagen (Australia). Pre-cast NuPage 4–12% Bis-Tris gels and pre-stained molecular weight markers were purchased from Invitrogen (Australia).
Recombinant F. hepatica procathepsin L1, L2 and L3 (FhCL1, FhCL2 and FhCL3) were produced in yeast as previously described [22], [23]. Briefly, P. pastoris (for FhCL1 and FhCL2 expression) and P. angusta (for FhCL3 expression) yeast transformants were cultured in 500 ml BMGY broth, buffered to pH 8.0, in 5 L baffled flasks at 30°C until an OD600 of 2–6 was reached. Cells were harvested by centrifugation at 2000× g for 5 min and protein expression induced by resuspending in 100 ml BMMY broth, buffered at pH 6.0 containing 1% methanol. Recombinant proteins were affinity purified from yeast using Ni-NTA-agarose. Recombinant propeptidases were dialysed against phosphate buffered saline (PBS) and stored at −20°C. The 37 kDa cathepsin L zymogens were autocatalytically activated and processed to 24.5 kDa mature enzymes by incubation for 2 h at 37°C in 0.1 M sodium citrate buffer (pH 5.0) containing 2 mM DTT and 2.5 mM EDTA. The mixture was then dialysed against PBS, pH 7.3. The proportion of functionally active recombinant protein in these preparations was determined by titration against E-64.
Initial rates of hydrolysis of the fluorogenic peptide substrates shown in table 2 were monitored by the release of the fluorogenic leaving group, NHMec, at an excitation wavelength of 380 nm and an emission wavelength of 460 nm using a Bio-Tek KC4 microfluorometer. kcat and Km values were determined using nonlinear regression analysis. Initial rates were obtained at 37°C over a range of substrate concentrations spanning Km values (0.2–200 µM) and at fixed enzyme concentrations (0.5–5 nM). Assays were performed in 100 mM sodium phosphate buffer (pH 6.0) containing 1 mM DTT and 1 mM EDTA. Rate constants for the inactivation of the Fasciola enzymes by Z-Phe-Ala-CHN2 and cathepsin K inhibitor II were determined from progress curves in the presence of substrate as previously described [22].
Calf skin collagen type I and bovine nasal septum collagen type II (solubilised in 0.2 M acetic acid at a concentration of 2 mg/ml) were dialysed for two days against 0.1 M sodium acetate (pH 5.5) or PBS (pH 7.0). Digestion reactions contained 10 µg of dialysed collagen substrates, 1 mM DTT and 1 mM EDTA and 2 µM activated FhCL1, FhCL2 or FhCL3 in a final volume of 100 µl of one of the above buffers at 28°C. For collagen type 1, reactions were performed for 3 h (pH 5.5) or 20 h (pH 7.0) whilst collagen type II was digested over 13–18 h. All reactions were stopped by the addition of 10 µM E-64. Digests were analyzed on reducing 4–12% NuPage Bis-Tris gels and visualised by staining with Flamingo fluorescent stain (Bio-Rad).
For digestion of collagen-like peptide substrates, 20 µg of Z-Gly-Pro-Leu-Gly-Pro and Z-Gly-Pro-Gly-Gly-Pro-Ala in DMSO were incubated with FhCL2 or FhCL3 (15 µM) in 100 mM sodium acetate buffer, pH 4.5, containing 1 mM EDTA and 2 µM DTT for 30 min at 37°C. Digestion reactions were halted by the addition of 10 µM E-64.
Recombinant FhCL2 and FhCL3 were removed from collagen type I digests using Ni-NTA agarose. The reactions were then spun at 13,000 rpm for 15 min to remove particulates and were concentrated to a final volume of 15 µl using a Concentrator 5301 (Eppendorf). Using an Eksigent AS-1 autosampler connected to a Tempo nanoLC system (Eksigent, USA), 10 µL of the sample was loaded at 20 µl/min with MS buffer A (2% acetonitrile+0.2% formic acid) onto a C8 trap column (Michrom, USA). After washing the trap for three minutes, the peptides were washed off the trap at 300 nL/min onto an IntegraFrit column (75 µm×100 mm) packed with ProteoPep II C18 resin (New Objective, Woburn, MA). Peptides were eluted from the column and into the source of a QSTAR Elite hybrid quadrupole-time-of-flight mass spectrometer (AB Sciex) using the following program: 5–50% MS buffer B (98% acetonitrile+0.2% formic acid) over 15 minutes, 50–80% MS buffer B over 5 minutes, 80% MS buffer B for 2 minutes, 80–5% for 3 min. The eluting peptides were ionised with a 75 µm ID emitter tip that tapered to 15 µm (New Objective) at 2300 V. An Intelligent Data Acquisition (IDA) experiment was performed, with a mass range of 375–1500 Da continuously scanned for peptides of charge state 2+–5+ with an intensity of more than 30 counts/s. Selected peptides were fragmented and the product ion fragment masses measured over a mass range of 50–1500 Da. The mass of the precursor peptide was then excluded for 15 seconds. Peak list files generated by MSX (Infochromics) were exported to a local PEAKS Studio v5.0 (Bioinformatics Solutions Inc.) search engine for protein database searching. MS/MS data was used to search a custom-made database containing only bovine collagen sequences. The enzyme specificity was set to “no enzyme” and propionamide (acrylamide) modification of cysteines was used as a fixed parameter and oxidation of methionines was set as a variable protein modification. The mass tolerance was set at 100 ppm for precursor ions and 0.2 Da for fragment ions. Only 1 missed cleavage was allowed. Matched peptides achieving a score >60% were accepted during PEAKs searches [16]. The matching peptides were then mapped onto the primary amino acid sequence of bovine collagen to identify FhCL2 and FhCL3 cleavage sites and to plot P2 residue preference for each enzyme.
For collagen-like peptide substrates, digests were concentrated and analysed by MS/MS as described above with the following modifications. The mass range of 150–600 Da was scanned for peptides of charge state 2+ with an intensity of more than 100 counts/s. Selected peptides were fragmented and the product ion fragment masses measured over a mass range of 50–600 Da. The mass of the precursor peptide was then excluded for 120 seconds. An inclusion list describing all possible substrate ions that could be produced by enzymatic cleavage of the peptide substrates was generated and programmed into the Analyst acquisition software. The resulting data files were manually interrogated to determine the presence of peptide ions described in the inclusion list. The MS/MS spectra of those peptides were de novo sequenced for b and y ion fragments describing the peptide substrate's sequence to a mass accuracy of approximately 0.2 Da.
For the MD simulations, starting coordinates for F. hepatica cathepsin L were taken from the 1.4 Å resolution crystal structure of a FhCL1 mutant zymogen, in which the active site Cys was replaced by Gly ([22]; PDB 2O6X). The prosegment (residues 1–100) was removed and the active site Gly mutation reversed to the wild type Cys. Initial coordinates for a template peptide substrate (Ala-Leu-Ala-Leu-Pro) were derived from X-ray structures of inhibitors bound to human cathepsin K ([31]; PDB 1NLJ) and bovine cathepsin B ([32]; PDB 1SP4) after structural alignment with FhCL1. This initial peptide was altered to Ala-Leu-Arg-Asn-Ala using the mutate function in Swiss-PdbViewer ([33]; http://spdbv.vital-it.ch/) and then minimized while bound to the wild-type FhCL1 using the equilibration protocol described below. The equilibrated peptide was then extended by one Ala residue at its N- and C-termini using the coordinate generation function in the psfgen program [34], and then re-equilibrated. The resultant peptide (ligand A) was used to generate all other substrate starting coordinates by using the mutate function in Swiss-PdbViewer. Mutations to FhCL1 were similarly generated using Swiss-PdbViewer. Rotamers for mutated enzyme and substrate side-chains were chosen by visual inspection and using the rotamer score provided in Swiss-PdbViewer. The N-terminal residue of the substrate was acetylated and the C-terminus N-methylamidated. Each complex was optimally oriented to minimize cell volume [35] and solvated in a truncated octahedral periodic cell with a minimum of 20 Å between periodic images of the protein. The system was neutralized with sodium ions.
MD simulations were carried out with NAMD 2.6 [34] using the CHARMM27 force field with φ/ψ cross-term map corrections [36]. Parameters for Hyp were from Veld and Stevens [37]. Water molecules were simulated with the TIP3P model [38]. Simulation conditions were maintained at 1.0 atm constant pressure by the Nosé-Hoover Langevin piston method [39], [40] and at 310 K constant temperature by Langevin dynamics with a damping coefficient at 5 ps−1. The time step used for the simulations was 1.5 fs. A cutoff of 12 Å, with a switching function between 10 and 12 Å, was used for short-range non-bonded interactions. Long-range electrostatic interactions were computed using the particle mesh Ewald method [41] with a grid density of approximately 1/Å. A multiple time-stepping algorithm was used with interactions involving covalent bonds and short-range non-bonded interactions computed every time step, while long-range electrostatic forces were computed every two time steps. SHAKE [42] and SETTLE [43] were applied to constrain the lengths of all bonds that involve hydrogen.
The solvated starting structure was minimized using conjugate gradient minimization to a 0.5 kcal/(mol·Å) r.m.s. gradient with all enzyme heavy atoms fixed, with the exception of side-chain atoms of mutated residues, which were unrestrained. In addition, in this phase of the equilibration, ligand atoms were not fixed and harmonic positional constraints of 100 kcal/(mol·Å2) force constant were placed on the Cα atoms of ligand residues 3–5 (P2, P1 and P1′). The unrestrained atoms were then further minimized during a 50 ps molecular dynamics run at 310 K. This starting model was then minimized with harmonic positional constraints on the NCαCO backbone of the protein and Cα atoms of ligand residues 3–5. A 100 kcal/(mol·Å2) force constant was used to minimise the system to a 0.5 kcal/(mol·Å) r.m.s. gradient. The constraints were gradually removed by subsequent minimizations to a 0.1 kcal/(mol·Å) r.m.s. gradient, scaling the initial force constants by factors of 0.5, 0.15, 0.05, and 0. The unrestrained minimized structure was then heated from 50 K to 310 K in steps of 25 K using velocity reassignment during a 30 ps molecular dynamics run. The equilibrated system was then used for production runs with no restraints. All systems were run for 12 ns. All simulations remained stable to completion. For analysis, the distance between the sulphur atom of the active Cys residue and the scissile backbone carbonyl carbon of the substrate (S-C distance) was recorded every 50 time-steps (0.075 ps); trajectory coordinates were recorded every 1000 time-steps (1.5 ps).
The free energy of binding of the peptide ligand to the peptidase contains an enthalpic and an entropic contribution. Free energy analysis of the production trajectories employed the single-trajectory MM/PBSA method combined with a determination of the change in configurational entropy using the harmonic approximation of normal-mode analysis [44], [45]. Snapshots from the MD trajectory, stripped of water and counterions, were analysed. The enthalpy of binding is composed of the change in the molecular mechanics free energy upon complex formation, and the solvated free energy contribution. The molecular mechanics free energy difference was calculated using the SANDER module in AMBER 9 [46], with no cutoff for the non-bonded energies and the AMBER ff03 force field to describe the protein and peptide ligands [47]. Compatible parameters for Hyp were not available and binding energies for ligand F were not calculated. The AMBER PBSA module was used for the evaluation of the electrostatic free energy of solvation. A grid density of 3/Å was employed for the cubic lattice, the internal and external dielectric constants were set to 1 and 80, respectively, and 1000 linear iterations were performed. The non-polar solvation free energy was calculated from the solvent accessible surface area using the MSMS program [48], with a probe radius of 1.4 Å, the surface tension set to 0.00542 kcal/(mol·Å2), and the off-set to 0.92 kcal/mol·m.
The changes in configurational entropy upon ligand association were estimated by an all-atom normal-mode analysis performed with the AMBER NMODE module. Prior to the normal mode calculations, the complex, receptor, and ligand were subjected to minimization with a distance dependent dielectric constant 4r and convergence tolerance tighter than a root-mean-squared gradient of drms 10−4 kcal/(mol·Å). Entropy and enthalpy calculations on all peptidase-ligand systems are performed separately and averaged over equally spaced snapshots, extracted over the final 4.005 ns of the production phase. The mean of the binding free enthalpies and entropies of all the snapshots were computed and then summed to obtain the binding free energy. For the enthalpy calculations, snapshots were taken every 10.5 ps (381 frames), for the entropy calculations, snapshots were taken every 190.5 ps (21 frames). VMD [49] was used to prepare the initial simulation system and analyse trajectories. Structural figures were prepared with PyMol [50]. Simulaid (http://atlas.physbio.mssm.edu/~mezei/) was used in the preparation of the truncated octahedral cell [35] and to convert the NAMD dcd format MD coordinate trajectories to AMBER format for the MM/PBSA analysis.
Functionally active recombinant forms of the major cathepsin L peptidases of F. hepatica, FhCL1, FhCL2 and FhCL3, were expressed in yeast and isolated to homogeneity as previously described [22], [23]. To compare their biochemical substrate specificity the kinetic parameters (Km, kcat, and kcat/Km) for each enzyme was determined against a range of small fluorogenic peptide (predominantly tripeptide) substrates (Table 2).
FhCL1 most efficiently cleaved substrates containing hydrophobic residues at the P2 position such as the dipeptides Z-Leu-Arg-NHMec (kcat/Km 1,492,354 M−1 s−1), Z-Phe-Arg-NHMec (kcat/Km 64,912 M−1 s−1) and tripeptide Boc-Val-Leu-Lys-NHMec (kcat/Km 54,266 M−1 s−1). In contrast, tripeptide substrates containing Pro at the P2 position, including Tos-Gly-Pro-Arg-NHMec (kcat/Km 671 M−1 s−1), Boc-Ala-Gly-Pro-Arg-NHMec (kcat/Km 673 M−1 s−1), Tos-Gly-Pro-Lys-NHMec (kcat/Km 612 M−1 s−1) and Boc-Val-Pro-Arg (kcat/Km 478 M−1 s−1), were cleaved relatively poorly (Table 2).
In comparison to FhCL1, substrates with Phe and Leu in the P2 position were much less effectively cleaved by FhCL2 and even less so by FhCL3. The kcat/Km values for FhCL2 and FhCL3 against Z-Phe-Arg-NHMec were 6- and 65-fold lower, respectively, than that observed for FhCL1. Similarly, the kcat/Km values for Z-Leu-Arg-NHMec were 3.5- and 66-fold lower than FhCL1 for FhCL2 and FhCL3 respectively. By contrast, FhCL2 and FhCL3 cleaved Pro-containing substrates much more readily than FhCL1 with kcat/Km values of 18,559 M−1 s−1 (28-fold greater, FhCL2) and 95,774 M−1 s−1 (142-fold increase, FhCL3) for Tos-Gly-Pro-Arg-NHMec; 58,027 M−1 s−1 (86-fold increase, FhCL2) and 60,763 M−1 s−1 (90-fold increase, FhCL3) for Boc-Ala-Gly-Pro-Arg-NHMec; 13,746 M−1 s−1 (22-fold increase, FhCL2) and 36,419 M−1 s−1 (60-fold increase, FhCL3) for Tos-Gly-Pro-Lys-NHMec and 21,193 M−1 s−1 (44-fold increase, FhCL2) and 1,564 M−1 s−1 (3-fold increase, FhCL3) for Boc-Val-Pro-Arg-NHMec (Table 2). Collectively, these data highlight significant differences in the substrate specificity of the three major F. hepatica cathepsin L peptidases. More specifically, the data demonstrates that FhCL3 prefers a bulky Pro residue in the P2 position of substrates over hydrophobic residues such as Leu or Phe, while FhCL2 can readily accept Pro despite preferring hydrophobic moieties at P2, and FhCL1 has an inverse preference to FhCL3.
Peptidyl diazomethyl ketones are irreversible inhibitors of cysteine peptidases [51]. Changes in rates of inactivation by these inhibitors have highlighted different specificities at subsites of cysteine peptidases such as cathepsin L and cathepsin B [52]. In this study, we measured the rates of inactivation of FhCL1, FhCL2 and FhCL3 by the cathepsin inhibitor Z-Phe-Ala-CHN2. Both FhCL1 and FhCL2 were rapidly inactivated by Z-Phe-Ala-CHN2 with the rate of inactivation of FhCL1 being almost 2-fold higher than that of FhCL2 (Table 3). This is in accordance with our previous data [22] and demonstrates that FhCL1 accommodates hydrophobic P2 residues more effectively than FhCL2. In contrast, the rate of inactivation of FhCL3 by Z-Phe-Ala-CHN2 was 20-fold times lower, showing that Z-Phe-Ala-CHN2 is a poor inhibitor of FhCL3 (Table 3). This is in agreement with our kinetic substrate data using peptidyl fluorogenic substrates (Table 2) that revealed the poor capacity of FhCL3 to accommodate hydrophobic residues in the P2 position.
The inhibitor known as cathepsin K Inhibitor II (Z-LNHNHCONHNHLF-Boc, CKII) is a potent time-dependent inhibitor of human cathepsin K; its selectivity for this enzyme is largely because of the effectiveness by which Leu occupies the S2 subsite [53]. FhCL1 and FhCL2 were both potently inhibited by cathepsin K inhibitor II with Ki values of 0.63 nM and 0.46 nM respectively. In contrast, CKII was 14-fold less effective against FhCL3 (Ki 336 nM) compared to FhCL1 and 20-fold less effective compared to FhCL2 (Table 3). The data are consistent with the kinetic data for hydrolysis of peptidyl fluorogenic substrates as both FhCL1 and FhCL2 had high kcat/Km values for Z-Leu-Arg-NHMec whereas that of FhCL3 against this substrate was much lower (Table 2).
The α chains of collagens are woven together to form triple helical, or Col, regions of collagen. These are flanked by non-collagenous, or non-helical, regions termed NC domains Type I and type II collagens are most abundant in nature and are the major components of vertebrate connective tissue. They share ∼70% primary sequence identity and are composed largely of repeating Gly-X-Y motifs [27]. FhCL1, FhCL2 and FhCL3 effectively degraded type I collagen at pH 5.5 which induces a denaturation of the protein's helical Col structure. However, FhCL1 was much less able to degrade type I collagen at pH 7.0, where its native structure is preserved, and its activity was limited to the β and γ chains of the NC domains leaving the α1 and α2 chains of the Col domain intact (Fig. 1A). By contrast, both FhCL2 and FhCL3 degraded native collagen at pH 7.0 and cleaved efficiently within the Col helical structures as revealed by the breakdown of the α1 and α2 chains (Fig. 1A).
To determine the relative activity of FhCL1, FhCL2 and FhCL3 for collagen type I, digests were performed at pH 7.0 over a time course (up to 18 h) at 28°C (Fig. 1B). Only FhCL3 was capable of completely digesting collagen type I after 18 h incubation in these conditions. FhCL2 digested collagen α chains to a lesser extent than FhCL3 while FhCL1 only digested the β11 and β12 dimers but not the collagen α chains (Fig. 1B). Similarly, only FhCL3 was capable of degrading type II collagen whilst FhCL2 displayed much less activity against this substrate at pH 7.0 (Fig. 1C). FhCL1 was unable to cleave within the tightly wound type II collagen helices under these conditions (Fig. 1C).
To identify the cleavage sites for FhCL2 and FhCL3 within collagen type I α1 and α2 chains, the 18 h reaction aliquots (shown in Fig. 1B) were analysed by tandem mass spectrometry to determine the masses and sequence identities of the resulting hydrolytic products. Liberated peptides were mapped onto the primary amino acid sequence of bovine collagen to identify the cleavage sites of the F. hepatica peptidases (Fig. 2). FhCL2 cleaved collagen type I at 43 sites within the α1 chain and 26 sites within the α2 chain while FhCL3 cleaved at 24 sites within the α1 chain and 24 sites within the α2 chain. Strikingly, only three of these cleavage sites were shared between FhCL2 and FhCL3, all of which occurred in the α1 chains (Fig. 2).
We examined the frequency of each amino acid in the P1, P2 and P3 position of the proteolytic cleavage sites identified in the collagen digests described above to determine preferences for binding their respective active site S1, S2 and S3 subsites (Fig. 3). While substrate residues present at the P2 position from the scissile bond interact with the S2 subsite of the active site of papain-like cysteine peptidases are considered most critical in determining the efficiency by which the P1-P1′ bond is cleaved [54], the binding of these residues are influenced by residues in the P3 positions. Consistent with our previous findings using positional scanning of synthetic combinatorial libraries the P1 position can be occupied by many different amino acids without a strong preference [22]. However, specificity is observed in the P2 position; Gly was most commonly found in the P2 position of the FhCL2 cleavages (27%), and this was followed by Leu (21%) and Pro (18%) (Fig. 3). By contrast, FhCL3 displayed a highly specific preference for Gly at the P2 position (44% of all cleavages) with a weak preference for all other amino acids including Leu and Pro in this position (3% for both residues, Fig. 3). The P3 and P4 positions were occupied by a wide range of amino acids.
To further investigate the cleavage of collagen by FhCL2 and FhCL3, the ability of both enzymes to cleave two small peptide substrates, Z-Gly-Pro-Leu-Gly-Pro and Z-Gly-Pro-Gly-Gly-Pro-Ala, mimicking the repeating Gly-X-Y motifs (where X is often Pro) that occur within the collagen primary sequence was followed by tandem mass spectrometry. The presence of several peptides matching hydrolytic cleavage products showed that FhCL2 and FhCL3 were able to digest both substrates (Fig. 4). The cleavage pattern of peptide Z-Gly-Pro-Leu-Gly-Pro was identical for both FhCL2 and FhCL3. However, while FhCL2 cleaved the peptide Z-Gly-Pro-Gly-Gly-Pro-Ala at three sites (with Gly or Pro in the P2 position), FhCL3 was unable to cleave at one of these three sites where Pro occupied the P2 position (Fig. 4).
In order to delineate the molecular basis of the ability of FhCL2 and FhCL3 to digest collagen, our recently determined crystal structure of FhCL1 was used as the starting point for a computational analysis of ligand binding. Complexes of FhCL1, with variations to key residues involved in substrate binding (summarised in Table 4), bound to different seven-residue peptides (Table 5), were analysed by performing MD simulations. Using the simulation trajectories, free energies of binding of the peptide substrates were calculated using the well-established MM-PBSA method [44], [45]. In addition, the distances between the nucleophilic sulphur atom of the active site Cys residue and the backbone carbonyl carbon atom of the scissile peptide bond, were examined over the course of the simulations. Since higher frequencies of close approach of these atoms would likely correlate with higher frequencies of formation of the transition state of the hydrolysis reaction this measure gives an indication of how well the substrate fits into the binding cleft and how readily it is cleaved [55]. Fig. 5 illustrates the critical residues of the active site investigated and their disposition in FhCL1 relative to the bound peptide substrate ligand A (AALR*NAA, shown as an example, asterisk represents position of scissile bond) and in FhCL2 bound to ligand C (AGPR*NAA). Table 5 presents the results of the binding energy calculations, as well as the average nucleophilic sulphur-scissile carbon (S-C) distances for the various peptidase-ligand complexes simulated. Fig. S1 illustrates the regions of the peptidase that contact the ligand during the FhCL1 ligand A simulation.
The MD simulations indicate that, for wildtype FhCL1, activity is greatest for substrates with Leu at P2 and that Arg is favoured at P1 (consistent with our substrate and inhibitor binding kinetics shown in Table 2 and 3). Thus, the results for the FhCL1-ligand A complex (Table 4; Fig. 5A) are taken as a benchmark against which the other results are compared. The free energy of ligand binding is related to the dissociation constant Kd by the formula ΔG = −RT ln Kd. Thus, the calculated binding energy for the FhCL1-ligand A complex of 10.83 kcal/mol corresponds to a Kd of 22.9 nM, while the approximate level of error in the free energy calculations of 1 kcal/mol corresponds to a 5-fold difference in Kd. Where differences between calculated binding energies are greater than the error bounds, the calculations are taken to predict differences in binding affinities. The calculations for ligand A (AALR*NAA), which has Leu at P2, thus discriminate between binding affinities for FhCL2 and FhCL3, predicting an approximately 5–10 fold difference in Kd, which correlates well with the inhibition constants determined for the CKII inhibitor (Table 3), which also has Leu at P2.
The calculations also agree with the experimental data in suggesting reduced activity of FhCL1 against a ligand with Pro at P2 (ligand B, AAPR*NAA) compared to Leu (ligand A). However, they also predict that for ligand B, FhCL1 has a higher binding affinity than FhCL2 and equal or greater activity than FhCL3. When ligand B is altered such that Gly is substituted for Ala at P3 (ligand C, AGPR*NAA), the binding affinities for the peptidases with FhCL2 or FhCL3 S2 subsites show a marked increase over those for ligand B. Thus, the data suggest that the collagenolytic activity of FhCL2 and FhCL3 may not be due simply to the P2 Pro-S2 subsite interaction, and that Gly at ligand residue P3 is also important, consistent with our earlier suggestions [23]. This inference is consistent with our previously reported experiments using combinatorial libraries [22] that showed FhCL2, strongly favoured a Gly at P3, and with our present data using native collagen digestion which indicated that Gly is favoured at P3 for both FhCL2 and FhCL3. Analysis of our collagen digest revealed that of the 11 cleavage sites for FhCL2 and FhCL3 (Tyr and Trp at position 67, respectively) containing Pro at P2 (seven for FhCL2 and four for FhCL3), 8 had Gly at P3. Given their similar active site residues to FhCL2 and FhCL3 we also analysed previous studies with human cathepsin K [56] and ginger rhizome GP2 [29] (also possess Tyr and Trp at position 67, respectively) and observed that of the 12 peptidase cleavage sites within native collagen type I containing Pro at P2 (eight for cathepsin K and four for GP2), 10 had Gly at P3. Examination of the simulation trajectories suggests that Gly (that lacks a side-chain) at P3, would offer minimal steric interference with the large active site Tyr or Trp side-chain at position 67 in FhCL2 and FhCL3, respectively, allowing the Tyr or Trp ring to form a “lid” over the ligand's P2 Pro ring, helping to sequester it in the S2 subsite (Fig. 6).
Analysis of the cleavage sites within native type I collagen show that both FhCL2 and FhCL3 have a strong preference for Gly at P2, most particularly for the latter enzyme (Figs. 2 and 3). To investigate the molecular basis of this preference, simulations were performed using ligands with Gly at P2 (ligands D and E). The simulations with Gly at P2 generally showed markedly greater S-C distances than were observed in the complexes with Leu or Pro at P2 (Table 5). This supports the idea that the P2-S2 interaction has a strong influence on the S-C interaction. The binding affinity of collagen-like ligand D (which has Ala at P3, PAGP*AGP) is substantially higher for FhCL2 compared wildtype FhCL1, but when in complex with FhCL3, ligand D essentially disengages. However, when ligand D is altered such that Leu occurs at P3 (ligand E, PLGP*AGP), binding affinity to FhCL3 is restored but greatly reduced in the complex with FhCL2. These results further support the idea that the interaction of ligand residue P3 with the peptidase is a significant factor in ligand binding, and possibly of greater importance when Gly is at P2.
Collagens comprise polypeptide chains containing the repeating triplet sequence Gly-Pro-Y where 4-hydroxyproline (Hyp) commonly occupies the Y position [57]. Thus, simulations of complexes with a ligand containing Hyp at P3 and P1′ and Gly at P2 (ligand F, PPGP*PGP) were performed. For the FhCL2 variant, the ligand began to disengage from the peptidase whilst for FhCL3 the ligand remained closely bound. Although the binding affinity for the FhCL3-ligand F complex was not calculated, the average S-C distance was much lower than for the other complexes with Gly at P2 (Table 5). Moreover, the plot of the S-C distance frequencies showed high frequencies of very close approach for the FhCL3-ligand F complex (Fig. 7). These data suggest that FhCL3 is able to digest collagen with Hyp-Gly at P3-P2 whereas FhCL2 cannot. This may explain why we observed a greater ability of FhCL3 to digest type I and II collagen compared to FhCL2 (Fig. 1).
The FhCL3-ligand F simulation trajectories revealed that the side-chain of Trp 67 occupies the FhCL3 S2 subsite and sits against the peptide backbone of the ligand P2 Gly (Fig. 6A). This may stop solvent from entering the S2 subsite and interacting with the ligand P2 Gly, thus “sealing” the ligand in the enzyme's binding cleft. A similar disposition of the Trp 67 side-chain was also observed in the FhCL3-ligand E simulations. Tyr at position 67 in FhCL2 behaves in a similar manner to the Trp 67 of FhCL3 when binding ligands with Gly at P2. Thus, Tyr 67 occupied the S2 subsite cleft and contacted the peptide backbone of the ligand P2 Gly in the FhCL2-ligand D complex (Fig. 6B). The position of the Tyr side-chain was further stabilised by a hydrogen bond between its hydroxyl oxygen and the backbone oxygen of residue 157.
F. hepatica has evolved a repertoire of cathepsin L peptidases as a result of gene duplication and diversification that exhibit subtle but distinct substrate specificities [1], [8], [14], [15]. The expression of different members of this peptidase family is temporally regulated suggesting that they perform precise functions at different stages of the parasites' development [13]. This idea is supported by our previous data showing that the predominant enzyme, FhCL1, secreted by the mature adult parasites, which are obligate blood-feeders, is adapted to the degradation of host haemoglobin; the S2 subsite of the FhCL1 active site, which contributes mostly to substrate binding, readily accommodates P2 residues such as Leu, Ala, Val and Phe that together represent >40% of the residues present in haemoglobin [3].
FhCL1 does not readily accept Pro into the S2 subsite as shown in this and other [22], [28] studies and thus it's activity against type I and II collagens observed here was restricted to the non-collagenous, NC, domains. By contrast, both FhCL2 and FhCL3 have evolved to accommodate Pro in the S2 subsite of their active sites; this property has been attributed to the presence of Tyr and Trp, respectively, at position 67 within the S2 subsite of these enzymes, a position that is occupied by Leu in FhCL1 [1], [14], [15], [22], [23]. In this study, our computational data show that Tyr and Trp at position 67 have the ability to function in distinct ways to accommodate either Gly or Pro residues at P2 and explains why FhCL2 and FhCL3 have an ability to degrade the Gly-X-Y containing Col helices of collagen. The results are also in accordance with our previous suggestion that the interaction of substrate residue P3 with the peptidase is a significant factor in substrate binding, in particular when Gly or Pro is at P3 [23]. This is also the case when we compare FhCL3 specificity towards synthetic peptides; Pro is readily accepted in P2 only when Gly is at P3 (Tos-Gly-Pro-Arg-NHMec, Tos-Gly-Pro-Lys-NHMec and Boc-Ala-Gly-Pro-Arg-NHMec) but not when Val is at P3 (Boc-Val-Pro-Arg-NHMec) (Table 2). Therefore, the collagenolytic activity of FhCL2 and FhCL3 is not due simply to the P2 Pro-S2 subsite interaction, and with Pro at P2, Gly at residue P3 is critical. A comparison of the cleavage sites of FhCL2, FhCL3, human cathepsin K [56] and ginger rhizome GP2 [29] revealed that many of their cleavage sites within collagen where Pro is at P2, a Gly is present at P3. A P3 Gly, which lacks a side-chain, offers minimal steric interference with the large active site Tyr or Trp side-chain at position 67, and allows the Tyr or Trp ring to form a “lid” over the ligand's P2 Pro ring, helping to sequester it in the S2 subsite (Fig. 6).
However, we observed that FhCL3 digested type I and II collagens more efficiently compared to FhCL2 and that these two enzymes cleave at mostly different sites (see Figs. 1 and 2). The computational data indicates that this may be, in part, because FhCL3 binds substrates containing a P3 and P1′ Hyp much tighter than FhCL2. A difference between these two enzymes was also observed using a peptide substrate that mimics the Gly-X-Y repeat in the collagen Col domain, Z-Gly-Pro-Gly-Gly-Pro-Ala; FhCL2 cleaves at three sites with Gly or Pro in the P2 position, whereas FhCL3 was unable to cleave at one of these three sites despite having Pro occupying the P2 position and Gly at the P1 position. This result contrasts with our data using fluorogenic peptide substrates which showed that FhCL3 cleaved the tripeptides Tos-Gly-Pro-Arg-NHMec and Tos-Gly-Pro-Lys-NHMec with 5- and 3-fold better efficiency, respectively, than FhCL2. On the other hand, the two enzymes exhibited equal efficiency for the substrate Boc-Ala-Gly-Pro-Arg-NHMec. The influence of P4 and P′ regions of the peptides on substrate binding in these two enzymes need greater attention in future studies when suitable reagents become available. Notwithstanding, it is clear that the modification within the active site of FhCL2 and FhCL3 (Tyr or Trp at position 67) has subtly altered the substrate specificity of the two enzymes such that they exhibit different substrate profiles without compromising their unique ability to degrade host native collagen.
FhCL3 is expressed by the invasive stage of F. hepatica which must quickly penetrate the wall of the intestine to enter its host [1], [13]. RNAi-mediated knockdown experiments have demonstrated that the secretion of this peptidase and a cathepsin B cysteine peptidase by these invasive parasites is critical to invasion of the intestinal tissue [2]. Once the intestine has been traversed expression of these enzymes is switched off and the parasite up-regulates expression and secretion of FhCL1 and FhCL2 which are required to facilitate tunnelling through the liver mass and feeding on host tissue (the parasite undergoes rapid growth at this stage) [16]. The collagenolytic activity of FhCL3 and FhCL2 is important in degrading the extracellular matrix of the tissues through which this parasite moves. While collagenase activity has been demonstrated in ginger rhizome cysteine peptidases [29], [58], only one other animal cysteine peptidase, human cathepsin K which functions in bone re-modelling [59], possesses collagenase activity. Accordingly, the evolution of this activity in F. hepatica must represent an important step in the development of a parasitic way of life.
|
10.1371/journal.pntd.0005624 | Radiological evolution of porcine neurocysticercosis after combined antiparasitic treatment with praziquantel and albendazole | The onset of anthelmintic treatment of neurocysticercosis (NCC) provokes an acute immune response of the host, which in human cases is associated with exacerbation of neurological symptoms. This inflammation can occur at the first days of therapy. So, changes in the brain cysts appearance may be detected by medical imaging. We evaluated radiological changes in the appearance of brain cysts (enhancement and size) on days two and five after the onset of antiparasitic treatment using naturally infected pigs as a model for human NCC.
Contrast T1-weighted magnetic resonance imaging with gadolinium was performed before and after antiparasitic treatment. Eight NCC-infected pigs were treated with praziquantel plus albendazole and euthanized two (n = 4) and five (n = 4) days after treatment; another group of four infected pigs served as untreated controls. For each lesion, gadolinium enhancement intensity (GEI) and cyst volume were measured at baseline and after antiparasitic treatment. Volume and GEI quantification ratios (post/pre-treatment measures) were used to appraise the effect of treatment. Cysts from untreated pigs showed little variations between their basal and post treatment measures. At days 2 and 5 there were significant increases in GEI ratio compared with the untreated group (1.32 and 1.47 vs 1.01, p = 0.021 and p = 0.021). Cyst volume ratios were significantly lower at days 2 and 5 compared with the untreated group (0.60 and 0.22 vs 0.95, p = 0.04 and p = 0.02). Cysts with lower cyst volume ratios showed more marked post-treatment inflammation, loss of vesicular fluid and cyst wall wrinkling.
A significant and drastic reduction of cyst size and increased pericystic enhancement occur in the initial days after antiparasitic treatment as an effect of acute perilesional immune response. These significant changes showed that early anthelmintic efficacy (day two) can be detected using magnetic resonance imaging.
| Neurocysticercosis (NCC) is a frequent parasitic infection of the human brain and the most common cause of adult onset epilepsy in developing countries. Acute inflammatory response in NCC plays an important role in the pathogenesis of symptoms by anthelminitic therapies. The anthelmintic recommended therapy for NCC has drawbacks as the exacerbation of inflammation around degenerating cysts provokes the appearance of symptoms at the first days of treatment. Radiological changes in the appearance of cysts usually are seen after months of therapy. To evaluate if significant radiological changes (enhancement and size) occur in the first days of therapy, we used a porcine NCC model and magnetic resonance imaging (MRI) with contrast solution. The major radiological changes observed after treatment with albendazole and praziquantel were an increase in enhancement and the significant reduction in cyst size by day 2 and more evident on day 5. Cysts with greater changes also experienced exacerbated inflammation, loss of vesicular fluid and wrinkling of the cyst wall. These results show an early therapeutic effect and the possible utility of repeat MRI imaging within a few days after starting treatment. Finally, these findings contribute to our understanding of the treatment induced early exacerbation of symptoms.
| Neurocysticercosis (NCC) is a neurological parasitic disease caused by the infection of the brain by the larval stage of Taenia solium [1]. NCC represents a serious and persisting public health problem because it is the most frequent cause of late-onset seizures in developing countries [1, 2].
Treatment with anthelmintic drugs such as praziquantel and/or albendazole has been associated with increased severity of symptoms within the first days of therapy [3–7]. Even though praziquantel and albendazole have different mechanisms of action [8, 9], both drugs cause the destruction of cysts and subsequent release of antigens, triggering the host immune response [7, 10–13]. Using the porcine NCC model and the antihelmintic drug praziquantel, this acute post-treatment inflammatory response was associated with pericystic inflammation [14] accompanied by an increase of vascular permeability, pro-inflammatory and regulatory cytokine profiles [15] during the second and fifth day. Using the same model, radiological changes in the appearance of brain cysts have been reported after two weeks of praziquantel treatment [16–18]. Similarly, the use of albendazole in the porcine model resulted in an increase of pro-inflammatory cytokines [14].
Medical imaging has been a useful tool in the diagnosis and medical follow-up of NCC patients [7]. Cyst appearance, size, perilesional enhancement and edema are imaging criteria to determine the radiological resolution of NCC after treatment [19]. The earliest radiological changes related to the size and appearances of brain cysts after conventional anthelmintic treatment has been reported during the first week of treatment in humans [20, 21] and in pigs after two weeks [16–18]. However, the radiological evolution of brain cysts during the first days of treatment, when perilesional inflammation establishes and symptoms increase in treated patients, has been scarcely explored.
In the present study, we evaluated the early radiological changes on MRI following the onset of antiparasitic treatment (days two and five) in pigs naturally infected with T. solium as a model for human NCC and confirmed the radiological findings with an ex-vivo histopathological examination.
A total of twelve pigs naturally infected with Taenia solium cysticercosis were obtained in endemic villages, transported to our facilities in Lima, and randomly divided in three groups, control or untreated, PZQ+ABZ 2d and PZQ+ABZ 5d, as follows: Four pigs remained untreated as a control group and 8 pigs were treated with the same combination of anthelmintic drugs and sacrificed at two (n = 4) and five (n = 4) days after treatment. The treatment consisted of combined therapy with praziquantel (Helmiben, Farmindustria, Peru) given for only the first day at 75 mg/kg/day, divided into three doses of 25 mg/kg administered every two hours [10], and albendazole (Zentel, GlaxoSmithKline, Peru) given daily until sacrifice at 15 mg/kg/day [22].
All pigs had pre and post-contrast MRI before treatment (Pre-treatment MRI) and on the day of sacrifice (Post-treatment MRI). Two hours before sacrifice, an Evans blue solution was infused as previously reported [15]. For all interventions, pigs were anesthetized with an intramuscular injection of a mixture of ketamine (Ket-A-100 50 mg/kg, Agrovet Market SA, Peru) and xylazine (Dormi-Xyl 2mg/kg, Agrovet Market SA, Peru) [15].
After euthanasia, the pig brains were placed on dry ice slabs and cut in 1-cm sections. Cysts with pericystic capsules were collected from both hemispheres for histopathology and RNA studies. Specimens from the right hemisphere were fixed in 10% neutral buffered formalin, embedded in paraffin and then sectioned coronally at 4 μm thickness. Conventional hematoxylin-eosin was performed on every slide and two sections were examined with conventional light microscopy. Microphotographs were taken at 15X magnification with a Carl Zeiss stereoscope with AxioVision software to obtain a single large image (“cyst map”) [22].
Pre- and post-treatment GEI, pre- and post-treatment cyst volume, cyst volume ratio, GEI ratio, Inflammatory Score Composite (ISC) and cyst damage score composite (CDSC) were all continuous parameters. Treatments groups were used as a categorical variable (untreated, been treated at 2d and 5d). Mann Whitney test was used to compare pre-GEI and pre Cyst volume between the different treatment groups. Pre-post treatment differences for GEI and cyst volume were analyzed by the Wilcoxon test in each treatment group, individually. To evaluate if the mean change in GEI and cyst volume from pre to post-treatment measures differed in the three groups, we used a generalized estimating equation (GEE) analysis. To verify those post-treatment differences (cyst volume and GEI) truly result from treatment rather than from left-over effects of (usually random) pre-test differences between groups, we used an analysis of covariance (ANCOVA) with pre-treatment measures as covariates. Finally, we used the Mann-Whitney U test to compare ratios (changes between pre- and post-treatment measures) of GEI and cyst volume between treatment groups. Since ratio analysis results were highly correlated with unstandardized group analyses, we used ratios for the correlations with histopathology. Spearman correlation was used to assess the relation between each radiological (GEI and cyst volume) and histopathological (ISC and CDSC) parameters. All statistical analyses were performed using software R program for Windows, version 3.2.2. Graphs were performed using the ggplot2 package [25]. Values of p under 0.05 were considered to be statistically significant.
The study was conducted in accordance with the National Institutes of Health/AALC guidelines, and was reviewed and approved by the Institutional Ethics Committee for Animal Use at Universidad Peruana Cayetano Heredia (assurance number A5146-0).
The study animals were seven male and five female pigs. Their weight range was variable (mean: 69.8 kg; range: 15–120 kg). A total of 328 brain cysts were obtained from the 12 pigs. The parasite cyst burden in each pig brain was also variable (mean: 27.3; range: 1–152) (Table 1) [22].
A number of estimates of GEI showed increases around cysts in treated pigs compared to cysts in control untreated pigs. At baseline (before antiparasitic treatment), cysts in pigs from the Control and PZQ+ABZ 2d groups had higher GEI than cysts from the PZQ+ABZ 5d group (30.22 and 28.32 vs. 24.1, p<0.05). Post-treatment GEI values were higher in both treated groups compared with control pigs (PZQ+ABZ 2d: 36.04 and PZQ+ABZ 5d: 35.8 vs. Control: 33.31, p<0.001). When comparing pre- and post-treatment GEI in each group, there were marginal differences in cysts from control animals (30.22 vs. 33.31, p = 0.048), while GEI around cysts in treated groups increased markedly (PZQ+ABZ 2d: 28.32 vs. 36.04, p<0.001 and PZQ+ABZ 5d: 24.1 vs. 35.8, p<0.001) (S1 Table). GEE analysis confirmed that the effect of treatment in increasing the enhancement around cysts changed from basal to days 2 and 5 (RC for interaction term between pre-post GEI measures and groups: 4.996, <0.001) (S2 Table).
Additionally, after adjusting for pre-treatment differences, GEI increased significantly in both treated groups (PZQ+ABZ 2d: 7.324, p-value = 0.001 and PZQ+ABZ 5d: 9.442, p-value<0.001) compared with the control group (S2 Table).
Ratio analysis was also used to assess the increases in enhancement between groups (across time). Individual cyst GEI ratio (post-/pre-treatment GEI) demonstrated a similar effect (mean ratios were 1.01 for cysts of control pigs, 1.32 for cysts of pigs in PZQ+ABZ 2d group, and 1.47 in PZQ+ABZ 5d group; p = 0.021 between groups and p = 0.387 comparing both treatment groups) (Table 2, Fig 1).
On baseline MRI (before antiparasitic treatment), cysts from the control and PZQ+ABZ 5d groups had larger volumes (106.16 mm3 and 114.18 mm3, respectively) than those from PZQ+ABZ 2d pigs (74.56 mm3) (p<0.05). On post-treatment MRI, cysts from PZQ+ABZ 2d and PZQ+ABZ 5d groups had lower cyst volume than cysts from the control group (48.64 mm3 vs. 97.92 mm3, <0.001 and 24.36 mm3 vs. 97.92 mm3, p = 0.03). Cyst volume also decreased in the 5-d treated cysts compared to the 2-d treated cysts (48.64 mm3 vs. 24.36 mm3, p<0.001) (S1 Table).
Similar to GEI, pre- and post-treatment cyst volumes in control pigs were similar (106.16 vs. 97.92, p = 0.045), while post-treatment cyst volumes were significantly smaller in treated animals (PZQ+ABZ 2d: 74.56 vs. 48.64, p<0.001 and PZQ+ABZ 5d: 114.18 vs. 24.36, p<0.001) (S1 Table). Adjustment for pre-treatment measures in ANCOVA confirmed that cysts from both treated groups had smaller volumes than cysts from the control group (PZQ+ABZ 2d: -62.117, p-value = 0.014 and PZQ+ABZ 5d: -95.032, p-value<0.001). Similar to enhancement, GEE analysis confirmed that the effect of treatment on cyst volume was more marked at day 5 (RC for interaction term between pre-post measures and groups: -48.201, <0.001) (S2 Table).
A similar effect was also seen when individual cyst volume ratios (post-/pre-treatment) were compared between groups. Cyst volume ratio was lower (more reduction) in cysts from both treated groups than in those from the control group (0.60 for ABZ+PZQ 2d and 0.22 for ABZ+PZQ 5d vs. 0.95 for controls, <0.05) (Fig 1), demonstrating cyst volume reduction after treatment. However, cysts from pigs in PZQ+ABZ 5d group had similar volume reduction than did cysts in the PZQ+ABZ 2d group (0.22 vs. 0.60, p = 0.248) (Table 3).
Further analysis demonstrated a significant negative relationship between cyst volume ratio with GEI ratio after 5 days of treatment (r = -0.412, p<0.001) (S3 Table, Fig 2) suggesting that cysts with more enhancement (GEI) experience greater reduction in volume.
To confirm the radiological changes, we performed an ex-vivo examination to measure inflammation and the cyst damage using the ISC (inflammatory score-composite) and the CDSC (cyst damage score-composite), then we correlated those histological parameters with GEI and cyst volume ratios (radiological parameters).
Cysts from right brain hemispheres (n = 165) were selected for histopathological studies. Of these, only 105 cysts had a complete cyst structure and capsule and were therefore evaluable. Both treated groups had higher ISC and CDSC than the control group (p<0.001, Mann Whitney test). Both scores were higher at 5d compared to 2d, but there were no significant differences in these variables between both treated groups (ISC: 352 vs 304, p = 0.364; CDSC: 388 vs 336, p = 0.405 for CDSC) (S4 Table).
Higher ISCs were significant and positively associated with GEI ratio (r = 0.002, p = 0.028), meaning that cysts with higher increases in enhancement have more post-treatment pericystic inflammation. However, there was no significant correlation between GEI and CDSC (r = -0.001, p = 0.286) (S3 Table).
Interestingly, there was a significant negative relationship between cyst volume ratio (post-/pre-treatment measure) and post-treatment inflammation (ISC) at day 5 (RC = -0.002, p = 0.004), suggesting that cysts with increased inflammation showed increased reduction in volume. Slides of cysts with high volume ratio (higher reduction of cyst volume) showed loss of vesicular area and excess cyst wall folding upon themselves or wrinkling accompanied by granulomatous reaction (Fig 3A and 3C). In both cases, eosinophils have invaded the parasite’s wall as an effect of treatment (Fig 3B and 3D). This eosinophilic invasion has been observed before at points of high inflammation [26] and it is a demonstration of an acute response.
However, there was no significant relationship between volume ratio and cyst damage (CDSC) in any group (S3 Table).
Combined treatment of parenchymal NCC with praziquantel and albendazole destroys brain cysts in humans and pigs [1, 10], which is associated with a better clinical evolution in cases of human NCC [10]. However, after anthelmintic treatment humans are not usually reimaged until six or 12 months after treatment so early effects are not measured.
Despite the efficacy of combined treatment, in humans therapy causes an exacerbation of symptoms, usually seizures, due to acute inflammatory response to degenerating or dying cysts [27]. To assess early radiological changes, we examined MRI parameters of enhancement and cyst size and confirmed those findings with an ex-vivo histopathology (tissue-based semi-quantitative estimates of inflammation, and cyst damage) in naturally T. solium-infected pigs treated with albendazole and praziquantel at 2 and 5 days post initiation of treatment, compared to untreated control animals.
Enhancement has been associated to the disruption of the BBB in porcine NCC [22, 28, 29] as it happens in other diseases such as multiple sclerosis [30–32], gliomas, metastases and abscesses [33]. Earlier studies in NCC employing contrast-enhanced computed tomography (CT) in pigs [16–18] described the appearance of pericystic enhancement two weeks following praziquantel treatment. In humans, anthelmintic treatment also exacerbates gadolinium (Gd) enhancement during the first days of therapy [7], causing a change from an initial ring pattern of enhancement to a disc pattern, as seen using Gd T1-MRI [34]. These results are coherent with the post-treatment increase of enhancement reported in this study. We observed that the effect of treatment on enhancement increases with time already on day 2 and is further increased on day 5. Also, there was a positive correlation between increase of enhancement and inflammation. As enhancement is associated with BBB disruption, the following or parallel process that occurs is the extravasation of immune cells into the injured area and the increase of the inflammatory response. This agrees with previous studies where pro-inflammatory cytokines [14] and eosinophils where more abundant in pericystic tissues where the BBB had been disrupted [26].
Unexpectedly, we found that cyst volume was reduced very early after the onset of antiparasitic treatment. Reductions in cyst volume were evident in both treated groups on day two and were more pronounced five days after treatment, when the median of cyst volume loss was almost 78% ([1–0.22]*100; pre vs. post-treatment). Changes in the size of brain cysts in pigs had previously been reported after two weeks of praziquantel treatment [16]; however our findings suggest that sizable changes in the cyst size occur already by the second day of treatment. These results might have been more marked because we used combined therapy, and are consistent with early cyst size decrease observed on day 3 [21] and after one week [20] of antiparasitic treatment in humans. The reduction of the size of the parasite likely results from treatment-induced cyst damage and associated increased permeability of the cyst membrane, with a consequent increase in density of the cyst contents due to the influx of host albumin, protein coagulation, and loss of water [35].
The reduction of the size of the cyst was also accompanied with increased enhancement and inflammation. A previous study from our group reported that enhancement was associated with granuloma formation [22]; in this study we found similar results but additionally accompanied with cyst reduction. However, there was no association with cyst damage score (extension of the damage). A possible explanation could be that the combined treatment damages the scolex first, before damage is histologically noticeable and extended at the cyst wall level. Only afterwards would the cyst shrink and release fluid through the most heavily damaged regions of its wall. Similarly, a previous study concluded that the scolex is the primary target and its dissolution carries the complete resolution of the cyst [36]. As for cysts with little or no enhancement with a negligible change in size, they would represent those cysts in patients that do not respond to drug therapy, although our suggestion is valid only up to five days.
Despite these significant findings, our study has some limitations. We used a small number of animals and the parasite load per pig brain was very variable, making it difficult to compare groups. However, we used three statistical analyses to handle baseline differences to truly measure the effect of treatment. The variable thickness of MRI scans introduced some noise in the measurements of enhancement and volume; nevertheless, cyst volume and enhancement were significantly different in the treated groups. Minor drawbacks include use of only one hemisphere for histopathological assessments; however, differences in cyst load between hemispheres were not discernable [37]. Also, we used only two representative slides to assess the immune response of the entire cysts, which, nevertheless, sufficed to show differences in inflammation and cyst damage with treatment and over time. Despite these limitations, the changes observed after treatment corroborate the increase of inflammation seen in post mortem histological studies in pigs treated with antiparasitic drugs compared to untreated animals [14, 15] and were also confirmed by pre- and post-treatment MRI observations of gadolinium enhancement made in the same pigs. Finally, our study found that combined albendazole plus praziquantel treatment produces a rapid and pronounced reduction of the cyst size in the initial days of the treatment and an acute inflammatory response characterized by an increase of Gd enhancement. This may lead to a release of cyst contains by the extreme cyst damage and a subsequent reduction of cyst size. These results define the pathophysiology of the early exacerbation of symptoms induced by treatment of human NCC, which may lead to earlier monitoring of NCC treatment and thus improved and safer interventions.
|
10.1371/journal.ppat.1006738 | RIG-I-like receptor activation by dengue virus drives follicular T helper cell formation and antibody production | Follicular T helper cells (TFH) are fundamental in orchestrating effective antibody-mediated responses critical for immunity against viral infections and effective vaccines. However, it is unclear how virus infection leads to TFH induction. We here show that dengue virus (DENV) infection of human dendritic cells (DCs) drives TFH formation via crosstalk of RIG-I-like receptor (RLR) RIG-I and MDA5 with type I Interferon (IFN) signaling. DENV infection leads to RLR-dependent IKKε activation, which phosphorylates IFNα/β receptor-induced STAT1 to drive IL-27 production via the transcriptional complex ISGF3. Inhibiting RLR activation as well as neutralizing antibodies against IL-27 prevented TFH formation. DENV-induced CXCR5+PD-1+Bcl-6+ TFH cells secreted IL-21 and activated B cells to produce IgM and IgG. Notably, RLR activation by synthetic ligands also induced IL-27 secretion and TFH polarization. These results identify an innate mechanism by which antibodies develop during viral disease and identify RLR ligands as potent adjuvants for TFH-promoting vaccination strategies.
| Strong antibody production is critical for effective immune responses against viral infections and is a primary factor in the development of successful vaccines. However, it is unclear how virus infection leads to effective antibody responses. Dengue virus (DENV) is known to induce potent antibody production, although the underlying mechanism is poorly understood. Dendritic cells (DCs) are professional sentinels of the immune system and crucial for induction of immune responses. Here we show that DENV infection of human DCs leads to robust antibody production by inducing a specific T helper cell type (also called follicular T helper or TFH) that specializes in stimulating antibody production by B cells. Our data show that DENV replication triggers a viral detection system consisting of sensors RIG-I and MDA5, which specifically induce factors such as IL-27 that are essential for TFH induction. Our data demonstrate that this viral detection system is especially powerful to induce antibody production. Indeed, synthetic molecules that trigger this viral detection mechanism induced superior antibody production compared to other activation signals. Thus, we have identified a viral detection mechanism that leads to strong antibody production and its importance in DENV infection as well as its potential in vaccinations.
| Dengue virus is a global mosquito-transmitted pathogen that infects 400 mln people annually [1]. The majority of patients experience only mild fever, but the disease can progress to life-threatening dengue shock syndrome and dengue hemorrhagic fever. With no specific antivirals or effective vaccine available there is urgent need to advance our understanding of the human immune response against DENV to improve vaccine development and identify molecular targets for drug development.
Antibodies are critical for the host immune response to control, eradicate and prevent (future) viral infections. Antibodies play a dual role in DENV pathology as neutralizing high-affinity antibodies are protective while cross-reacting antibodies can possibly enhance disease of heterologous DENV strains via antibody-dependent enhancement of infection [2]. However, it is unclear how antibodies are induced upon DENV infection. High-affinity antibodies are formed by B cells in germinal centers (GC) during somatic hypermutation [3]. TFH cells are critical for the formation and maintenance of GCs, and induce B cell proliferation and Ig isotype class switching by producing IL-21. TFH cells selectively stimulate high-affinity B cell entry into GCs to promote effective antibody-mediated responses [3–5]. Formation of TFH cells is driven by transcription factor Bcl-6, which induces IL-21 production and expression of chemokine receptor CXCR5, which is pivotal for TFH migration into GCs [6]. DCs are essential for TFH differentiation from naïve CD4+ T cells and previously we have shown that in humans IL-27 is pivotal for TFH formation while IL-6 enhances TFH formation in response to fucosylated parasitic/bacterial ligands [7].
DCs are equipped with numerous sensors including Toll-like receptors (TLRs) and RLRs that sense viral particles or viral replication products to induce innate signaling and drive TH polarization for tailored adaptive immune responses. DENV can both activate TLRs as well as RLRs, depending on the cell type, leading to cytokine secretion and type I IFN production [8–10]. TLR3 resides in endosomes while RLRs are cytoplasmic receptors. Both their activation by viral RNA leads to IFN-β induction via Tank binding protein 1 (TBK1) and transcription factor IRF3. However, RLR signaling also involves IkB kinase (IKK)-related kinase IKKε, which functions in concert with TBK1 to activate IRF3 [11,12]. In parallel with IRF3-dependent IFN-β transcription, TLR3 and RLRs activate NFκB signaling to induce cytokine expression and the combined effect of type I IFN and cytokines determines which differentiation program is initiated in naïve CD4+ T cells. However, it remains unclear how viral sensing in DCs leads to TFH development and subsequent B cell activation.
Here we show that DENV-infected DCs instruct naïve CD4+ T cells to differentiate into Bcl-6+CXCR5+PD-1+ IL-21-secreting TFH cells. DENV RNA replication in both monocyte-derived and primary skin DCs triggers RLR RIG-I and MDA5 leading to IFN-β transcription and IFN-α/βR activation. Notably, RLR-induced IKKε activation modulates IFNα/βR signaling by phosphorylating STAT1. This results in the formation of transcriptional complex ISGF3 instead of STAT1 homodimers, which is pivotal for IL-27 production by DCs and TFH formation. Inhibiting RLR signaling by silencing adapter protein MAVS abrogates IL-27 production and TFH polarization by DENV-infected DCs. In addition, direct RLR activation by synthetic ligands is sufficient to induce IL-27 transcription and TFH formation.
DCs drive TH differentiation and therefore we examined DC-induced immune responses upon DENV infection. DENV efficiently infected human DCs and induced DC maturation as indicated by increased surface expression of CD83 and CD86 (S1 Fig). To investigate TFH differentiation, DENV-infected DCs were co-cultured with naïve CD4+ T cells and TFH induction was determined by measuring expression of CXCR5 and PD-1, which are both expressed by lymph node TFH cells in vivo [13,14]. Strikingly, DENV-infection of DCs induced a robust CXCR5+PD-1+ subset of differentiated TH cells (Fig 1A and 1C), which expressed high levels of TFH-specific transcription factor Bcl-6 (Fig 1B and 1D). T cell differentiation induced by DENV-infected DCs also resulted in strong secretion of IL-21, which is the main effector cytokine of TFH cells (Fig 1E). To investigate whether DENV-induced TFH cells have the capacity to activate B cells, we co-cultured DENV-differentiated TH cells with CD19+ B cells and measured antibody production. Remarkably, differentiated TH cells from DENV-infected, but not mock-treated DCs, induced secretion of both IgM and IgG by B cells (Fig 1F). Blocking DENV RNA replication and infection of DCs (S2 Fig) with DENV RNA replication inhibitor SDM25N [15] abolished the formation of IL-21-secreting CXCR5+PD-1+Bcl-6+ TFH cells (Fig 1A and 1C–1E). These data strongly indicate that DENV replication in DCs induces a TH differentiation program leading to TFH induction and B cell activation.
We set out to identify the molecular mechanism in DCs essential for TFH differentiation. We investigated induction of type I IFN responses upon DENV infection. DENV induced IFN-β transcription in DCs at 18 hours post infection (h.p.i.), which increased over time and correlated with DENV RNA replication (Fig 2A). Next we measured induction of antiviral IFN stimulated genes (ISGs), which are induced by IFNα/βR signaling and indicative of functional type I IFN responses [16]. DENV-infection of DCs induced expression of ISGs MxA, APOBEC3G, ADAR1 and TRIM5α 24 h.p.i. (Fig 2B). These responses depended on DENV RNA replication as the inhibitor SDM25N abrogated the induction of IFN-β and ISGs (Fig 2C). IRF7 is crucial for the induction of IFN-α, which is required for enhancing type I IFN responses [17]. Interestingly, DENV infection induced IRF7 and IFN-α expression after IFN-β induction (Fig 2D and 2E). Both IRF7 and IFN-α expression depended on IFNα/βR signaling, while IFN-β expression was not decreased by blocking IFNα/βR antibodies at early time points (Fig 2D and 2E). IFN-β increased at 32 h.p.i. upon blocking IFNα/βR, probably because of increased DENV replication due to absence of antiviral ISGs (Fig 2E). Notably, silencing IRF7 using RNA interference abrogated IFN-α expression (Fig 2F). Thus, DENV RNA replication induces functional type I IFN responses that are initiated by IFN-β. The induction of IFNα/βR-dependent IRF7 expression subsequently drives IFN-α expression to increase and prolong type I IFN responses against DENV.
Both TLRs and RLRs have been implicated in DENV sensing [8–10]. To elucidate the PRR involved in DENV-induced IFN responses in DCs, we silenced adapter molecules TRIF/MYD88 and MAVS, which are essential for TLR and RLR signaling, respectively (S3 Fig). Silencing MAVS strongly decreased IFN-β and ISG expression in DENV-infected DCs, in contrast, neither silencing of TRIF nor MYD88 affected type I IFN responses (Fig 3A, S2 Fig) even though their silencing abrogated responses to known TLR ligands (S4 Fig). Notably, silencing RIG-I and MDA5 alone or together decreased DENV-induced IFN-β as well as ISG expression (Fig 3B, S2 Fig). RLR triggering leads to activation and phosphorylation of IKKε and TBK1, which target transcription factor IRF3 for nuclear translocation to drive IFN-β expression. DENV infection induced phosphorylation of TBK1 and IKKε, which was dependent on DENV RNA replication (Fig 3C, 3D and 3E). Silencing of TBK1 and IKKε or treatment with the TBK1/IKKε inhibitor BX795 strongly decreased DENV-induced IFN-β expression and supports an important role for these kinases in the type I IFN response against DENV (Fig 3F and 3G). Moreover, DENV infection resulted in nuclear translocation of IRF3 (Fig 3H, 3I and 3J) and silencing IRF3 abrogated IFN-β expression (Fig 3K). These data show that RIG-I and MDA5 sense DENV RNA replication and induce type I IFN responses via TBK1 and IKKε-mediated IRF3 signaling.
Parallel to type I IFN responses, RIG-I and MDA5 induce type III IFN responses which have been implicated in DC migration, viral suppression and modulation of T and B cell responses [18–21]. Therefore, we investigated type III IFN responses by analyzing the expression of IFN-λ genes (IFNL1-4). DENV induced the expression of IFNL1 and IFNL2 and these responses were not affected by type I IFN as blocking IFNα/βR signaling did not affect IFNL1 or IFNL2 expression (Fig 4A). Blocking IFNLR did also not impact DENV-induced IFN-α or IFN-β responses indicating that type I and type III responses operate independently (Fig 4B). IFN-λ is known to suppress viral replication and therefore we examined the effect of blocking IFNLR antibodies on DENV RNA replication. Blocking IFNLR increased DENV RNA replication but not to a similar extent as blocking IFNα/βR (Fig 4C). These data suggest that although type I and type III IFN suppress DENV replication, type I IFN is more effective than type III IFN. Next, we investigated how IFN-λ is induced by DENV. Interestingly, the replication inhibitor SDM25N abrogated DENV-induced IFNL1 and IFNL2 expression (Fig 4D). Moreover, IFNL1 and IFNL2 were induced by MAVS, RIG-I and MDA5 as silencing MAVS or RIG-I and MDA5 together strongly decreased both IFNL1 and IFNL2 expression by DENV infection (Fig 4E). These data strongly suggest that RIG-I and MDA5 triggering by DENV replication induces type III IFN responses that operate independently of type I IFN to suppress viral replication.
Next, we investigated induction of cytokines involved in TFH differentiation. Mice lacking IL-27R have impaired TFH formation [22] and we have recently shown that IL-27 is crucial for TFH polarization by human DCs in response to fucosylated parasitic/bacterial ligands [23]. In addition, Activin A and IL-12 are known to be important factors to drive human TFH formation [24,25]. Therefore, we examined whether DENV infection leads to Activin A, IL-12 or IL-27 expression. IL-12 is heterodimeric protein consisting of subunit p35 and p40. Although DENV induced low levels of IL-12p35, we were unable to detect IL-12p40 expression and this resulted in a lack of IL-12p70 protein (S5A and S5B Fig). We were also unable to detect increased Activin A expression in DENV-infected cells, while LPS strongly induced both IL-12p70 and Activin A (S5A and S5B Fig). Direct stimulation of RIG-I and MDA5 with RLR ligand poly(I:C)Lyovec also did not induce IL-12p70 production or Activin A expression (S5A and S5B Fig), indicating that these factors are more associated with TLR-induced TFH formation and that RLR-mediated TFH differentiation depends on other factors. Interestingly, DENV infection of DCs induced IL-27 production, which decreased by silencing either MAVS or both RIG-I and MDA5 (Fig 5A). IL-27 is a heterodimeric cytokine consisting of subunit p28 and Epstein-Barr virus-induced gene 3 (EBI3) [26]. DENV infection induced expression of both IL-27p28 and EBI3, which was dependent on RLR signaling as well as viral replication (Fig 5B; S5C Fig). Furthermore, blocking IFNα/βR antibodies abrogated IL-27p28 expression without affecting EBI3 expression (Fig 5C), supporting differential regulation of IL-27p28 and EBI3 [27,28] and an important role for crosstalk between RLR and IFNα/βR signaling. Interestingly, IL-27p28 expression was not affected by blocking IFNLR antibodies indicating that specific IFN signaling is necessary for IL-27p28 expression (Fig 5D).
Il-27p28 contains an IFN-stimulated response element (ISRE), which is induced by IFN-stimulating gene factor 3 (ISGF3), a complex of STAT1, STAT2 and IRF9 [27]. Differential signaling by IFNα/βR triggering leads to induction of either STAT1 homodimers or ISGF3 induction [29]. As ISGF3 formation is controlled by IKKε-dependent phosphorylation of STAT1 at Ser708, which prevents STAT1 homodimer formation [30] and IKKε also enhances the transactivation capacity of ISGF3 by phosphorylating STAT1 at Ser727 [23], we investigated whether ISGF3 was involved in IL-27 induction upon DENV infection. We observed STAT1 phosphorylation at Ser708 and Ser727 which was abrogated after treatment with TBK1/IKKε inhibitor BX795 (Fig 5E, 5F and 5G). We also observed nuclear translocation of IRF9 in DENV infected DCs (Fig 5H, 5I and 5J). Moreover, both IKKε and IRF9 were crucial for IL-27p28 expression since silencing of IKKε or IRF9 strongly decreased IL-27p28 expression (Fig 5K). These data show that RLR activation by DENV leads to IFN-β production and subsequent IFNα/βR triggering, and that IKKε activation modulates IFNα/βR signaling to drive IL-27 production.
We next set out to investigate the importance of RLR sensing in TFH formation by DENV-infected DCs. Silencing of MAVS abrogated the formation of CXCR5+PD-1+ TFH cells by DENV-infected DCs (Fig 6A). Moreover, silencing of MAVS diminished Bcl-6 expression in T cells polarized by DENV-infected DCs (Fig 6B). These data strongly suggest that MAVS activation by DENV is pivotal for the formation of TFH cells and identifies an important role for RLR activation in TFH cells induction. To examine this, we transfected DCs with poly(I:C) or 5’pppRNA to activate RIG-I and MDA5 or RIG-I alone, respectively, and investigated TFH differentiation. Strikingly, stimulation of DCs with either poly(I:C)Lyovec or 5’pppRNA-Lyovec induced CXCR5+PD-1+ TFH formation and increased Bcl-6 expression (Fig 6E and 6F). Notably, TH cells polarized by poly(I:C) or 5’pppRNA transfected DCs activated B cells to produce IgM and IgG (Fig 6G). In contrast, LPS-stimulated DCs did neither induce TFH formation nor B cell activation (Fig 6E–6G). We next investigated the importance of IL-27 secretion by DCs for TFH formation. Interestingly, both poly(I:C) and 5’pppRNA transfected DCs produced IL-27 (Fig 6C and 6D) and neutralizing antibodies against IL-27 abrogated CXCR5+PD-1+ TFH formation (Fig 6F). Neutralizing IL-27 also diminished the capacity of TH cells polarized by poly(I:C) or 5’pppRNA transfected DCs to activate B cells to produce IgM and IgG (Fig 6G). These data strongly indicate that RLR activation, either by DENV or synthetic ligands, drives IL-27-dependent TFH polarization.
DENV infection is initiated after a mosquito bite in the skin and swift immune responses against DENV depends on activation of skin DCs [31]. Human skin harbors several DC subsets including CD14+ and CD1c+ dermal DCs of which CD14+ dermal DCs are specialized in the induction of TFH cells [32]. Therefore, we isolated CD14+ and CD1c+ dermal DCs from human skin and investigated if dermal DCs mount immune responses against DENV. Interestingly, DENV induced type I IFN responses in CD1c+ as well as CD14+ dermal DCs, although the induction of type I IFN was more robust in CD14+ dermal DCs than in CD1c+ dermal DCs (Fig 7A). Notably, CD14+ dermal DCs specifically expressed IL-27p28 in response to DENV infection (Fig 7A). We next set out to investigate the importance of DENV replication in IL-27p28 expression by CD14+ dermal DCs. Remarkably, DENV RNA replication inhibitor SDM25N abrogated the induction of IL-27p28 as well as IL-27EBI3 in DENV infected CD14+ dermal DCs (Fig 7B). These data indicate that human dermal DCs mount type I IFN responses against DENV and that DENV replication is essential to trigger IL-27 expression in CD14+ dermal DCs.
TFH cells play a key role in antibody-mediated responses during viral infections or vaccination. Although numerous studies have demonstrated the importance of TFH cells in GC reactions [3–5], little is known about the factors that drive TFH differentiation from naïve CD4+ T cells upon viral infection. Here we show that DENV induces RLR-crosstalk with IFNα/βR signaling leading to IL-27 secretion, which is pivotal for the formation of IL-21-producing CXCR5+PD-1+Bcl-6+ TFH. DENV replication triggered RLR signaling leading to IKKε activation, which is essential for RLR-IFNα/βR crosstalk by phosphorylating STAT1 and inducing ISGF3 formation. Notably, direct activation of RLRs in DCs using poly(I:C)Lyovec or 5’pppRNA-Lyovec also induced IL-27 secretion, TFH polarization and IgM and IgG production by B cells. These data strongly suggest that RLRs are efficient in the induction of TFH responses via their crosstalk with IFNα/βR signaling, and links viral recognition to induction of robust antibody responses.
Viral RNA is a potent pattern-associated molecular pattern that can activate numerous receptors and induce strong immune responses; both TLRs and RLRs have been implicated in DENV sensing [8–10]. DENV particles contain positive single-stranded RNA that can be directly sensed by TLR7 as shown in macrophages [8]. However, we did not find a role for TLR7 in recognizing DENV by DCs as MYD88 silencing did not affect type I IFN responses, probably because TLR7 signaling requires IRF7, which is constituently expressed in plasmacytoid DCs but minimally in other DC subsets [17,33]. Indeed, our data show that IRF7 is minimally expressed by DCS and that this transcription factor is induced by DENV infection of DCs via IFN-β. DENV replication leads to double stranded RNA intermediates, which can be sensed by TLR3 and RLRs. Studies in multiple cell lines have shown that TLR3 can be involved in IFN-β production in response to DENV infection although it is unclear how cytoplasmic RNA is transferred to endosomal TLR3 [9,34]. Our data strongly suggest that RIG-I and MDA-5 but not TLR3 are involved in sensing of DENV and subsequent induction of type I and type III IFN responses. Although DENV is known to block type I IFN responses by inhibiting RLR-MAVS interaction, TBK1 and IRF3 phosphorylation and IFNα/βR signaling via STAT2 degradation [35–38], our data suggest that the inhibition does not effectively occur in DCs. Both type I IFN and type III IFN suppressed viral replication of DENV and preventing antiviral ISG induction by IFNα/βR or IFNLR increased DENV RNA replication. Notably, the increase in DENV RNA replication resulted in an increase in IFN-β suggesting that DENV RNA and IFN-β levels are correlated. It is likely that sensing of RNA products precedes activity of de novo produced viral proteins that block RLR signaling.
Recently, it was shown that Measles virus (MV) directly affects RLR activation. RIG-I and MDA5 activation is tightly regulated and requires dephosphorylation by PP1 phosphates for activation [39]. MV replication is sensed by RLRs and leads to type I IFN responses. However, MV triggers C-type lectin receptor DC-SIGN signaling leading to kinase Raf-1 activation [23]. Raf-1 subsequently induces association of inhibitor protein I-1 with PP1 to lower RIG-I and MDA5 dephosphorylation and type I IFN induction [40]. Although DENV also binds to DC-SIGN [41,42], we did not observe inhibition of RLR activation upon infection. In contrast to MV infection, which activates RLRs very rapidly [43], DENV infection is only sensed after 18 hours when innate signaling by DC-SIGN is probably not effective anymore. Therefore, RIG-I and MDA5 activation is not only important for TFH polarization and antibody production but also to limit viral replication in DCs and possible viral transmission to other cells. Indeed, our data show that RLR-dependent induction of type I IFN and type III IFN suppressed DENV.
We have recently shown that IL-27 is important in the differentiation of TFH and IL-27 induction is dependent on the formation of ISGF3 [7]. Although several TLRs can induce IL-27 transcription, the levels are not sufficient to induce TFH polarization and requires IKKε activation by other receptors for STAT1 phosphorylation and ISGF3 formation [7]. RLRs are therefore unique in their ability to activate both IKKε and IFNα/βR signaling for efficient IL-27 transcription and TFH polarization. These underlying mechanisms might also apply to other viruses as Measles virus, Influenza virus, Rubella virus, HIV-1 and Hepatitis C virus activate RLRs during infection [44–47]. Indeed, our data strongly suggest the use of RLR-ligands as adjuvants for human vaccination strategies, which has been shown to be successful in animal models [48–50].
In addition to IL-27, Activin A and IL-12 have been identified as important cytokines to drive human TFH differentiation [24]. IL-12 expression is strongly induced by TLR signaling, while it is inhibited by RIG-I mediated IRF3 activation [51]. Our data also show that RLR triggering by synthetic ligands or DENV does not lead to IL-12p70 production, in contrast to strong IL-12p70 production by TLR4 activation. We obtained similar results for Activin A, suggesting that Activin A and IL-12p70 could be important for TLR-mediated TFH formation while IL-27 is crucial for RLR-mediated TFH formation.
In the natural course of infection, skin DCs are the first immune cells to encounter DENV after a blood meal of an infected mosquito [31,52]. Effective control of viral propagation from the site of infection requires robust type I IFN responses to suppress viral replication. Our data show that both CD14+ and CD1c+ dermal DCs mount type I IFN responses against DENV. Interestingly, DENV specifically induced IL-27 expression in CD14+ dermal DCs, which are known to be effective inducers of TFH differentiation [32]. Our data show that the induction of IL-27 by DENV in CD14+ dermal DCs critically depends on DENV RNA replication and thereby supports a key function of cytoplasmic sensors of DENV RNA replication in the induction of TFH responses by primary human skin DCs in response to DENV infection.
Our data shows that IFN-α/β induced IL-27 expression is pivotal for TFH formation by DCs while direct stimulation of naïve CD4+ cells with type I IFN-α/β is thought to inhibit TFH formation [24,53]. These studies indicate that IFN signaling, depending on the cell-type and time, can have different effects on TFH differentiation. In addition, direct IFN-α/β stimulation does not lead to IL-27 transcription without IKKε activation to modulate IFNα/βR signaling.
Developing effective DENV vaccines has been hampered by the formation of non-neutralizing antibodies that have the potential to enhance disease [2,54]. A subunit vaccine based on the neutralizing epitope of DENV envelop protein could circumvent the formation of non-neutralizing antibodies [55,56]. However, subunit vaccines usually have low immunogenicity and induce only weak antibody responses. We propose that a subunit vaccine containing the neutralizing DENV epitope in combination with RLR-based adjuvants is a potent strategy to induce high levels of DENV neutralizing antibodies.
In conclusion, we have identified an innate mechanism in DCs that drives TFH polarization during viral infection. Adjuvants targeting this innate mechanism have the potential to improve vaccination strategies for DENV and other pathogens.
This study was done in accordance with the ethical guidelines of the Academic Medical Center and human material was obtained in accordance with the AMC Medical Ethics Review Committee (i.e. Institutional Review Committee) according to the Medical Research Involving Human Subjects Act. Buffy coats obtained after blood donation (Sanquin) or skin tissue are not subjected to informed consent according to the Medical Research Involving Human Subjects Act and the AMC Medical Ethics Review Committee. All samples were handled anonymously.
Peripheral blood monocytes were isolated from buffy coats of healthy donors (Sanquin) by Lymphoprep (Axis-Shield) gradient followed by Percoll (Amersham Biosciences) gradient steps. Monocytes were differentiated into immature DCs in the presence of 500 U/ml IL-4 and 800 U/ml GM-SCF (both Invitrogen) for 6–7 days in RPMI supplemented with 10% fetal calf serum, 10 U/ml penicillin, 10 mg/ml streptomycin (all Invitrogen) and 2 mM L-glutamine (Lonza). This study was done in accordance with the ethical guidelines of the Academic Medical Center.
Dermal DCs (DDCs) were isolated from human skin tissue obtained from healthy donors after corrective breast or abdominal surgery. A dermatome (Zimmer) was used to produce 0.3 mm skin grafts that were treated with dispase (1U/ml, Roche) for 45 min at 37°C to separate dermis and epidermis. Dermal tissue was floated on medium for 16h. Migrated cells were collected and separated based on CD14 (130-050-201, Miltenyi) and CD1c (130-090-506, Miltenyi) expression using magnetic beads according to the manufactures instruction. Isolated cells were analyzed for HLA-DR-PE/Cy7 (1:200, 560651, BD), CD11c-Alexa647 (1:100, 2108100, BioLegend), CD14-PerCP (1:10, 345786 BD) and CD1c-APC/Cy7 (1:50, 331519, BioLegend) expression on a BD Canto II (S6 Fig). CD14+ DDCs were characterized as HLA-DR+CD11c+CD14+CD1c+ and CD1c+ DDCs as HLA-DR+CD11c+CD1c+CD14-. Purity of sorted cells was over 95%.
DCs were stimulated with 1 μg/ml poly(I:C)LyoVec LMW or 10 μg/ml 5’ppp-dsRNA-LyoVec (both Invivogen) unless stated otherwise. DENV replication inhibitor SDM25N (10μM, Tocris Bioscience),blocking IFNα/βR antibody or blocking IFNLR antibody (20 μg/ml, both PBL Interferon Source) were added simultaneous with DENV to DCs.
DCs were transfected with 500 nM short interfering RNAs (siRNAs) using the Neon Transfection System (ThermoFisher) according to the manufacturer’s instructions. In brief, DCs were washed with PBS, resuspended in Buffer R (ThermoFisher) and divided over different siRNAs. DCs were transfected with a single pulse of 1500V for 20 ms, mixed with complete RMPI and incubated for 48h before stimulation. SMARTpool siRNA used were MAVS (M-024237-02), TRIF (M-012833-02), MYD88 (M-004769-01), RIG-I (M-012511-01), MDA5 (M-013041-00), TBK1 (M-003788-02), IKKε (M-003723-02), IRF9 (M-020858-02), IRF3 (M-006875-02), and non-targeting siRNA (D-001206-13) as control (all Dharmacon). Silencing was confirmed by real-time PCR, flow cytometry and immunoblot (S3 Fig).
Naive CD4+ T cells were isolated from buffy coats of healthy blood donors (Sanquin) with human CD4+ T-cell isolation kit II (Miltenyi) by negative selection and subsequent depletion of CD45RO+ memory T cells using phycoerythrin (PE)-conjugated anti-CD45RO (80μg ml-1; R0843; Dako) and anti-PE beads (Miltenyi). B cells were isolated from buffy coats of healthy blood donors (Sanquin) with human B-cell isolation kit II (Miltenyi) by negative selection. This study was approved by the Medical Ethics Review Committee of the AMC.
DCs were either silenced for indicated proteins or treated with SDM25N and stimulated for 48h as indicated. DCs were combined with allogeneic naïve CD4+ T cells (5,000 DCs/20,000 T cells) in the presence of 10 pg/ml Staphylococcus aureus enterotoxin B (Sigma). SDM25N (1 μM, Tocris Bioscience) was added to cocultures of SDM25N-treated DCs to maintain inhibition of DENV replication. Neutralizing antibodies against IL-27 (5 μg/ml, AF2526; R&D Systems) or normal goat IgG (AB-108-C; R&D Systems) as isotype control was added at the start of DC-T cell coculture. After 3 days, cells were further cultured in the presence of 10 U/ml IL-2 (Chiron). Resting T cells were restimulated with 100 ng/ml PMA and 1 μg/ml ionomycin (both Sigma) for 24h. For flow cytometry analysis of restimulated T cells, cells were stained with Alexa Fluor 647-conjugated anti-CXCR5 (1:800; 558113; BD Pharmingen) and PerCP-Cy5.5-conjugated α-PD-1 (1:50; 561273; BD) before fixation in 2% para-formaldehyde for 20 min, followed by permeabilization in 50% methanol at -20°C for 45 min. Cells were stained with anti-Bcl-6 (1:50; ab19011; Abcam), followed by incubation with PE-conjugated anti-rabbit (1:200; 711-116-152, Jackson ImmunoResearch). Cells were analyzed on a FACS Canto II (BD Biosciences). Supernatants of restimulated T cells were harvested after 24h and IL-21 expression was analyzed by ELISA (eBioscience).
T-cell dependent B-cell activation was assessed by coculturing resting differentiated T cells restimulated with 1 μg/ml anti-CD3 (1XE, Sanquin) and 2 μg/ml anti-CD28 (15E8, Sanquin) with allogeneic B cells (100,000 T cells/50,000 B cells). Supernatants were harvested after 7 days for analysis of IgM and IgG production by ELISA (eBioscience).
DENV-2/16681 was added to 80% confluent C6/36 cells at an MOI of 0.01 in RPMI medium RPMI supplemented with 2% fetal calf serum, 10 U/ml penicillin, 10 mg/ml streptomycin (all Invitrogen) and 2 mM L-glutamine (Lonza). After 5–7 days, supernatant was harvested and cleared from cellular debris by centrifugation and subsequent filtration using a 0.2 μM filter. Supernatant was aliquoted, snap-frozen in liquid nitrogen and stored at -80°C. Viral titers were determined as described previously[57].
DCs were infected with DENV at an MOI of 1 unless stated otherwise. Infection was determined after 36-48h by flow cytometry. Cells were fixed in 4% para-formaldehyde for 15 min followed by permeabilization in PBS supplemented with 0.1% saponin for 10 min. Cells were stained with anti-NS3 (1:800, SAB2700181, Sigma) followed by PE-conjugated anti-rabbit (1:200; 711-116-152, Jackson ImmunoResearch) in combination with APC-conjugated CD83 (1:25, 551073, BD Pharmingen) and FITC-conjugated CD86 (1:25, 555657, BD Pharmingen). Cells were analyzed on a FACS Canto II (BD Biosciences).
TBK1, IKKε and STAT1 phosphorylation was determined by flow cytometry and immunoblot. For flow cytometry, cells were fixed in 4% para-formaldehyde for 15 min followed by permeabilization in 90% methanol at -20°C for 45 min. Cells were stained with phospho-specific antibodies against TBK1 Ser172 (1:50, 5483S, Cell Signaling), IKKε Ser172 (1:50, 06–1340, Millipore), STAT1 Ser708 (1:100, provided by M. Gale, Jr, University of Washington School of Medicine, Seattle, WA, (Perwitasari et al., 2011) or STAT1 Ser727 (1:200, 9177; Cell Signaling), followed by PE-conjugated anti-rabbit (1:200; 711-116-152, Jackson ImmunoResearch). Cells were analyzed on a FACS Calibur (BD Biosciences).
For immunoblot, whole cell extracts were prepared using Ripa lysis buffer (Cell Signaling Technology) and protein were resolved by SDS-PAGE and detected with anti-TBK1 Ser172 (1:1000, 5483S, Cell Signaling) anti-IKKε Ser172 (1:1000, 06–1340, Millipore), anti-STAT1 Ser708 (1:1000, provided by M. Gale, Jr, University of Washington School of Medicine, Seattle, WA, (Perwitasari et al., 2011) or anti-STAT1 Ser727 (1:1000, 9177; Cell Signaling). Membranes were also probed with anti-TBK1 (1:1000, 3504, Cell Signaling), anti-IKKε (1:1000, 2905, Cell Signaling) or anti-STAT1 (1:1000, 9172, Cell Signaling) as loading control. Primary antibody was detected using HRP-conjugated secondary antibody (1:2000, 21230, Pierce)
mRNA was isolated using mRNA capture kit (Roche) and cDNA was synthesized with reverse transcriptase kit (Promega). PCR amplification was performed in the presence of SYBR Green in an ABI 7500 Fast PCR detection system (Applied Biosystems). Specific primers were designed using Primer Express 2.0 (Applied Biosystems; S1 Table). Expression of target genes was normalized to GAPDH (Nt = 2Ct(GAPDH)–Ct(target)) and set at 1 in DENV-infected DCs for each donor within one experiment.
Nuclear translocation of IRF3 and IRF9 was determined by confocal microscopy, immunoblot and ELISA. For Confocal microscopy, cells were allowed to adhere to poly-l-lysine coated glass slides for 20 min at 37°C before fixation in 2% para-formaldehyde for 20 min followed by permeabilization using 0.2% Triton for 10 min. Cells were stained with anti-IRF3 (1:100, D83B9, Cell Signaling) or anti-IRF9 (5 μg/ml, sc-496X, Santa Cruz) and anti-DENV E protein (1:400, 3H5-1, Millipore) followed by Alexa Fluor 488-conjugated anti-mouse (1:400, A11029, Invitrogen) and Alexa Fluor 546-conjugated anti-rabbit (1:400, A11035, Invitrogen). Nuclei were stained using Hoechst (1:10,000, Molecular Probes). Cells were analyzed on a Leica TCS SP8 X mounted on a Leica DMI6000 inverted microscope and data was processed using Leica LAS-X software.
For immunoblot, nuclear and cytoplasmic extracts were prepared using NucBuster protein extraction kit (Novagen). Proteins were resolved by SDS-PAGE and detected by immunoblotting with anti-iRF3 (1:1000, 4302; Cell Signaling) or anti-IRF9 (1:1000, sc-496, Santa Cruz). Membranes were also probed with anti-β-actin (1:2500, sc-81178; Santa Cruz) to ensure equal protein loading. Detection was performed as described above. Specific increase of IRF3 and IRF9 in the nuclear fraction without increase in the cytoplasmic fraction underscores specificity of the fractionation. IRF3 and IRF9 levels in nuclear extracts was also determined using ELISA (IRF3, SEB589HU, USCN Life Sciences; IRF9, MBS921012, MyBiosource).
Statistical analyses were performed using the Student’s t-test for paired observations. Statistical significance was set at P<0.05.
|
10.1371/journal.pcbi.1006633 | Deep image reconstruction from human brain activity | The mental contents of perception and imagery are thought to be encoded in hierarchical representations in the brain, but previous attempts to visualize perceptual contents have failed to capitalize on multiple levels of the hierarchy, leaving it challenging to reconstruct internal imagery. Recent work showed that visual cortical activity measured by functional magnetic resonance imaging (fMRI) can be decoded (translated) into the hierarchical features of a pre-trained deep neural network (DNN) for the same input image, providing a way to make use of the information from hierarchical visual features. Here, we present a novel image reconstruction method, in which the pixel values of an image are optimized to make its DNN features similar to those decoded from human brain activity at multiple layers. We found that our method was able to reliably produce reconstructions that resembled the viewed natural images. A natural image prior introduced by a deep generator neural network effectively rendered semantically meaningful details to the reconstructions. Human judgment of the reconstructions supported the effectiveness of combining multiple DNN layers to enhance the visual quality of generated images. While our model was solely trained with natural images, it successfully generalized to artificial shapes, indicating that our model was not simply matching to exemplars. The same analysis applied to mental imagery demonstrated rudimentary reconstructions of the subjective content. Our results suggest that our method can effectively combine hierarchical neural representations to reconstruct perceptual and subjective images, providing a new window into the internal contents of the brain.
| Machine learning-based analysis of human functional magnetic resonance imaging (fMRI) patterns has enabled the visualization of perceptual content. However, prior work visualizing perceptual contents from brain activity has failed to combine visual information of multiple hierarchical levels. Here, we present a method for visual image reconstruction from the brain that can reveal both seen and imagined contents by capitalizing on multiple levels of visual cortical representations. We decoded brain activity into hierarchical visual features of a deep neural network (DNN), and optimized an image to make its DNN features similar to the decoded features. Our method successfully produced perceptually similar images to viewed natural images and artificial images (colored shapes and letters), whereas the decoder was trained only on an independent set of natural images. It also generalized to the reconstruction of mental imagery of remembered images. Our approach allows for studying subjective contents represented in hierarchical neural representations by objectifying them into images.
| While the externalization of states of the mind is a long-standing theme in science fiction, it is only recently that the advent of machine learning-based analysis of functional magnetic resonance imaging (fMRI) data has expanded its potential in the real world. Although sophisticated decoding and encoding models have been developed to render human brain activity into images or movies, the methods are essentially limited to image reconstructions with low-level image bases [1, 2], or to matching to exemplar images or movies [3, 4], failing to combine the visual features of multiple hierarchical levels. While several recent approaches have introduced deep neural networks (DNNs) for the image reconstruction task, they have failed to fully utilize hierarchical information to reconstruct visual images [5, 6]. Furthermore, whereas categorical decoding of imagery contents has been demonstrated [7, 8], the reconstruction of internally generated images has been challenging.
The recent success of DNNs provides technical innovations to study the hierarchical visual processing in computational neuroscience [9]. Our recent study used DNN visual features as a proxy for the hierarchical neural representations of the human visual system and found that a brain activity pattern measured by fMRI could be decoded (translated) into the response patterns of DNN units in multiple layers representing the hierarchical visual features given the same input [10]. This finding revealed a homology between the hierarchical representations of the brain and the DNN, providing a new opportunity to utilize the information from hierarchical visual features.
Here, we present a novel approach, named deep image reconstruction, to visualize perceptual content from human brain activity. This technique combines the DNN feature decoding from fMRI signals with recently developed methods for image generation from the machine learning field (Fig 1) [11]. The reconstruction algorithm starts with a given initial image and iteratively optimizes the pixel values so that the DNN features of the current image become similar to those decoded from brain activity across multiple DNN layers. The resulting optimized image is considered as a reconstruction from the brain activity. We optionally introduced a deep generator network (DGN) [12] to constrain the reconstructed images to look similar to natural images by performing optimization in the input space of the DGN.
We trained the decoders that predicted the DNN features of viewed images from fMRI activity patterns following the procedures of Horikawa & Kamitani (2017) [10]. In the present study, we used the VGG19 DNN model [13], which consisted of sixteen convolutional layers and three fully connected layers and was pre-trained with images in ImageNet [14] to classify images into 1,000 object categories (see Materials and Methods: “Deep neural network features” for details). We constructed one decoder for a single DNN unit to predict outputs of the unit. We trained decoders corresponding to all the units in all the layers (see Materials and Methods: “DNN feature decoding analysis” for details).
The feature decoding analysis was performed with fMRI activity patterns in visual cortex (VC) measured while subjects viewed or imagined visual images. Our experiments consisted of the training sessions in which only natural images were presented and the test sessions in which independent sets of natural images, artificial shapes, and alphabetical letters were presented. In another test session, a mental imagery task was performed. The decoders were trained using the fMRI data from the training sessions, and the trained decoders were then used to predict DNN feature values from the fMRI data of the test sessions (the accuracies are shown in S1 Fig).
Decoded features were then forwarded to the reconstruction algorithm to generate an image using variants of gradient descent optimization (see Material and Methods: “Reconstruction from a single DNN layer” and “Reconstruction from multiple DNN layers” for details). The optimization was performed to minimize the error between multi-layer DNN features decoded from brain activity patterns and those calculated from the input image by iteratively modifying the input image. For natural image reconstructions, to improve the “naturalness” of reconstructed images, we further introduced the constraint using a deep generator network (DGN) derived from the generative adversarial network algorithm (GAN) [15], which is known to capture a latent space explaining natural images [16] (see Material and Methods: “Natural image prior” for details).
Examples of reconstructions for natural images are shown in Fig 2 (see S2 Fig for more examples, and see S1 Movie for reconstructions through the optimization processes). The reconstructions obtained with the DGN capture the dominant structures of the objects within the images. Furthermore, fine structures reflecting semantic aspects like faces, eyes, and texture patterns were also generated in several images. Our extensive analysis on each of the individual subjects demonstrated replicable results across the subjects. Moreover, the same analysis on a previously published dataset [10] also replicated qualitatively similar reconstructions to those in the present study (S3 Fig).
To investigate the effect of the DGN, we evaluated the quality of reconstructions generated both with and without using it (Fig 3A and 3B; see S4 Fig for individual subjects; see Material and Methods: “Evaluation of reconstruction quality”). While the reconstructions obtained without the DGN also successfully reproduced rough silhouettes of dominant objects, they did not show semantically meaningful appearances (see S5 Fig for more examples; also see S6 Fig for reconstructions from different initial states for both with and without the DGN). Evaluations using pixel-wise spatial correlation and human judgment both showed almost comparable accuracy for reconstructions with and without the DGN (accuracy of pixel-wise spatial correlation, with and without the DGN, 76.1% and 79.7%; accuracy of human judgment, with and without the DGN, 97.0% and 96.0%). However, reconstruction accuracy evaluated using pixel-wise spatial correlation showed slightly higher accuracy with reconstructions performed without the DGN than with the DGN (two-sided signed-rank test, P < 0.01), whereas the opposite was observed for evaluations by human judgment (two-sided signed-rank test, P < 0.01). These results suggest the utility of the DGN that enhances the perceptual similarity of reconstructed images to target images by rendering semantically meaningful details in the reconstructions.
To characterize the ‘deep’ nature of our method, the effectiveness of combining multiple DNN layers was tested using both objective and subjective assessments [5, 17, 18]. For each of the 50 test natural images, reconstructed images were generated with a variable number of multiple layers (Fig 4A; DNN1 only, DNN1–2, DNN1–3, …, DNN1–8; see S7 Fig for more examples). In the objective assessment, the pixel-wise spatial correlations to the original image were compared between two combinations of DNN layers. In the subjective assessment, an independent rater was presented with an original image and a pair of reconstructed images, both from the same original image but generated with different combinations of multiple layers, and was required to indicate which of the reconstructed images looked more similar to the original image. While the objective assessment showed higher winning percentages for the earliest layer (DNN1) alone, the subjective assessment showed increasing winning percentages for a larger number of DNN layers (Fig 4B). Our additional analysis showed poor reconstruction quality from individual layers especially from higher layers (see S8 Fig for reconstructions from individual layers). These results suggest that combining multiple levels of visual features enhanced the perceptual reconstruction quality even though the pixel-wise accuracy is lost.
Given the true DNN features, instead of decoded features, as the input, the reconstruction algorithm produces almost complete reconstructions of original images (S8 Fig), indicating that the DNN feature decoding accuracy would determine the quality of reconstructed images. To further confirm this, we calculated the correlation between the feature decoding accuracy and the reconstruction quality for individual images (S9 Fig). The analyses showed positive correlations for both the objective and subjective assessments, suggesting that improving feature decoding accuracy could improve reconstruction quality.
We found that the luminance contrast of a reconstruction was often reversed (e.g., the stained-glass images in Fig 2), presumably because of the lack of (absolute) luminance information in the fMRI signals, even in the early visual areas [19]. Additional analyses revealed that the feature values of filters with high luminance contrast in the earliest DNN layers (conv1_1 in VGG19) were better decoded when they were converted to absolute values (Fig 5A and 5B), demonstrating a clear discrepancy between the fMRI and raw DNN signals. The large improvement levels demonstrate the insensitivity of fMRI signals to pixel luminance, suggesting the linear-nonlinear discrepancy of DNN and fMRI responses to pixel luminance. This discrepancy may explain the reversal of luminance observed in several reconstructed images. While this may limit the potential for reconstructions from fMRI signals, the ambiguity might be resolved by modelling DNNs to fill the gaps between signals of DNNs and fMRI. Alternatively, further emphasis of the high-level visual information in hierarchical visual features may help to resolve the ambiguity of luminance by incorporating information on semantic context.
To confirm that our method was not restricted to the specific image domain used for the model training, we tested whether it was possible to generalize the reconstruction to artificial images. This was challenging, because both the DNN and our decoding models were solely trained on natural images. The reconstructions of artificial shapes and alphabetical letters are shown in Fig 6A and 6B (also see S10 Fig and S2 Movie for more examples of artificial shapes, and see S11 Fig for more examples of alphabetical letters). The results show that artificial shapes were successfully reconstructed with moderate accuracy (Fig 6C left; 70.5% by pixel-wise spatial correlation, 91.0% by human judgment; see S12 Fig for individual subjects) and alphabetical letters were also reconstructed with high accuracy (Fig 6C right; 95.6% by pixel-wise spatial correlation, 99.6% by human judgment; see S13 Fig for individual subjects). These results indicate that our model did indeed ‘reconstruct’ or ‘generate’ images from brain activity, and that it was not simply making matches to exemplars. Furthermore, the successful reconstructions of alphabetical letters demonstrate that our method can expand the possible states of visualizations, with advance in resolution over reconstructions performed in previous studies [1, 20].
To assess how the shapes and colors of the stimulus images were reconstructed, we separately evaluated the reconstruction quality of each of shape and color by comparing reconstructed images of the same colors and shapes. Analyses with different visual areas showed different trends in reconstruction quality for shapes and colors (Fig 7A and see S14 Fig for more examples). Human judgment evaluations suggested that shapes were reconstructed better from early visual areas, whereas colors were reconstructed better from the mid-level visual area V4 (Fig 7B and see S15 Fig for individual subjects; ANOVA, interaction between task type [shape vs. color] and brain areas [V1 vs. V4], P < 0.01), although the interaction effect was marginal when considering evaluations by pixel-wise spatial correlation (P = 0.06). These contrasting patterns further support the success of shape and color reconstructions and indicate that our method can be a useful tool to characterize the information content encoded in the activity patterns of individual brain areas by visualization.
Finally, to explore the possibility of visually reconstructing subjective content, we performed an experiment in which participants were asked to produce mental imagery of natural and artificial images shown prior to the task session. The reconstructions generated from brain activity due to mental imagery are shown in Fig 8 (see S16 Fig and S3 Movie for more examples). While the reconstruction quality varied across subjects and images, rudimentary reconstructions were obtained for some of the artificial shapes (Fig 8A and 8B for high and low accuracy images, respectively). In contrast, imagined natural images were not well reconstructed, possibly because of the difficulty of imagining complex natural images (Fig 8C; see S17 Fig for vividness scores of imagery). While the pixel-wise spatial correlation evaluations of reconstructed artificial images did not show high accuracy (Fig 8D; 51.9%; see S18 Fig for individual subjects), this may have been due to the possible disagreements in positions, colors and luminance between target and reconstructed images. Meanwhile, the human judgment evaluations showed accuracy higher than the chance level, suggesting that imagined artificial images were recognizable from the reconstructed images (Fig 8D; 83.2%; one-sided signed-rank test, P < 0.01; see S18 Fig for individual subjects). Furthermore, separate evaluations of color and shape reconstructions of artificial images suggested that shape rather than color had a major contribution to the high proportion of correct answers by human raters (Fig 8E; color, 64.8%; shape, 87.0%; two-sided signed-rank test, P < 0.01; see S19 Fig for individual subjects). Additionally, poor but sufficiently recognizable reconstructions were obtained even from brain activity patterns in the primary visual area (V1; 63.8%; three subjects pooled; one-sided signed-rank test, P < 0.01; see S20 Fig for reconstructed images and S21 Fig and S22 Fig for quantitative evaluations), possibly supporting the notion that low-level visual features are encoded in early visual cortical activity during mental imagery [21]. Taken together, these results provide evidence for the feasibility of visualizing imagined content from brain activity patterns.
We have presented a novel approach to reconstruct perceptual and mental content from human brain activity combining visual features from the multiple layers of a DNN. We successfully reconstructed viewed natural images, especially when combined with a DGN. The results from the extensive analysis on each subject were replicated across different subjects. Reconstruction of artificial shapes was also successful, even though the reconstruction models used were trained only on natural images. The same method was also applied to mental imagery, and revealed rudimentary reconstructions of mental content.
Our method is capable of reconstructing various types of images, including natural images, colored artificial shapes, and alphabetical letters, even though each component of our reconstruction model, the DNN models and the DNN feature decoders, was solely trained with natural images. The results strongly demonstrated that our method was certainly able to ‘reconstruct’ or ‘generate’ images from brain activity, differentiating our method from the previous attempts to visualize perceptual contents using the exemplar matching approach, which suffers from restrictions derived from pre-selected image/movie sets [3, 4].
We introduced the GAN-based constraint using the DGN for natural image reconstructions to enhance the naturalness of reconstructed images, rendering semantically meaningful details to the reconstructions. A variant of the GAN-based approach has demonstrated the utility in a previous face image reconstruction study, too [22]. GAN-derived feature space appears to provide efficient constraints on resultant images to enhance the perceptual resemblance to the image set on which a GAN is trained.
While one of the strengths of the present method is its generalizability across image types, there remains room for substantial improvements in reconstruction performance. Because we used the models (DNNs and decoders) trained with natural ‘object’ images from the ImageNet database [14], whose images contain objects around the center, it would not be optimal for the reconstruction of other types of images. Furthermore, because we used the DNN model trained to classify images into 1,000 object categories, the representations acquired in the DNN would be specifically suited to the particular objects. One could train the models with diverse types of images, such as scenes, textures, and artificial shapes, as well as object images, to improve general reconstruction performance. If the target image type is known in prior, one can use a specific set of images and a DNN model training task that are matched to it.
Other DNN models with different architectures could also be used to improve general reconstruction performance. As the reconstruction quality is positively correlated with the feature decoding accuracy (S9 Fig), DNNs with highly decodable units are likely to improve reconstructions. Recent studies evaluated different types of DNNs in term of the prediction accuracy of brain activity given their feature values (or the encoding accuracy) [23–25]. Although it remains to be seen how closely the encoding and decoding accuracies are linked, it is expected that more ‘brain-like’ DNN models would yield high-quality reconstructions.
Our approach provides a unique window into our internal world by translating brain activity into images via hierarchical visual features. Our method can also be extended to decode mental contents other than visual perception and imagery. By choosing an appropriate DNN architecture with substantial homology with neural representations, brain-decoded DNN features could be rendered into movies, sounds, text, or other forms of sensory/mental representations. The externalization of mental contents by this approach might prove useful in communicating our internal world via brain–machine/computer interfaces.
All subjects provided written informed consent for participation in our experiments, in accordance with the Declaration of Helsinki, and the study protocol was approved by the Ethics Committee of ATR.
Three healthy subjects with normal or corrected-to-normal vision participated in our experiments: Subject 1 (male, age 33), Subject 2 (male, age 23) and Subject 3 (female, age 23). This sample size was chosen on the basis of previous fMRI studies with similar experimental designs [1, 10].
Visual stimuli consisted of natural images, artificial shapes, and alphabetical letters. The natural images were identical to those used in Horikawa & Kamitani (2017) [10], which were originally collected from the online image database ImageNet (2011, fall release) [14]. The images were cropped to the center and resized to 500 × 500 pixels. The artificial shapes consisted of a total of 40 combinations of 5 shapes and 8 colors (red, green, blue, cyan, magenta, yellow, white, and black), in which the shapes were identical to those used in Miyawaki et al. (2008) [1] and the luminance was matched across colors except for white and black. The alphabetical letter images consisted of the 10 black letters, A, C, E, I, N, O, R, S, T, and U.
We conducted two types of experiments: image presentation experiments and a mental imagery experiment. The image presentation experiments consisted of four distinct session types, in which different variants of visual images were presented (training natural images, test natural images, artificial shapes, and alphabetical letters). All visual stimuli were rear-projected onto a screen in the fMRI scanner bore using a luminance-calibrated liquid crystal display projector. To minimize head movements during fMRI scanning, subjects were required to fix their heads using a custom-molded bite-bar individually made for each subject. Data from each subject were collected over multiple scanning sessions spanning approximately 10 months. On each experimental day, one consecutive session was conducted for a maximum of 2 hours. Subjects were given adequate time for rest between runs (every 5–8 min) and were allowed to take a break or stop the experiment at any time.
The image presentation experiments consisted of four distinct types of sessions: training natural-image sessions, test natural-image sessions, artificial-shape sessions, and alphabetical-letter sessions. Each session consisted of 24, 24, 20, and 12 separate runs, respectively. For these four sessions, each run comprised 55, 55, 44, and 11 stimulus blocks, respectively, with these consisting of 50, 50, 40, and 10 blocks with different images, and 5, 5, 4, and 1 randomly interspersed repetition blocks where the same image as in the previous block was presented (7 min 58 s for the training and test natural-image sessions, 6 min 30 s for the artificial-shape sessions, and 5 min 2 s for the alphabetical-letter sessions, for each run). Each stimulus block was 8 s (training natural-images, test natural-images, and artificial-shapes) or 12 s (alphabetical-letters) long, and was followed by a 12-s rest period for the alphabetical-letters, while no rest period was used for the training natural-images, test natural-images, and artificial-shapes. Images were presented at the center of the display with a central fixation spot and were flashed at 2 Hz (12 × 12 and 0.3 × 0.3 degrees of visual angle for the visual images and fixation spot respectively). The color of the fixation spot changed from white to red for 0.5 s before each stimulus block began, to indicate the onset of the block. Additional 32- and 6-s rest periods were added to the beginning and end of each run respectively. Subjects were requested to maintain steady fixation throughout each run and performed a one-back repetition detection task on the images, responding with a button press for each repeated image, to ensure they maintained their attention on the presented images (mean task performance across three subjects: sensitivity 0.9820; specificity 0.9995; pooled across sessions). In one set of training natural-image session, a total of 1,200 images were presented only once. This set of training natural-image session was repeated five times (1,200 × 5 = 6,000 samples for training). In the test natural-image, artificial-shape, and alphabetical-letter sessions, 50, 40, and 10 images were presented 24, 20, and 12 times each respectively. The presentation order of the images was randomized across runs.
In the mental imagery experiment, subjects were required to visually imagine (recall) one of 25 images selected from those presented in the test natural image and artificial shape sessions of the image presentation experiment (10 natural images and 15 artificial images). Prior to the experiment, subjects were asked to relate words to visual images, so that they could recall the visual images from word cues. The imagery experiment consisted of 20 separate runs, with each run containing 26 blocks (7 min 34 s for each run). The 26 blocks consisted of 25 imagery trials and a fixation trial, in which subjects were required to maintained a steady fixation without any imagery. Each imagery block consisted of a 4-s cue period, an 8-s mental imagery period, a 3-s evaluation period, and a 1-s rest period. Additional 32- and 6-s rest periods were added to the beginning and end of each run respectively. During the rest periods, a white fixation spot was presented at the center of the display. At 0.8 s before each cue period, the color of the fixation spot changed from white to red for 0.5 s, to indicate the onset of the blocks. During the cue period, words specifying the visual images to be imagined were visually presented around the center of the display (1 target and 25 distractors). The position of each word was randomly changed across blocks to avoid cue-specific effects contaminating the fMRI response during mental imagery periods. The word corresponding to the image to be imagined was presented in red (target) and the other words were presented in black (distractors). Subjects were required to start imagining a target image immediately after the cue words disappeared. The imagery period was followed by a 3-s evaluation period, in which the word corresponding to the target image and a scale bar was presented, to allow the subjects to evaluate the correctness and vividness of their mental imagery on a five-point scale (very vivid, fairly vivid, rather vivid, not vivid, cannot correctly recognize the target). This was performed by pressing the left and right buttons of a button box placed in their right hand, to change the score from its random initial setting. As an aid for remembering the associations between words and images, the subjects were able to use control buttons to view the word and visual image pairs during every inter-run-rest period.
fMRI data were collected using a 3.0-Tesla Siemens MAGNETOM Verio scanner located at the Kokoro Research Center, Kyoto University. An interleaved T2*-weighted gradient-echo echo planar imaging (EPI) scan was performed to acquire functional images covering the entire brain (TR, 2000 ms; TE, 43 ms; flip angle, 80 deg; FOV, 192 × 192 mm; voxel size, 2 × 2 × 2 mm; slice gap, 0 mm; number of slices, 76; multiband factor, 4). High-resolution anatomical images of the same slices obtained for the EPI were acquired using a T2-weighted turbo spin echo sequence (TR, 11000 ms; TE, 59 ms; flip angle, 160 deg; FOV, 192 × 192 mm; voxel size, 0.75 × 0.75 × 2.0 mm). T1-weighted magnetization-prepared rapid acquisition gradient-echo (MP-RAGE) fine-structural images of the entire head were also acquired (TR, 2250 ms; TE, 3.06 ms; TI, 900 ms; flip angle, 9 deg, FOV, 256 × 256 mm; voxel size, 1.0 × 1.0 × 1.0 mm).
The first 8 s of scans from each run were discarded to avoid MRI scanner instability effects. We then used SPM (http://www.fil.ion.ucl.ac.uk/spm) to perform three-dimensional motion correction on the fMRI data. The motion-corrected data were then coregistered to the within-session high-resolution anatomical images with the same slices as the EPI, and then subsequently to the whole-head high-resolution anatomical images. The coregistered data were then re-interpolated to 2 × 2 × 2 mm voxels.
Data samples were created by first regressing out nuisance parameters from each voxel amplitude for each run, including any linear trend and the temporal components proportional to the six motion parameters calculated during the motion correction procedure. After that, voxel amplitudes were normalized relative to the mean amplitude of the initial 24-s rest period of each run and were despiked to reduce extreme values (beyond ± 3 SD for each run). The voxel amplitudes were then averaged within each 8-s (training natural image-sessions) or 12-s (test natural-image, artificial-shape, and alphabetical-letter sessions) stimulus block (four or six volumes), and within the 16-s mental imagery block (eight volumes, mental imagery experiment), after shifting the data by 4 s (two volumes) to compensate for hemodynamic delays.
V1, V2, V3, and V4 were delineated following the standard retinotopy experiment [26, 27]. The lateral occipital complex (LOC), fusiform face area (FFA), and parahippocampal place area (PPA) were identified using conventional functional localizers [28–30] (See S1 Supporting Information for details). A contiguous region covering the LOC, FFA, and PPA was manually delineated on the flattened cortical surfaces, and the region was defined as the higher visual cortex (HVC). Voxels overlapping with V1–V3 were excluded from the HVC. Voxels from V1–V4 and the HVC were combined to define the visual cortex (VC). In the regression analysis, voxels showing the highest correlation coefficient with the target variable in the training image session were selected to decode each feature (with a maximum of 500 voxels).
We used the Caffe implementation of the VGG19 deep neural network (DNN) model [13], which was pre-trained with images in ImageNet [14] to classify 1,000 object categories (the pre-trained model is available from https://github.com/BVLC/caffe/wiki/Model-Zoo). The VGG19 model consisted of a total of sixteen convolutional layers and three fully connected layers. To compute outputs by the VGG19 model, all visual images were resized to 224 × 224 pixels and provided to the model. The outputs from the units in each of the 19 layers (immediately after convolutional or fully connected layers, before rectification) were treated as a vector in the following decoding and reconstruction analysis. The number of units in each of the19 layers is the following: conv1_1 and conv1_2, 3211264; conv2_1 and conv2_2, 1605632; conv3_1, conv3_2, conv3_3, and conv3_4, 802816; conv4_1, conv4_2, conv4_3, and conv4_4, 401408; conv5_1, conv5_2, conv5_3, and conv5_4, 100352; fc6 and fc7, 4096; and fc8, 1000. In this study, we named five groups of convolutional layers as DNN1–5 (DNN1: conv1_1, and conv1_2; DNN2: conv2_1, and conv2_2; DNN3: conv3_1, conv3_2, conv3_3, and conv3_4; DNN4: conv4_1, conv4_2, conv4_3, and conv4_4; and DNN5: conv5_1, conv5_2, conv5_3, and conv5_4), and three fully-connected layers as DNN6–8 (DNN6: fc6; DNN7: fc7; and DNN8: fc8). We used the original pre-trained VGG19 model to compute the feature unit activities, but for analyses with fMRI data from the mental imagery experiment, we changed the DNN model so that the max pooling layers were replaced by average pooling layers, and the ReLU activation function was replaced by a leaky ReLU activation function with a negative slope of 0.2 (see Simonyan & Zisserman (2015) [13] for the details of the original DNN architecture).
We used a set of linear regression models to construct multivoxel decoders to decode the DNN feature vector of a seen image from the fMRI activity patterns obtained in the training natural-image sessions (training dataset). In this study, we used the sparse linear regression algorithm (SLR) [31], which can automatically select important voxels for decoding by introducing sparsity into a weight estimation through Bayesian estimation of parameters with the automatic relevance determination (ARD) prior (see Horikawa & Kamitani (2017) [10] for a detailed description). The training dataset was used to train the decoders to decode the values of individual units in the feature vectors of all DNN layers (one decoder for one DNN feature unit), and the trained decoders were then applied to the test datasets. For details of the general procedure of feature decoding, see Horikawa & Kamitani (2017) [10].
For the test datasets, fMRI samples corresponding to the same stimulus or mental imagery were averaged across trials to increase the signal-to-noise ratio of the fMRI signals. To compensate for possible differences in the signal-to-noise ratio between training and test samples, the decoded features of individual DNN layers were normalized by multiplying them by a single scalar, so that the norm of the decoded vectors of individual DNN layers matched with the mean norm of the true DNN feature vectors computed from independent 10,000 natural images. This norm-corrected vector was then subsequently provided to the reconstruction algorithm (See Supporting Information for details of the norm-correction procedure).
Given a DNN feature vector decoded from brain activity, an image was generated by solving the following optimization problem [11].
v*=argminv12∑i=1Il(ϕi(l)(v)−yi(l))2
(1)
=argminv12‖Φ(l)(v)−y(l)‖22
(2)
where v∈R224×224×3 is a vector whose elements are pixel values of an image (224 × 224 × 3 corresponds to height × width × RGB color channel), and v* is the reconstructed image. ϕi(l):R224×224×3→R is the feature extraction function of the i-th DNN feature in the l-th layer, with ϕi(l)(v) being the output value from the i-th DNN unit in the l-th layer for the image v. Il is the number of units in the l-th layer, and yi(l) is the value decoded from brain activity for the i-th feature in the l-th layer. For simplicity, the same cost function was rewritten with a vector function in the second line. Φ(l):R224×224×3→RIl is the function whose i-th element is ϕi(l) and y(l)∈RIl is the vector whose i-th element is yi(l).
The above cost function was minimized by either a limited-memory BFGS algorithm (L-BFGS) [32–34] or by a gradient descent with momentum algorithm [35], with L-BFGS being used unless otherwise stated. The obtained solution was taken to be the image reconstructed from the brain activity (see Supporting Information for details of optimization methods).
To combine the DNN features from multiple layers, we took a weighted sum of the cost functions for individual DNN layers, given by
v*=argminv12∑l∈Lβl‖Φ(l)(v)−y(l)‖22
(3)
where L is a set of DNN layers and βl is a parameter that determines the contribution of the l-th layer. We set βl to 1/‖y(l)‖22 to balance the contributions of individual DNN layers. This cost function was minimized by the L-BFGS algorithm. The DNN layers included in L were combined. In the main analyses, we combined all convolutional (DNN1–5) and fully connected layers (DNN6–8), unless otherwise stated.
To improve the ‘naturalness’ of reconstructed images, we modified the reconstruction algorithm by introducing a constraint. To constrain the resulting images from all possible pixel contrast patterns, we reduced the degrees of freedom by introducing a generator network derived using the generative adversarial network algorithm (GAN) [15], which has recently been shown to have good performance in capturing a latent space explaining natural images [16]. In the GAN framework, a set of two neural networks, which are called a generator and a discriminator, are trained. The generator is a function to map from a latent space to the data space (i.e. pixel space), and the discriminator is a classifier that predicts whether a given image is a sample from real natural images or an output from the generator. The discriminator is trained to increase its predictive power, while the generator is trained to decrease it. We considered constraining our reconstructed images to be in the subspace consisting of the images that could be produced by a generator trained to produce natural images [12, 36]. This is expressed by
z*=argminz12∑l∈Lβl‖Φ(l)(G(z))−y(l)‖22
(4)
and
v*=G(z*).
(5)
G is the generator, as the mapping function from the latent space to the image space, which we have called a deep generator network (DGN). In our reconstruction analysis, we used a pre-trained DGN which was provided by Dosovitskiy & Brox (2016; available from https://github.com/dosovits/caffe-fr-chairs; trained model for fc7) [36].
The above cost function for the reconstruction with respect to z was minimized by gradient descent with momentum. We used the zero vector as the initial value. To keep z within a moderate range, we restricted the range of each element of z following the method of a previous study [36].
Reconstruction quality was evaluated by either objective or subjective assessment [5, 17, 18]. For the objective assessment, we performed a pairwise similarity comparison analysis, in which a reconstructed image was compared with two candidate images (its original image and a randomly selected image), to test whether its pixel-wise spatial correlation coefficient (Pearson correlation between vectorized pixel values) with the original image was higher than that for a randomly selected image. For the subjective assessment, we conducted a behavioral experiment with a group of 13 raters (5 females and 8 males, aged between 19 and 37 years). On each trial of the experiment, the raters viewed a display presenting a reconstructed image (at the bottom) and two candidate images (displayed at the top; the original image and a randomly selected image), and were asked to select the candidate image most similar to the reconstructed one presented at the bottom. Each trial continued until the raters made a response. For both types of assessments, the proportion of trials, in which the original image was selected as more similar was calculated as a quality measure. In both objective and subjective assessments, each reconstructed image was tested with all pairs of the images from the same types of images (natural-images, artificial-shapes, and alphabetical-letters for images from the image presentation sessions, and natural-images and artificial-shapes for images from the mental imagery session; e.g., for the test natural-images, one of the 50 reconstructions was tested with 49 pairs, with each one consisting of one original image and another image from the rest of 49, resulting in 50 × 49 = 2,450 comparisons). The quality of an individual reconstructed image was evaluated by the percentage of correct answers that was calculated as the proportion of correct trials among all trials where the reconstructed image was tested (i.e., a total of 49 trials for each one of the test natural-images). The resultant percentages of correct answers were then used for the following statistical tests.
To compare the reconstruction quality across different combinations of DNN layers, we also used objective and subjective assessments. For the subjective assessment, we conducted another behavioral experiment with another group of 7 raters (2 females and 5 males, aged between 20 and 37 years). On each trial of the experiment, the raters viewed a display presenting one original image (at the top) and two reconstructed images of the same original image (at the bottom) obtained from different combinations of the DNN layers, and were asked to judge which of the two reconstructed images was better. This pairwise comparison was conducted for all pairs of the combinations of DNN layers (28 pairs), and for all stimulus images presented in the test natural-image session (50 samples). Each trial continued until the raters made a response. We calculated the proportion of trials, in which the reconstructed image obtained from a specific combination of DNN layers was judged as the better one, and then this value was treated as the winning percentage of this combination of DNN layers. For the objective assessment, the same pairwise comparison was conducted using pixel-wise spatial correlations, in which pixel-wise spatial correlations to the original image were compared between two combinations of DNN layers to judge the better combination of DNN layers. The results obtained from all test samples (50 samples from the test natural-image dataset) were used to calculate the winning percentage of each combination of DNN layers in the same manner with the subjective assessment.
These assessments were performed individually for each set of reconstructions from the different subjects and datasets (e.g., test natural-images from Subject 1). For the subjective assessments, one set of reconstructed images was tested with at least three raters. The evaluation results from different raters were averaged within the same set of reconstructions and were treated in the same manner as the evaluation results from the objective assessment.
We used two-sided signed-rank tests to examine differences in assessed reconstruction quality according to the different conditions (N = 150, 120, and 45 for the test-natural images, artificial shapes, and imagery images, respectively) and used ANOVA to examine interaction effects between task types and brain areas for artificial shapes (F (1,1) = 28.40 by human judgment; F (1,1) = 3.53 by pixel-wise spatial correlation). We used one-sided signed-rank tests to examine the significance of correct classification accuracy by the human judgment for evaluations of the imagery image reconstructions (N = 45).
|
10.1371/journal.pgen.1007693 | Non-proteolytic activity of 19S proteasome subunit RPT-6 regulates GATA transcription during response to infection | GATA transcription factors play a crucial role in the regulation of immune functions across metazoans. In Caenorhabditis elegans, the GATA transcription factor ELT-2 is involved in the control of not only infections but also recovery after an infection. We identified RPT-6, part of the 19S proteasome subunit, as an ELT-2 binding partner that is required for the proper expression of genes required for both immunity against bacterial infections and recovery after infection. We found that the intact ATPase domain of RPT-6 is required for the interaction and that inhibition of rpt-6 affected the expression of ELT-2-controlled genes, preventing the appropriate immune response against Pseudomonas aeruginosa and recovery from infection by the pathogen. Further studies indicated that SKN-1, which is an Nrf transcription factor involved in the response to oxidative stress and infection, is activated by inhibition of rpt-6. Our results indicate that RPT-6 interacts with ELT-2 in vivo to control the expression of immune genes in a manner that is likely independent of the proteolytic activity of the proteasome.
| The conserved GATA transcription factor ELT-2 plays an important role in the control of genes required for both defense and recovery from infection. We show that RPT-6, a component of the 19S subunit, physically interacts with ELT-2 in vivo, controlling the expression of ELT-2-dependent genes and the response of the nematode Caenorhabditis elegans to bacterial infection. The proteolytic activity of the proteasome has surfaced as a key regulator of gene expression, but our results provide evidence indicating that a non-canonical activity of the 26S proteasome subunit plays an important role in the control of gene expression during the response to bacterial infection.
| The regulation of gene transcription plays crucial roles in the control of an array of critical biological processes, including immune responses against microbial infections [1,2]. At the heart of immune activation are transcription factors, which directly bind to specific DNA motifs in promoter regions to control gene transcription. The GATA transcription factor ELT-2 is a major component of innate immunity and lies downstream of the conserved PMK-1/p38 mitogen-activated protein kinase (MAPK) signaling pathway in the nematode Caenorhabditis elegans [3]. Indeed, ELT-2 has been demonstrated to control immune responses against several human bacterial pathogens, including Pseudomonas aeruginosa, Salmonella enterica, Enterococcus faecalis, and Cryptococcus neoformans. [4–7]. Furthermore, ELT-2 has recently been shown to be involved in the control of host changes that also take place during recovery from bacterial infections [6,7]. Despite the several aforementioned studies describing the importance of ELT-2 in the regulation of myriad target genes required for response and recovery from infections, nothing is known about the mechanisms involved in the control of ELT-2 transcriptional activity.
The 26S proteasome has surfaced as a key regulator of gene expression, mostly via its proteolytic activity. However, increasing evidence suggests a non-proteolytic role of the 26S proteasome or its sub-complex, notably the 19S regulatory subunit, in the control of gene transcription [8,9]. This non-canonical activity of the 26S proteasome has been linked to the control of various aspects of gene transcription, including initiation and elongation steps, and chromatin remodeling [10–12]. Recent biochemical and genetic studies have shown that the proteasome can physically interact with transcription factors and regulate their interactions with coactivators as well as promoter regions, all leading to the control of gene activation [8,10,13]. For example, in yeast and mammalian cells, SUG1/RPT6 and some other proteasome subunits can physically interact with transcription factors to control gene transcription in a non-proteolytically fashion [8,14–17]. Also, viral gene transcription has been associated with the non-proteolytic activity of the 19S subunits during gene expression [11,18], but the role of non-canonical functions of the proteasome in the control of defense against pathogens and activation of innate immunity has not been studied.
We identified RPT-6, a component of the 19S proteasome subunit, as a binding partner of ELT-2. We showed that inhibition of rpt-6 leads to inactivation of ELT-2-regulated immune genes during P. aeruginosa infection, resulting in enhanced susceptibility to the pathogen, similar to that observed in animals deficient in elt-2. We also demonstrated that both elt-2 and rpt-6 work together to control recovery after P. aeruginosa infection has been cleared. Disruption of the proteasome complex, but not the inhibition of proteasomal activity, prevents ELT-2 transcriptional activation of immune genes. Finally, we demonstrate that both RPT-6 and ELT-2 physically interact in vivo in C. elegans and that this interaction is affected by other components of the proteasome complex. Our findings show that the proteasome interacts with the transcription factor ELT-2 to control the activation of immune genes, extending our knowledge of the role of the proteasome in gene transcription to innate immunity.
As a first step to further understand the mechanisms by which ELT-2 controls gene expression, we attempted to identify potential interacting proteins with ELT-2/GATA during infection of C. elegans. A C. elegans strain carrying a stably integrated ELT-2::GFP transgene was exposed to P. aeruginosa prior immunoprecipitation of GFP-tagged ELT-2. A total of 14 candidate ELT-2-interacting proteins were identified using liquid chromatography-tandem mass spectrometry (LC–MS/MS) (S1 Table)
To validate the role of the potential ELT-2 partners in the control of innate immunity, we studied the susceptibility to P. aeruginosa of animals in which 10 candidate genes were inhibited by RNAi using all the commercially available clones to inhibit genes in C. elegans. Because ELT-2 is essential for C. elegans larval development [19], we reasoned that genes encoding the candidate binding partner would also be essential. Thus, we performed RNAi to downregulate elt-2 and candidate genes at late larval stage 4 (L4). RNAi against four of the genes led to susceptibility to PA14 infection in the animals. However, only rpt-6(RNAi) displayed a very robust enhanced pathogen susceptibility that was comparable to that exhibited by elt-2(RNAi) animals (Table 1, S1A Fig). As ELT-2 transcriptional activity has been linked to recovery from acute bacterial infection [6,7], we reasoned that potential ELT-2 partners may also affect recovery from bacterial infection. We studied the survival of animals in which the candidate genes were knocked down by RNAi, infected with P. aeruginosa, and treated with streptomycin. We found that rpt-6(RNAi) animals failed to recover after P. aeruginosa infection (Table 1, S1B Fig). RNAi downregulation of gpb-1 also enhanced susceptibility to infection and affected recovery. Inhibition of gpb-1 causes defects in the body wall muscles [20], which may prevent bacterial clearance and affect recovery. Thus, this gene was not further analyzed and we focused on rpt-6, which like elt-2, controls both response to infection and recovery.
To study the mechanism by which RPT-6 controls immune activation, we first used a PF55G11.2::gfp transcriptional reporter strain that expresses GFP under the control of the promoter of F55G11.2, which is an ELT-2-dependent gene [3,4]. Inactivation of elt-2 by RNAi usually leads to downregulation of the basal expression of F55G11.2 [3]. Upon rpt-6 RNAi, we observed a reduction of GFP expression in PF55G11.2::gfp animals similar to that observed when elt-2 is also inhibited (Fig 1A; Fig 1B). Next, we used qRT-PCR to examine the effect of rpt-6 inactivation on the induction of selected immune genes that contain the TGATAA ELT-2 binding motif in their proximal promoter regions and that are usually activated during P. aeruginosa infection [7]. Four out of the five examined immune genes failed to be activated during P. aeruginosa infection in both elt-2(RNAi) animals and rpt-6(RNAi) animals (Fig 1C). Figs 1D and S1A show that although rpt-6(RNAi) animals were not as susceptible as elt-2(RNAi) animals to P. aeruginosa infection, they were significantly more susceptible than the control animals (P<0.0001). The overall expression of immune genes in elt-2(RNAi) animals is not different from that of elt-2(RNAi);rpt-6(RNAi) animals (S2 Fig). S3A and S3B Fig shows that elt-2 and rpt-6 were effectively inhibited in the elt-2(RNAi) and elt-2(RNAi);rpt-6(RNAi) animals. Next, we aimed to determine whether rpt-6 also played a role during recovery from bacterial infection. While RNAi downregulation of rpt-6 prevented the activation of two out of the four tested genes involved in recovery from P. aeruginosa infection (Fig 1E), rpt-6(RNAi) animals failed to recover from the infection (Fig 1F). These results suggest that rpt-6 and elt-2 may work together to control the expression of a subset of immune genes.
Knockdown of elt-2 or rpt-6 leads to larval arrest [21]. To study whether the enhanced pathogen susceptibility of rpt-6(RNAi) animals is a consequence of the animals being sickly, we performed lifespan assays on rpt-6(RNAi) and elt-2(RNAi) animals. The result shows that both elt-2(RNAi) and rpt-6(RNAi) animals displayed comparable lifespans, although a small reduction was observed compared to control animals (S4 Fig). However, the rate of death of elt-2(RNAi) and rpt-6(RNAi) animals exposed to P. aeruginosa is higher than that of animals exposed to killed E. coli (Fig 1D and S4 Fig). These results indicate that the enhanced pathogen susceptibility is not simply a consequence of the animals being sickly, and are in agreement with other studies indicating that elt-2 also controls lifespan in C. elegans [5,22,23].
Finally, we were interested in determining whether other components of the 26S proteasome also affected ELT-2 transcriptional activity. Thus, we examined the fluorescence expression of PF55G11.2::gfp animals after knocking down the proteasome genes rpt-3, pbs-2, pas-6, and rpn-11. RNAi of these genes, which belong to the three 26S sub-complexes (19S lid, 19S base, and 20S core), reduced PF55G11.2::gfp fluorescence (S5A Fig). We also examined whether another component of the 26S proteasome, RPN-11, was involved in the control of expression of ELT-2-dependent genes. As shown in S5B Fig, rpn-11(RNAi) animals failed to activate ELT-2 regulated-immune genes during P. aeruginosa infection, with the exception of irg-6. Consistent with these results, rpn-11(RNAi) animals were also susceptible to P. aeruginosa infection (S5C Fig). Thus, we concluded that rpt-6 and other components of the 26S mediate ELT-2 transcriptional activity against P. aeruginosa infection.
We observed that although rpt-6(RNAi) animals were significantly susceptible to P. aeruginosa infection, they were not as susceptible to the pathogen as elt-2(RNAi) animals (Fig 1D). However, rpt-6(RNAi) animals exhibited a downregulation of immune genes comparable to that of elt-2(RNAi) animals (Fig 1C). SKN-1/Nrf is a transcription factor that is involved in the response to oxidative stress and protection against bacterial infection via the PMK-1/p38 MAPK pathway [24]. Because SKN-1 is known to be activated when the proteasome system is perturbed by the inactivation of proteasome genes [25,26], we reasoned that SKN-1 activation might confer some protection against P. aeruginosa infection in rpt-6(RNAi) animals. Therefore, we examined the expression of selected SKN-1 target genes [24]. We found that in the absence of infection, gst-4, gst-5, and gst-10 were upregulated in rpt-6(RNAi) animals compared with the control animals (Fig 2A). Upon exposure to P. aeruginosa, the expression of gst-4 and gst-5 was still upregulated in rpt-6(RNAi) animals. In contrast, the expression of these SKN-1 reporter genes was significantly downregulated in elt-2(RNAi) animals during infection (Fig 2B). Because skn-1 transcripts were not upregulated in rpt-6(RNAi) (Fig 2A; Fig 2B), SKN-1 might be activated post-transcriptionally by the disruption in the proteasome caused by rpt-6 inhibition. To confirm the idea that SKN-1 partially protects rpt-6(RNAi) animals from P. aeruginosa infection, we used RNAi to inhibit both rpt-6 and skn-1. As shown in Fig 2C, these animals were as susceptible to P. aeruginosa infection as elt-2(RNAi) animals, indicating that activation of SKN-1 partially compensated for the immune deficiency of rpt-6(RNAi) animals via an ELT-2-independent mechanism.
Because ELT-2-dependent genes are downregulated during infection in rpt-6(RNAi) animals, we wanted to investigate whether this downregulation of ELT-2 target genes was linked to the inhibition of proteasomal activity. We inhibited the proteasome activity using the proteasome inhibitor bortezomib. First, we ascertained that the proteasome was indeed inhibited in rpt-6(RNAi) animals as well as in bortezomib-treated animals by using the sur-5::UbV-gfp reporter strain. Normally, inhibition of the ubiquitin–proteasome system (UPS) leads to stabilization of the UbV-GFP fusion protein due to the absence of protein degradation [27]. RNAi of rpt-6 in the reporter strain led to an elevated UbV-GFP fluorescent signal (S6A and S6B Fig), indicating an inhibition of proteasome activity in rpt-6(RNAi) animals. Likewise, treatment of the reporter strain with bortezomib resulted in the stabilization of the UbV-GFP fluorescent signal (S6C–S6F Fig), indicating inhibition of the proteasomal activity. However, similar treatment with bortezomib had no significant effect on the expression of ELT-2-dependent immune genes activated during P. aeruginosa infection (Fig 3A). These results indicate that inhibition of proteasome activity is unlikely to impact ELT-2 transcriptional activation of immune genes during infection.
Next, we investigated whether rpt-6 RNAi prevented ELT-2 transcriptional activity by altering its nuclear localization. We utilized an elt-2::gfp reporter strain to quantify both the numbers of nuclei as well as the intensity of GFP fluorescence. As expected, elt-2(RNAi) animals did not show ELT-2::GFP expression, but ELT-2::GFP nuclear localization was still present in rpt-6(RNAi) animals (Fig 3B). The estimation of nuclear ELT-2::GFP in rpt-6(RNAi) animals showed that both the numbers and intensity of nuclear ELT-2::GFP were not significantly different from those observed in control animals (Fig 3C; Fig 3D). Taken together, these results showed that the proteolytic function of the proteasome did not mediate the ELT-2 transcriptional activation of immune genes.
Our co-immunoprecipitation and proteomic analysis indicated that ELT-2 and RPT-6 physically interacted to control gene expression. To confirm the ELT-2/RPT-6 interaction, we employed a bimolecular fluorescence complementation (BiFC) assay, which allows for the determination of physical interactions of proteins in living cells through direct visualization [28,29]. The BiFC constructs are engineered to individually express, in response to heat-shock, GFP protein fragments translationally fused with RPT-6 and ELT-2. Generally, interacting proteins bring the non-fluorescent fragments into close proximity for reconstitution and fluorescence. Twelve hours after heat shock, we observed fluorescence, indicating a physical interaction between RPT-6 and ELT-2 in vivo (Fig 4A). Animals carrying BiFC constructs without elt-2 did not exhibit fluorescence. Knockdown of either rpt-6 or elt-2 by RNAi resulted in the absence of fluorescence, further confirming that the presence of the two proteins is required for the GFP reconstitution (Fig 4A). We wondered whether the ability of other components of the proteasome complex to affect ELT-2 activity may be a consequence of these components affecting the ELT-2/RPT-6 interaction. RNAi of selected proteasome genes reduced the extent of the ELT-2/RPT-6 in vivo interaction to varying degrees (Fig 4B; Fig 4C).
The ATPase activity of the 19S proteasome has been implicated in the mediation of the interaction between proteasome subunits and transcription factors [30,31]. Thus, we asked whether the active site of the ATPase domain of RPT-6 played a role in mediating the interaction with ELT-2 in vivo, and we found that GFP fluorescence is significantly reduced when a single amino acid in the ATPase domain of RPT-6 is mutated (Fig 4D; Fig 4E). Taken together, these results indicate that the ATPase activity of RPT-6 is needed for its interaction with ELT-2 in vivo and that this interaction required other components of the 26S proteasome complex.
ELT-2/GATA transcription plays a crucial role in the control of immunity in C. elegans against both bacterial and fungal pathogens by cooperating with the PMK-1/p38 MAPK and SKN-1/Nrf pathways [3,5,32]. However, no information is available concerning the proteins that interact with ELT-2 to participate in the control of gene expression. It is known that transcription factors can recruit and associate with co-regulatory proteins during gene transcription. In mammalian cells, the immunoproteasome, the proteasome-isoform of the standard proteasome, is known to mediate the immune response, but mainly through its proteolytic function, due to its higher degradation capability [33]. Here we showed that RPT-6, a component of 19S, could non-proteolytically mediate ELT-2 transcriptional activity and physically interact with ELT-2.
Our findings showed that inactivation of rpt-6 inhibited the expression of ELT-2 target genes during infection. Our previous work has demonstrated that ELT-2 activates not only immune genes in response to pathogens, but also genes responsible for recovery from acute bacterial infections. Our new findings indicate that mediation of ELT-2 transcriptional activity by the proteasome is not limited to the immune response, but it is also involved in the recovery from bacterial infection. Inhibition of proteasome activity by bortezomib did not prevent the transcriptional activation of ELT-2 target genes during infection, indicating that RPT-6 control of ELT-2 targets is unlikely related to the proteolytic activity. We cannot completely rule out the possibility that a repressor of ELT-2, that is commonly targeted for proteolysis, reduces ELT-2 activity in rpt-6(RNAi) animals. However, we believe that the failure of bortezomib to alter ELT-2-mediated gene expression at the same concentrations at which it significantly inhibits the proteasome, makes this possibility unlikely.
Accumulating evidence indicates that different components of the proteasome interact with transcription factors to control gene expression in a manner that is independent of the canonical proteasomal activity [8,16,17,34,35]. The current notion is that components of the proteasome are recruited to the site of transcription for the subsequent mobilization of other transcriptional machinery, including chromatin remodeler and stabilization of enhanceosome at the site of transcription [16,17,34,35]. Similarly, our findings show that the inactivation of RTP-6 affects ELT-2-mediated transcription. Our results also suggest that other components of the proteasome affect both ELT-2 transcriptional activity and the ELT-2/RPT-6 interaction in vivo. Indeed, the three subunits of the proteasome have been shown to interact in yeast with activated GAL10, a galactose metabolism structural component [34].
Perturbation of core cellular activities, including proteasomal function, induces the expression of detoxification and innate immune response genes [36]. Here we show that inhibition of the RPT-6 subunit of the proteasome represses the expression of innate immune genes. Our results indicate that this effect is not related to the proteasomal activity, but rather due to the lack of physical interaction between RPT-6 and ELT-2. Thus, disruptions in the proteasome have different effects on other transcription factors. Proteasomal perturbations upregulate the expression of genes that are reporter of the activity of other transcription factors, including EGL-9, ZIP-2, ELT-3 and SKN-1 [26,36–39]. Consistently with these observations, we found that SKN-1-controlled genes were induced in rpt-6(RNAi) animals.
In conclusion, our results demonstrate that the proteasome is an important player in transcriptional activation during the immune response in C. elegans. We identified RPT-6, a subunit of the 19S proteasome, as the binding partner of the ELT-2/GATA transcription factor. Inactivation of rpt-6 and other components of the proteasome attenuates the activation of immune genes that are regulated by ELT-2. However, proteasome inhibition does not abolish the activation of immune genes, indicating that this control of innate immunity by the proteasome occurs in a non-proteolytic manner. In addition, ELT-2 and RPT-6 physically interact, a phenomenon that is affected by other components of the proteasome and that it requires the ATPase activity of RPT-6. It is noteworthy that both ELT-2/GATA and the proteasome are evolutionarily conserved. The two closest homologs to ELT-2 in humans, GATA6 and GATA4, also participate in the control of immunity. The former possesses an immune protective function against P. aeruginosa infection in human lung epithelial cells, while the latter affects intestinal immunity [4,40]. The non-proteolytic interaction between the proteasome and ELT-2 described herein may be part of a conserved mechanism involved in the control of immune responses.
All C. elegans strains used were grown and maintained on standard NGM-OP50 plates at 15°C. Bristol N2, HH142 [fer-1(b232ts)], OP56 gaEx290 [elt-2::TY1::EGFP::3xFLAG(92C12) + unc-119(+)], SD1949 glo-4(ok623) V; gaIs290 [elt-2::TY1::EGFP::3xFLAG(92C12) + unc-119(+)] were obtained from the Caenorhabditis Genetics Center. PF55G11.2::gfp worms were provided by Michael Shapira (University of California Berkeley, Berkeley, California), and hhIs64 [unc-119(+); sur-5::UbV-GFP]III; hhIs73 [unc-119(+); sur-5::mCherry] was constructed by T. Hoppe, University of Cologne. The following bacterial strains were used in this study: E. coli OP50-1 [SmR], P. aeruginosa PA14, and E. coli strain HT115 pL4440-RNAi (HT115-RNAi) [AmpR, TetR]. In all experiments, fer-1 animals were grown at the 15°C permissive temperature to gravid adults. Egg laying and subsequent growth for assays were carried out at the 25°C non-permissive temperature.
Strain OP56 carrying the ELT-2::GFP transgene was crossed with fer-1 to obtain the transgenic strain with fer-1 background, this was done to avoid progeny issues in the subsequent assay. Approximately 2,000 young adult animals grown at 25°C were washed with M9 and transferred unto the P. aeruginosa PA14 plate for 12 hours at 25°C. The animals were harvested, washed several times in M9 buffer, and frozen in PBS buffer. The worm pellet was later sonicated, and the proteins were immunoprecipitated using GFP-Trap A beads (Chromotek, Germany) at 4°C for 4 hours following the manufacturer’s guidelines. The complexes were washed with 50 mM NH4HCO3 and sent to Duke University Proteomics Core Facility for proteomic analysis. In-gel trypsin digestion was performed according to a standard protocol (https://genome.duke.edu/sites/genome.duke.edu/files/In-gelDigestionProtocolrevised_0.pdf). Qualitative LC/MS/MS was performed on the sample using a nanoAcquity UPLC system (Waters Corp) coupled to a Thermo QExactive Plus high-resolution accurate mass tandem mass spectrometer (Thermo). Raw data were processed using the Mascot Distiller and Mascot Server (v2.5, Matrix Sciences), and Scaffold v4 (Proteome Software, Inc) was used for curation. The protein and peptide threshold FDR were set to 5% and 1%, respectively.
Bacterial clones obtained from the Ahringer library and empty RNAi vector (EV) were grown in LB broth containing 100 μg/ml ampicillin at 37°C for 9 hours, concentrated, and spread onto NGM plates containing 100 μg/ml ampicillin plus 3 mM isopropyl 1-thio-β-D-galactopyranoside. RNAi-expressing bacteria were incubated at 37°C overnight and further overnight at room temperature to produce a thick bacterial lawn. Because most of the genes were lethal when knocked down during the larval stage, RNAi was initiated at L4, excluding skn-1, which was started during egg laying. Young adult fer-1 animals grown at the 25°C non-permissive temperature for 36 hours were fed bacteria expressing dsRNA for 36 hours at 25°C. Co-RNAi was performed by mixing the respective bacteria clones 1:1 before seeding. The nematodes were then used for subsequent assays. Control animals were grown on empty vector in all cases. All RNAi clones were verified by DNA sequencing.
P. aeruginosa killing: PA14 grown in LB broth for 15 hours was seeded onto modified NGM agar medium (0.35% instead of 0.25% peptone), and the plates were incubated overnight at 37°C. Synchronized animals were transferred onto PA14-seeded plates and incubated at 25°C. Animals were transferred onto a new pathogen lawn every day and scored every 12 hours. Animals were considered dead when they failed to respond to touch. Each experiment performed independently in triplicate, consisting of 50 animals in each experiment.
P. aeruginosa recovery assay: After infection for 12 hours as described above, infected or control animals grown on E. coli OP50 were rinsed by transferring them unto 120 μl M9 plus 300 μg/ml streptomycin on NGM-E. coli OP50 plates. The animals were allowed to swim out of the solution and unto the OP50 lawn, and they were subsequently transferred to modified NGM plates containing 300 μg/ml streptomycin seeded with E. coli OP50. Animals were transferred onto new plates containing antibiotics and seeded with E. coli OP50 daily and scored every 24 hours. Animals were considered dead when they failed to respond to touch. All experiments were performed independently in triplicate, consisting of 50 animals in each experiment.
Approximately 2,000 RNAi or control animals were washed with M9, transferred onto P. aeruginosa PA14-modified NGM plates for 12 hours at 25°C, and harvested in M9 buffer. For recovery, after12 hours of infection, animals were washed with several changes of M9 and M9 plus 300 μg/ml streptomycin and transferred onto E. coli OP50 plates containing 300 μg/ml streptomycin for 6 hours at 25°C. Harvested animals were washed with M9 and frozen in TRIzol (Life Technologies, Carlsbad, CA). Total RNA extraction and qRT-PCR using 2 μg total RNA were carried out as previously described [41]. All primer sequences used are available upon request. Primer sequences used for SKN-1 targets were from Hoeven et al. [24].
Bortezomib (Millipore, Temecula, CA, USA), a proteasome inhibitor, dissolved in DMSO to a stock concentration of 5 mM (w/v), was spread onto three-day-old NGM E. coli OP50-seeded plates to a final concentration of 100 nM. Control plates contained only DMSO or NGM E. coli OP50. All plates contained a final concentration of 0.002% DMSO. Approximately 1,500 L4 stage animals were added onto individual plates and incubated for 13 hours at 20°C. Because bortezomib is a reversible inhibitor, to ensure that the proteasome was still inhibited, we flooded the overnight incubated PA14-seeded plates with bortezomib (100 nM final concentration) or DMSO. Drug and DMSO-treated animals were transferred onto PA14-seeded plates, while the animals grown on NGM E. coli OP50 only were transferred onto E. coli OP50 plates. All plates were incubated at 25°C for 4 hours. After infection, animals were harvested, rinsed, and frozen in TRIzol (Life Technologies, Carlsbad, CA) for RNA isolation.
To construct plasmids for the BiFC assay for protein interaction, elt-2 and rpt-6 cDNA (GE Healthcare Dharmacon Inc.) were subcloned into pCE-BiFC-VN173 and pCE-BiFC-VC155 plasmids (Addgene, Cambridge, MA), respectively, both of which contain the heat shock promoter Phsp-16.41. Full-length elt-2 cDNA was subcloned in-frame into pCE-BiFC-VN173 between SmaI and AgeI, while the full-length rpt-6 cDNA was also subcloned in-frame into pCE-BiFC-VC155 between SmaI and KpnI. The generation of the RPT-6 ATPase mutant construct was performed using site-directed mutagenesis on the pCE-BiFC-rpt-6::VC155 to replace lysine at position 206 of RPT-6 with arginine. The BiFC plasmid constructs were injected into N2 worms at 15 ng/μl each, together with pRF4(rol-6) at 100 ng/μl (co-injection marker) [29]. To detect the interaction, transgenic worms carrying the BiFC plasmid constructs were raised to young adults at 20°C, heat shocked for 3 h at 33°C, and allowed to recover for 12 hours at 20°C. Direct visualization of fluorescent signals of the induced expression of fusion proteins (ELT-2 and RPT-6) was captured using a Leica M165 FC fluorescence stereomicroscope. The BiFC assay involving RNAi of other components of the 26S for estimation of the ELT-2/RPT-6 interaction was performed independently in triplicate with 50 animals. One-way ANOVA Dunnett's multiple comparisons test was employed for evaluation.
|
10.1371/journal.pntd.0000287 | Landscape Composition and Spatial Prediction of Alveolar Echinococcosis in Southern Ningxia, China | Alveolar echinococcosis (AE) presents a serious public health challenge within China. Mass screening ultrasound surveys can detect pre-symptomatic AE, but targeting areas identified from hospital records is inefficient regarding AE. Prediction of undetected or emerging hotspots would increase detection rates. Voles and lemmings of the subfamily Arvicolinae are important intermediate hosts in sylvatic transmission systems. Their populations reach high densities in productive grasslands where food and cover are abundant. Habitat availability is thought to affect arvicoline population dynamic patterns and definitive host–intermediate host interactions. Arvicoline habitat correlates with AE prevalence in Western Europe and southern Gansu Province, China.
Xiji County, Ningxia Hui Autonomous Region, borders southern Gansu. The aims of this study were to map AE prevalence across Xiji and test arvicoline habitat as a predictor. Land cover was mapped using remotely sensed (Landsat) imagery. Infection status of 3,205 individuals screened in 2002–2003 was related, using generalised additive mixed models, to covariates: gender; farming; ethnicity; dog ownership; water source; and areal cover of mountain pasture and lowland pasture. A Markov random field modelled additional spatial variation and uncertainty. Mountain pasture and lowland pasture were associated with below and above average AE prevalence, respectively.
Low values of the normalised difference vegetation index indicated sub-optimality of lowland pasture for grassland arvicolines. Unlike other known endemic areas, grassland arvicolines probably did not provide the principal reservoir for Echinococcus multilocularis in Xiji. This result is consistent with recent small mammal surveys reporting low arvicoline densities and high densities of hamsters, pikas and jerboas, all suitable intermediate hosts for E. multilocularis, in reforested lowland pasture. The risk of re-emergence is discussed. We recommend extending monitoring to: southern Haiyuan County, where predicted prevalence was high; southern Xiji County, where prediction uncertainty was high; and monitoring small mammal community dynamics and the infection status of dogs.
| In humans, larvae of the fox tapeworm Echinococcus multilocularis typically infect the liver where metastasis, calcification and necrosis cause the zoonotic disease alveolar echinococcosis (AE). Treatment is difficult. Early detection greatly increases patient life expectancy but under-detection is a problem. Understanding the ecological conditions that elevate AE risk would help identify at-risk communities. Voles and lemmings of the subfamily Arvicolinae are important intermediate hosts in most AE endemic areas, and arvicoline habitat has been proposed as a predictor of AE risk. Using a model of spatial autocorrelation with land cover identified from satellite remote sensing imagery, we identified AE hotspots in southern Ningxia Hui Autonomous Region (NHAR), China. Hotspots were not located near optimal arvicoline habitats. Thus, non-arvicolines provide principal reservoirs in NHAR and the range of ecological conditions sustaining E. multilocularis transmission in China is greater than previously thought. We also show: social factors explain higher prevalence in females than males; dogs increase infection risk; and we argue that water source quality is important via interaction with other environmental variables. Our map of AE prevalence represents the current state-of-the-art regarding the spatial distribution of AE in southern NHAR and provides an important baseline for future monitoring programs there.
| Biological mechanisms known to affect space-time dynamics of infectious diseases include: habitat changes affecting vector breeding sites or reservoir host distributions; niche invasion; biodiversity change including keystone predator loss and rapid magnitudinal increases in reservoir host populations; genetic change in vectors or pathogens; and environmental contamination with infectious agents [1]. Each of these mechanisms may be affected by ecosystem change and the last 50 years have seen the greatest changes in ecosystem structure and function in human history [2]. The Millennium Ecosystems Assessment of the World Health Organisation has listed over thirty infectious diseases known to be affected by ecosystem changes [1]. The list provides compelling evidence that ecological factors affect transmission of many of the most dangerous pathogens and zoonoses. However, the list was incomplete. Its failure to mention the fatal parasitic disease alveolar echinococcosis (AE) reflects that this very dangerous zoonosis is indeed a neglected disease. Despite being globally rare, AE places a serious burden on affected communities in endemic areas and remains very difficult to treat. As for many zoonoses, incidence rates of AE are affected by ecosystem changes. This paper explores the statistical relationships between land cover and human AE prevalence in southern Ningxia Hui Autonomous Region (NHAR) (Fig. 1), China. The identified statistical relationships are then used to map AE prevalence across the area.
Alveolar echinococcosis arises from infection with larvae of the fox tapeworm Echinococcus multilocularis [3]. In Europe, prevalences in foxes and the geographical range of the worm have increased giving rise to fears of AE emergence [4],[5],[6]. In central China, high AE prevalences have been reported from: Tibetan pastoral communities of northwest Sichuan [7],[8],[9]; Han communities of southern Gansu [10]; and Hui communities of southern NHAR [11]. A distribution map human AE across China is provided by [12]. The central Chinese endemic area has been described in meta-population terms with the grasslands of northwest Sichuan sustaining a large and stable meta-focus of E. multilocularis transmission that feeds peripheral areas where stability is lower by function of reduced availability of, and connectivity between, patches of optimal intermediate host habitat [13]. An average Tibetan pastoralist of northwest Sichuan is estimated to lose 0.81 Disability Adjusted Life Years (DALYs) to alveolar and cystic echinococcosis [14]. Compared to an average 0.18 DALYs lost in the general Chinese population due to all communicable and non-communicable ailments combined [14] it is clear that echinococcosis is a major burden for communities in endemic areas of China and poses a public health problem of primary importance.
Severe liver pathology and metastasis of multilocular cysts during long and asymptomatic incubation periods renders treatment of symptomatic AE cases extremely difficult [3]. Early detection is critical regarding patient life expectancy. Mass screening programs do successfully detect early cases [15],[16], but identifying target areas from hospital records is unreliable in relation to AE [11]. Reliance upon hospital records suffers two problems: i) under-detection in remote areas with limited access to medical facilities or poor knowledge of AE; ii) slow response to epidemiological shifts affected by environmental change. Public health managers would benefit from predictive models that could identify undetected or emergent AE hotspots. Moreover, understanding the links between land cover and small mammal communities is essential regarding development of effective environmentally based disease control strategies.
A classic observation is that voles or lemmings of the subfamily Arvicolinae frequently function as key intermediate hosts for E. multilocularis in sylvatic systems [17]. Due to the specific habitat requirements of arvicolines, it has been hypothesised that landscape composition may provide a useful predictor of the spatial distribution of E. multilocularis and AE [18],[19]. In eastern France, regular population outbreaks of the vole Arvicola terrestris occur in areas abundant with large open patches of pasture and positive correlations between percentage cover of grassland and E. multilocularis infection in humans and foxes have been shown [20],[21],[22]. In Zhang County, Gansu, China, trapping frequencies of the vole Microtus limnophilus and hamster Cricetulus longicaudatus were greatest in grass and shrub patches generated by successional growth following deforestation [23]. Population outbreaks of both species had been reported from the area [24] and species in these genera are highly susceptible to E. multilocularis infection [25],[26]. While C. longicaudatus was also abundant in lower prevalence agricultural areas, percentage cover of optimal M. limnophilus habitat correlated positively with human AE prevalence [10],[12],[27]. In Tibetan pastoral communities of northwest Sichuan modern private fencing practices reduce availability of, and increase grazing pressure on, common lands, improving habitat suitability for Microtus fuscus and the lagomorphs Ochotona curzoniae and Ochotona cansus. Tall grass within fenced areas provides sufficient nutrition and protection to support large populations of arvicolines such as M. limnophilus [28]. A correlation between the area of fenced pasture and human AE prevalence has been reported [29]. Details relating to rodent population dynamics and landscape composition are found in [30],[31],[32],[33],[34],[35].
Since economic reform in 1978 China has faced continuous change. The Chinese population has increased from 980 million to 1.3 billion, the urban proportion of which has risen from 20% to 36% and is projected to reach 60% by 2020 [36]. Land cover change in the 1990s encompassed a monolithic 2.99 million hectare increase in cropland and a 0.82 million hectare increase in urban area [36]. In the same period NHAR lost 236700 hectares of grassland and 9200 hectares of unused land and gained 223900, 10100 and 11800 thousand hectares of cropland, woodland and urban land respectively [36]. Proximity to the Gobi desert makes desertification risk in NHAR high [37]. This, and the need for improved flood and erosion management in Yellow River catchments [38], has led to vigorous promotion of afforestation programs [39].
The NHAR is located on the Chinese Loess Plateau. One third of its 5040000 population belongs to the Hui minority [5]. While cystic echinococcosis is found across NHAR, AE has so far only been reported in Xiji, Guyuan and Haiyuan Counties (Fig. 1) [40],[41],[11]. This region lies on the fringes of the known endemic area of central China [42],[12]. The absence of AE north of Haiyuan is attributable to the hot and dry climate [43] providing unsuitable conditions for E. multilocularis egg survival [44]. Dominant geographical features of southern NHAR include: the Liupan mountain range (2927 m); the Yueliang mountain range (2626 m); and the Nanhua mountains (2941 m) just south of Haiyuan City (Figs. 1 & 2). Mountain vegetation is dominantly lush grassland, although some residual forest patches persist in the southern Liupan. Elsewhere land cover is dominated by agricultural fields with hill tops reserved for pasture, albeit of much lower quality than in the mountains.
The current work investigates the hypothesis that AE in Xiji County is attributable to its mountain areas where lush grasslands provide optimal habitat for large, possibly cyclical, arvicoline populations, and relatively temperate climates favour E. multilocularis egg survival. This extrapolation of results from Zhang County [10],[12],[27] assumes that the endemic transmission of the two areas functions within comparable sets of environmental conditions. Specifically, the landscape - disease correlates observed in Zhang are tested as predictors of AE prevalence and maps of predicted prevalence and associated uncertainties are presented.
Records from Xiji (35°33′N - 36°13′N×105°18′E - 106°04E) County Hospital dating from 1985 motivated a mass screening program in 2002–03 described in [11]. Stations for performing abdominal ultrasound screening were established at medical centers or schools in 26 villages. Many people from surrounding villages also participated and in total 4778 individuals in the age range 5–83 years were screened. Approval for the surveys was given by the Ethics Committee of Ningxia Medical College, and written consent was obtained from all adult participants and from parents of minors aged 5 years or older who agreed to participate. Personal details, responses to a knowledge, attitudes and practices questionnaire and AE status were recorded in EpiInfo [45]. These records were combined with spatial coordinates for each village enabling storage in GRASS GIS [46] and spatial analysis in R [47]. The current study used a subset of this data corresponding to those 247 villages which could be geolocated within the study area. Since no AE cases were observed among the 1426 participating students analyses were further restricted to the non-student subset of data. This sub-sample, consisting of 3205 individuals from 152 villages, is summarised in Table 1.
Since latency of AE in humans is long, hepatic lesions detected during mass-screening likely arose from infection events occurring ten years or more prior to screening. In order to analyse the effects of land cover on AE distribution it was pertinent to work with archived satellite remote sensing data. For this a Landsat Multi-Spectral Scanner (MSS) image (acquisition date, April 1978) was used. Five ground control points (GCPs) were collected across the study area by hand-held global positioning system (GPS) for the purpose of geocorrection. These GCPs, typically road junctions or river bridges, were unidentifiable in the 60 m resolution MSS image. Therefore, a Landsat Enhanced Thematic Mapper (ETM) image with a 15 m panchromatic layer (acquisition date, June 2001) was obtained. The GCPs indicated that the pre-purchase automated geocorrection of the ETM image was subject to a georectification error of 250 m in the north west of the study area. A fist order polynomial geocorrection model with nearest neighbour resampling was used to reduce this error (RMSE = 11 m). The MSS image was then georectified to the corrected ETM using larger features identifiable in both images.
Land cover was assessed via a photographic survey in July 2002. Over 250 photographs of the landscape were taken across Xiji, Guyuan and Haiyuan counties. The point from which each photo was taken was recorded using a hand-held GPS receiver and the orientation of the camera was measured with a hand-held compass. These data enabled identification of over 140 homogeneous patches within the ETM image. Training area identification for classification of the MSS image required collection of historical information. This was achieved by discussing land cover change history with local farmers and pastoralists. Training areas were then identified on the basis of these local reports, field data from 2002 and image analysis. A land cover map (Fig. 2) was derived from the MSS image using supervised maximum likelihood classification [48] with the following classes: water bodies; forest; agricultural fields; bare soils; mountain grasslands and lowland pasture. A qualitative assessment of the classified image indicated good correspondence with the available information on historical land cover. All image processing was performed using Erdas Imagine 8.4 [49]. Under the hypothesis, mountain grasslands were expected to have provided both optimal key reservoir habitat and optimal climatic conditions regarding egg longevity. Therefore, that cover class was expected to be positively associated with areas of greatest human AE risk.
Areal cover for each land cover class was estimated as the proportion of pixels belonging to the class in question within a circular buffer centred at a given pixel. This was repeated for every pixel in the image, a technique sometimes referred to as moving window analysis. Such metrics vary with respect not only to landscape composition but also to the buffer radius R. It is generally impossible to anticipate a priori a suitable value for R so a set of values, R∈{500m,1000m,2000m,3000m,…,20000m}, was considered. The most pertinent value for R was identified via Akaike Information Criterion (AIC) based model selection (described below).
Generalised additive models (GAMs) with a logistic link function [50] were used to investigate correlations between risk factors and AE status (presence/absence of hepatic AE) of subjects. The model included the factors: gender (female), ethnic group (Hui), occupation (farmer), dog ownership and water source (tap or well). Non-linearity between age and the logit of prevalence was modelled using a cubic regression spline as described in [50]. Landscape effects were then investigated by adding areal cover estimates of mountain pasture and lowland pasture to the model as linear effects. The optimal R for each class was estimated by an exhaustive comparison of AIC among all possible models. This procedure was also repeated with areal cover estimates of forest included in the model.
The model outlined above was then analysed in a Bayesian context using the software BayesX [51]. This enabled two additional sources of variation to be investigated in a generalised additive mixed model (GAMM) approach: within-village random effects to account for village specific peculiarities in prevalence arising from unobserved village specific factors; and a spatial random effect to account for additional spatial autocorrelation. The spatial random effect was modelled as a Markov random field [52],[53] on a 53×44 grid with a 2 km×2 km pixel resolution. Non-linearity in age was modelled using a degree 3 P-spline, with second order random walk penalty against over fitting, on 20 equidistant knots [54]. Non-informative inverse gamma priors were assumed for variance components of the P-spline and random effects with hyper-parameters a = 0.001 and b = 0.001. Models with no random effects, village random effects only, spatial random effects only and both village and spatial random effects were compared with and without pasture and forest areal cover estimates. Buffer radii for the three land cover classes were fixed at the optimal values identified above. The parameter space of the Bayesian GAMMs was sampled using Markov Chain Monte Carlo (MCMC) techniques. For each model, MCMC was run for 1200000 iterations, discarding an initial burn-in period of 100000 iterations and sub-sampling every 10th sample thereafter. Model comparison was performed using the deviance information criterion (DIC) [55] with model selection based upon the model returning the lowest DIC.
A map of AE prevalence, corresponding to the age group for which the expected value of the P-spline was zero, was derived from posterior means of all other model parameters (except age), population means for individual-level risk factors and the areal cover layers. To avoid visual artifacts arising from the resolution differences of the land-cover data and the MRF, simple kriging [56] was used to interpolate MRF posterior means, evaluated at all pixel centroids and sampled villages, to a 60 m resolution grid. This interpolation used a linear semi-variogram model without a nugget and a fixed range of 60 km. Uncertainty in the spatial random effect was visualised by mapping the range of the 95% credibility interval of the MCMC samples of the MRF. As a model check, standardised village residuals were calculated according to [57] and a semi-variogram [58] was used to check for the absence of systematic spatial variation in model residuals. Geostatistical analyses were performed using the R implementation of the gstat package (http://www.gstat.org/s.html). Maps were plotted using the R function spplot.
Prevalence of alveolar echinococcosis in the studentless subset of data was 3.0% (2.45–3.66, 95% C.I.) (Table 1). Reported occupations among the 96 AE positive subjects were: 91 farmers; three housewives (aged 42,46 and 50); one cadre (aged 50); and one worker (aged 33). There were five cases in the under 30 age group (15, 18, 23, 23 and 25 years). Univariate analysis of the studentless sub-sample detected no significant difference in AE prevalence between farmers and non-farmers (χ2 = 0.0802, p = 0.777). Significantly higher AE prevalence was detected among Hui than among Han (χ2 = 6.78, p = 0.0092). AE prevalence was lower among subjects with access to tap or well water (χ2 = 5.89, p = 0.0152). Evidence of a sex difference in prevalence was weaker in the studentless subset (χ2 = 3.6, p = 0.057) than in the unfiltered subset (χ2 = 8.15, p = 0.0043). There was a sex bias in the student population with a greater number of students being male than expected under an equality null hypothesis (χ2 = 96.8, p<2.2×10−16). There was a sex bias in the farming population with a greater number of farmers being female than expected under equality conditions (χ2 = 115.4264, p<2.2×10−16). Dog ownership was reported more frequently among Han than Hui (χ2 = 107.3, p<2.2×10−16).
In interviews in 2002 local farmers reported that during the late 1970s valleys and lower slopes were generally used for agricultural crop production while upper slopes and hill tops were reserved for grazing. At that time there were no livestock restrictions. Twenty to fifty sheep per family was not uncommon and grazing pressure had been intense. Farmers also reported that the last patches of “forest” had been located on hill tops. However, these “forests”, by comparison with known forest patches in the southern Liupan, were not evident in the MSS imagery suggesting either small patch size or low tree density. Farmers reported that the number of sheep per household was capped in the late 1990's when incentives for converting grazing land to tree or shrub plantations were put in place.
The classified image is presented, with some summary statistics on percentage cover in Fig. 2. Two classes dominated the classification: agriculture and lowland pasture. The former was dominant in valleys and on hillsides while the latter was dominant on hill tops and in areas fringing the larger mountains. At higher elevations the Liupan, Yueliang and Nanhua mountains were dominated by mountain pasture and, particularly in the southern Liupan, harboured almost all the remaining forest in the area. Given the absence of archived land cover data, quantitative accuracy assessment was not possible. However, the classified image corresponded well with both field observations made in 2002 and anecdotal reports of historical land cover provided by local pastoralists and farmers. The mean normalised difference vegetation index, calculated from a Landsat MSS image acquired on June 1975, was 0.34 (s.d. = 0.13) and 0.14 (s.d. = 0.09) in mountain and lowland pastures respectively (t = 536.0, df = 152101.4, p<2.2×10−16).
The buffer radii maximising the likelihood of the GAM were found to be 6 km, 15 km and 18 km for lowland pasture, mountain pasture and forest cover respectively. Omitting forest cover resulted in optimal radii of 7 km and 15 km for lowland pasture and mountain pasture classes respectively. Models including the spatial random effect consistently returned lower DICs than their non-spatial counterparts (Table 2). The lowest DIC was returned by the model with all three areal cover components. However, this model was not selected for inference or prediction because of fears that the spatial relation between forest and sampling points was causing over fitting on a small subset of the data, namely the villages of the Liupan mountains in Guyuan and Longde counties closest to the largest forest patches and in which few AE cases were observed (Fig. 2). Village-level random effects failed to reduce the DIC of models containing a spatial random effect. Therefore model 2 in Table 2 was chosen for further inference and prediction.
Non-linear age-specific adjustments to the logit of prevalence are represented by the posterior mean and the 80% and 95% credibility intervals of the P-spline in Fig. 3. The fitted spline indicates a linear augmentation of prevalence on the logit scale in the 5–50 years age range, although uncertainty in the 5–20 years range was large. The linear trend plateaus at about 60 years. The expected value of the spline was closest to zero for 38 year olds. i.e. prevalence in that age category was representative of the average situation while prevalences in younger and older age groups were lower and higher than average respectively. Moreover, prevalence at 38 years could be predicted simply without adjustment for the non-linear age effect.
The mean, standard deviation, median and 95% credibility intervals of posterior samples for regression coefficients are shown in Table 3. The ranges of the 95% credibility intervals of posterior samples were strictly positive for dog ownership and lowland pasture and strictly negative for the intercept and mountain pasture. The 95% credibility intervals of coefficients of all other factors were not exclusive of zero. Posterior means and standard deviations of variance parameters for the P-spline and Markov random field were 0.0675 (s.d. 0.138) and 4.89 (s.d. 2.21) respectively.
Figure 4 presents the predicted prevalence among 38 year olds. The range of the 95% credibility intervals for each pixel of the spatial random effect is presented, with sample size, in Fig. 5. The principal hotspot was nested between the Liupan and Yueliang mountain ranges and lies within an area where the population was dominantly Hui (Fig. 4). Two lesser hotspots were also evident: the second, approximately 30 km west of the first, was situated in a dominantly Han area; and the third, situated between and a little south of the first and second, was in an ethnically mixed area. The lowest predicted prevalences were associated with the Liupan and Yueliang mountains (Fig. 4) despite relatively intense sampling in southwest Guyuan County (Fig. 5).
No AE cases were detected in students. Coefficient estimation for that subpopulation was neither necessary nor feasible, thus students were removed from further analyses. The remaining subsample consisted largely of farmers and no further statistical difference in AE prevalence between farmers and non-farmers was detected. Higher AE prevalence in Hui than in Han and among those without access to tap or well water was observed by univariate but not by multivariate statistics. These discrepancies probably arose from the non-random spatial distribution of the two ethnic groups (Fig. 4) and water source quality within the study area (Fig. 5). The spatial random effect apparently nullified the water source and ethnicity effects suggesting spatial heterogeneity in effect size. It is likely that a larger proportion of Hui than Han were infected by function of the spatial distribution of the two ethnic groups relative to areas ecologically favourable (by which we include water source quality) to E. multilocularis transmission. Redundancy of ethnicity in a spatial model and the fact that the high endemicity area in Zhang and the secondary hotspot here were both Han areas negates possible arguments of large between group genetic differences in susceptibility. Spatial heterogeneity in effect size could arise naturally for ecological reasons, e.g. interaction with spatially heterogeneous variables such as transmission intensity or density of E. multilocularis eggs in the environment. The role of unmeasured socio-economic factors in this interaction is impossible to assess here.
The only putative risk factor to correlate with AE prevalence in the Bayesian analysis was dog ownership. This corroborates previous studies suggesting that domestic dogs play an important role in the epidemiology of AE. Higher AE prevalence among females is frequently reported [59]. Both accelerated growth of Echinococcus cysts in immunosuppressed pregnant women [60],[61] and risk behaviour frequency (dog contact, farming,…) [10],[11] have been suggested as explanations, but the relative contributions of biological and sociological factors are unclear. Here, a sex difference was observed in the unfiltered dataset, but evidence for the sex difference became weak after students were removed from the analysis. The observed sex biases in the student and farming populations sufficiently explain the observed sex difference in AE prevalence.
Classical prevalence estimates obtained from small samples typically suffer low precision and are misleading. The village Xiping (UTM 572244, 3957984) gives a perfect example, a sample of just one individual provided a classical prevalence estimate of 100% (Figs. 4 and 5). One of the greatest strengths of spatial statistics over classical statistics is that spatial models utilise neighbouring information to tighten confidence intervals at each point. A good spatial model smooths out spuriously large variation arising from sampling without over-smoothing the true variation in the underlying phenomena being studied. The predicted prevalence surface presented here clearly achieves this goal, village-level prevalence estimates were restricted from the unrealistic range 0%–100% to a range comparable to other areas within the central Chinese endemic area (Fig. 4) [10],[29]. Over-smoothing appears to have been avoided given the size of the identified clusters relative to the spatial distribution of sampled villages. Estimated prevalence cannot be properly assessed without reference to a measure of uncertainty. There is a clear inverse relation between local sampling density and uncertainty (Fig. 5). Areas north and west of the Yueliang and near the Longde County - Gansu border are predicted by extrapolation of landscape indices to have high prevalence (Fig. 4). There is currently no data from these areas and follow up studies would help assess the validity of these extrapolations where uncertainties are large.
A hotspot located between the Liupan and Yueliang mountains was identified. This hotspot includes Nanwan village (UTM 583340, 3986433) where a large familial cluster of AE cases has been described [62]. Here we show that clustering has occurred at a higher organisational level than the family. The hotspot was approximately 10–15 km in diameter (Fig. 4). Further research is required to identify the causative factors of this hotspot. A direct effect of climate is an unlikely explanation since the hotspot is not within the mountains where E. multilocularis egg survival would be greatest. The relative roles of fox, dog and intermediate host densities and interactions with socio-economic factors must be identified. In a previous study, a family in which four in eight members were infected reported hunting Spermophilus for food and feeding uncooked viscera to their dogs [59]. Future studies must assess whether this practice was: unique to this family; common within the hotspot; or was wide spread across Xiji County.
The landscape analysis suggests that the environmental conditions favouring E. multilocularis transmission in Xiji differ from those favouring transmission in southern Gansu. Areal cover of mountain pasture around villages did correlate to AE prevalence, but with a negative coefficient. By contrast, abundance of the more degraded lowland pasture was associated with higher human AE prevalence. Note that these observed patterns depend upon the choices of scale that define the current sampling design. A hypothetical extension of the study area by an order of magnitude might result in contrary findings regarding the role of the Liupan mountain range. The greatest difficulty in interpreting this result arises from the time lags, which are unavoidable given the slow development of AE in humans, inherent in the current study. The lack of archived ground data makes it is hard to know exactly what constituted vegetation cover in “lowland pasture” areas. The available evidence comes from field observations (2001–03), image analysis of archived remote sensing data (1975 & 1978) and anecdotal accounts of local farmers. On this basis, “lowland pasture” most likely represents areas of grass, heavily grazed by sheep and goats and possibly interspersed with sub-pixel remnants of forest or shrub cover. Despite this uncertainty, one result is clear, the hypothesis of the current study fails to describe the Xiji endemic zone and small mammals of the Liupan grasslands were not the principal reservoirs of E. multilocularis linked to AE infection in humans. So were the principal intermediate host reservoirs in Xiji County different to those of Zhang where Microtus limnophilus was central to the eco-epidemiology of human AE?
A small mammal survey in 2003 recorded 16 species belonging to 7 families and identified five small mammal assemblages [63]. Forest, shrub and grasslands located in mountains provided greater diversity and lower trapping frequencies than lower elevation habitats. Microtus fortis and the wood-mouse Apodemus peninsulae were only trapped in forest and dense shrub where they dominated trapping results. Mountain grasslands, dominated by the lagomorph Ochotona huangensis, provided the highest densities of Apodemus agrarius and the jumping mouse Eozapus setchuanus was trapped uniquely there. Diversity was lowest and trapping frequency greatest in newly aforested set-aside where hamsters (Cricetulus longicaudatis) were dominant. This habitat provided the greatest trapping frequencies for Mus musculus, Ochotona daurica and the semi-desert jerboas Dipus sagitta and Allactaga sibirica. These species were also present in ploughed fields where the hamsters C. longicaudatus and Tscherskia triton and the zokor Eyospalax fronteria (previously Myospalax fontanierii), all known agricultural pests, were dominant. Dipus sagitta, A. sibirica and O. daurica were not trapped in areas of more advanced afforestation where hamsters and the Sciuridae Spermophilus alashanicus were dominant. Importantly, Arvicolidae were not trapped in large numbers in any habitat [63].
Susceptibility to E. multilocularis is undocumented for many of these species so inference must be made at higher levels biological organisation. Cyst fertility is poor in Apodemus [64] and Spermophilus [17],[65]. Fertile cysts in nine Eyospalax fronteria from the area contained few protoscoleces [40] and in Zhang E. fronteria was more abundant in lower prevalence agricultural areas [23],[27]. Allactaga elater has been found naturally infected in Azerbaidzhan [32]. The gerbil Meriones unguiculatus which has similar habitat preferences to jerboas has been reported in the area [40]. Gerbils [66],[67] and hamsters [26] provide excellent laboratory models for E. multilocularis and natural infections have been found in M. unguiculatus in NHAR [17] and in C. kamensis in Sichuan [unpub data]. Ochotona daurica has been found infected in Tuva, southern central Russia [17] and O. curzoniae is frequently present in fox faeces from NW Sichuan where it is predated preferentially, even when hamsters are visibly abundant [unpub data].
Each habitat in Xiji hosts potential E. multilocularis intermediate hosts. Foxes predate preferentially when a preferred prey species becomes readily available so high densities of preferred susceptible prey provide optimal conditions for E. multilocularis transmission [19]. Low densities and high diversity of small mammals in mountain habitats may explain the negative correlation between this habitat and AE prevalence. By contrast AE prevalence is rarely high in large agricultural expanses despite high densities of hamsters and zokors, suggesting low levels of predator prey interaction in these landscapes of intense human activity. In Xiji, AE risk was positively correlated with habitat best described as non-montane, non-agricultural with abundant bare soil, thin shrub cover and probably supporting large densities of hamsters, jerboas, O. daurica, zokors and mice, all of which may contribute to transmission.
The current work indicates that E. multilocularis can sustain transmission through small mammal communities that are not dominated by large cyclic populations of arvicolines. The meta-population dynamics of E. multilocularis across central China functions through a diversity of intermediate host communities. This diversity is largely unknown, the most up-to-date reference being [68], an atlas of the sylvatic mammals of China. But low spatial precision and a lack of information regarding dynamic patterns compromises its utility for modelling. It seems species level population data spanning large areas of China is not available, although such a dataset would be invaluable regarding the ecological management of small mammals and their related diseases.
During the 1990's rodenticides were applied liberally across much of central northern China [69]. Secondary poisoning destroyed dog populations in both Xiji and Zhang [11],[62],[13]. Many predator species disappeared [69] which would have inadvertently reduced transmission and may explain the absence of AE in children aged less than 15 years. Rodenticides are now heavily controlled and the domestic dog population is growing again, in part, courtesy of illegal dog trafficking with dogs from Tibetan areas being particularly appreciated. So the scene is set for a re-emergence of AE in Xiji and serology studies in school children have detected anti-E. multilocularis antibodies suggesting transmission is currently active in the area [70].
To conclude, landscape analyses indicate that mountain grasslands correlated negatively to AE prevalence in Xiji. This suggests transmission of E. multilocularis in Xiji, somewhat uniquely, functions principally through non-arvicoline species. The meta-population dynamics of E. multilocularis in central China functions across a diversity of eco-zones and small mammal communities which remain largely unknown. One principal and two lesser hotspots, each approximately 15 km in diameter, were identified. This spatial clustering appears to be caused by complex interactions between social and environmental factors that warrant further study. Spatial heterogeneity in water source quality apparently contributed, via interaction with other factors, to the observed distribution of AE. Sex biases in the student and farming populations appears to explain higher AE prevalence in females and dog ownership was a positive risk factor. Extrapolation of landscape trend terms identified three areas of above average prevalence, west of the southern Liupan and west and north of the Yueliang mountains, although uncertainty was large beyond the corpus of the sampling design. The prevalence and uncertainty maps represent the current state of the art regarding what is known of AE distribution in Xiji and provide an important baseline for future epidemiological monitoring and eco-epidemiological investigation. Future mass screening could focus on villages with a history of low quality water supply and should extend the study to: southern Haiyuan County where high prevalence is predicted; and southern Xiji County where uncertainties were large and a small number of cases have been detected despite low sample sizes. Extending the study area would also extend the range of environmental variation covered within the dataset which would be important regarding further data mining for ecological covariates and interactions.
|
10.1371/journal.ppat.1003051 | GABAergic Signaling Is Linked to a Hypermigratory Phenotype in Dendritic Cells Infected by Toxoplasma gondii | During acute infection in human and animal hosts, the obligate intracellular protozoan Toxoplasma gondii infects a variety of cell types, including leukocytes. Poised to respond to invading pathogens, dendritic cells (DC) may also be exploited by T. gondii for spread in the infected host. Here, we report that human and mouse myeloid DC possess functional γ-aminobutyric acid (GABA) receptors and the machinery for GABA biosynthesis and secretion. Shortly after T. gondii infection (genotypes I, II and III), DC responded with enhanced GABA secretion in vitro. We demonstrate that GABA activates GABAA receptor-mediated currents in T. gondii-infected DC, which exhibit a hypermigratory phenotype. Inhibition of GABA synthesis, transportation or GABAA receptor blockade in T. gondii-infected DC resulted in impaired transmigration capacity, motility and chemotactic response to CCL19 in vitro. Moreover, exogenous GABA or supernatant from infected DC restored the migration of infected DC in vitro. In a mouse model of toxoplasmosis, adoptive transfer of infected DC pre-treated with GABAergic inhibitors reduced parasite dissemination and parasite loads in target organs, e.g. the central nervous system. Altogether, we provide evidence that GABAergic signaling modulates the migratory properties of DC and that T. gondii likely makes use of this pathway for dissemination. The findings unveil that GABA, the principal inhibitory neurotransmitter in the brain, has activation functions in the immune system that may be hijacked by intracellular pathogens.
| Toxoplasma gondii is an obligate intracellular protozoan parasite and an important food- and water-borne human and veterinary pathogen. Toxoplasmosis is normally self-limiting but severe manifestations occur upon congenital transmission to the developing fetus or during infection in immune-compromised individuals. Toxoplasma invades a variety of cell types and mounting evidence shows that certain white blood cells, e.g. dendritic cells, can shuttle parasites in the infected host by a Trojan horse type of mechanism. Dendritic cells are considered the gatekeepers of the immune system but can, paradoxically, also mediate dissemination of the parasite. Previous work has shown that Toxoplasma induces a hypermigratory state in dendritic cells when they become infected. Here, we show that, shortly after infection by the parasite, dendritic cells start secreting γ-aminobutyric acid (GABA), also known as the major inhibitory neurotransmitter in the brain. We show that dendritic cells express GABA receptors, as well as the machinery to synthesize and transport GABA. When GABA synthesis, transport or receptor function was inhibited, the migration of infected dendritic cells was impaired. In a mouse model of toxoplasmosis, treatment of infected dendritic cells with GABA inhibitors resulted in reduced propagation of the parasite. This study establishes that GABAergic signaling modulates the migratory properties of dendritic cells and that the intracellular pathogen Toxoplasma gondii sequesters the GABAergic signaling of dendritic cells to assure propagation.
| Toxoplasma gondii is an obligate intracellular parasite that infects warm-blooded vertebrates. It infects approximately 25% of the global human population [1]. Initial infection occurs orally or congenitally, whereby the formed tachyzoite stages disseminate widely in the organism. Although principally asymptomatic in humans, infection can cause severe neurological complications in immune-compromised individuals, disseminated congenital infections in the developing fetus, and ocular manifestations in otherwise healthy individuals [1]. T. gondii enters host cells by active penetration, a rapid process that is dependent on the actin-myosin cytoskeleton of the parasite, and does not rely on the host cell machinery for uptake [2]. T. gondii can invade and multiply inside any nucleated cell type, including blood leukocytes, and a preference to infect myeloid leukocytes in vitro has been reported [3]. Following primary infection, T. gondii strikes a fine balance between eliciting an effective immune response and establishing a silent, life-long infection [4]–[6]. Acute infection triggers a robust Th1 polarized immune response with efficient activation of antigen presenting cells, including dendritic cells (DC) [7], [8].
DC are a fundamental component of the immune response but also a putative gate to immune evasion and persistence for pathogens [9]. DC serve as sensors in peripheral tissues that allow processing and presentation of antigens for initiation of adaptive immune responses and pathogen clearance. The mechanisms underlying DC migration are complex and the molecular traffic signals that govern DC migration are not fully understood [10]. One of the hallmarks of mature DC is the expression of the C-C chemokine receptor 7 (CCR7). Binding to its ligands (CCL19 and CCL21) guides the migrating cells to the lymph nodes where adaptive immune response is initiated [11]. In order to avoid clearance by the immune system, intracellular parasites, bacteria, fungi and virus have evolved diverse strategies to subvert this central function of DC [9], [12].
Mounting evidence indicates that DC play a pivotal role during T. gondii infection as mediators of essential immune responses [8], [13] and as parasite carriers that facilitate the dissemination of the infection [14]–[17]. In this context, T. gondii induces a hypermotility state in infected DC that contributes to parasite dissemination in vivo [14], [15]. Interestingly, this strategy for dissemination appears to be conserved among other members of the Apicomplexan parasite family, e.g. Neospora caninum [18]. Yet, the molecular mechanism controlling the parasite-induced hypermigratory phenotype in DC remains unknown. Given its characteristics, i.e. random directional hypermotility in absence of chemotactic cues, alternative/non-classical pathways are likely to be involved [4].
γ-aminobutyric acid (GABA) is one of the major neurotransmitters in the CNS [19], acting via activation of GABAA receptors [20] and to a lesser extent GABAB receptors [21]. GABA is shuttled in and out of cells via GABA transporters (GAT) of the solute carrier family 6 [22]. GABAergic cells synthesize GABA via glutamate decarboxylases (GAD) [23]. In contrast to its role as an inhibitory neurotransmitter, GABA plays an excitatory role during neuronal development [24], [25]. In fact, mounting evidence indicates that neurotransmitters, including GABA, have a motogenic function and participate outside the CNS in diverse functions including cell migration, immunomodulation, and metastasis [26], [27]. GABA, its synthesis enzymes GAD, GABA receptors and transporters have been found in a variety of tissues outside the CNS, such as the pancreatic islets and testes [28], [29].
Using in vitro models and in vivo bioluminescence imaging (BLI) in a mouse model of toxoplasmosis, we demonstrate that DC are GABAergic cells and that GABA modulates the hypermigratory phenotype observed in Toxoplasma-infected DC. During in vivo infections, the GABAergic system of infected DC is likely used to facilitate parasite dissemination.
To address the GABAergic response of mouse DC upon infection, GABA was quantified in the cell supernatant. Challenge of DC with freshly egressed T. gondii tachyzoites led to a significant increase of GABA in the supernatant, while heat inactivated parasites, parasite lysate or LPS did not increase GABA secretion relative to non-infected DC (Figure 1A). Moreover, secretion of GABA from DC challenged with freshly egressed tachyzoites rapidly increased over time, even prior to parasite replication, and augmented over 24 h (Figure 1B). In contrast, the GABA-precursor glutamate exhibited a modest transient increase in the supernatant following infection, which was redundant by 24 h (Figure S1). We next assessed if GABA secretion was induced in infected DC or uninfected bystander DC. GABA secretion rapidly augmented with MOI over time (Figure 1C) and supernatants from infected DC did not induce significant GABA secretion in DC (Figure 1D). Moreover, fluorescence-activated cell sorting of DC populations challenged with GFP-expressing T. gondii showed that GABA secretion occurred essentially in GFP+ cells (Figure 1E). Altogether, this shows that the observed elevation of GABA secretion emanates from infected DC and that GABA secretion of by-stander DC and DC in complete medium (CM) are similar. Next, 9 human donors were assessed. Monocyte-derived DC from all donors responded with increased amounts of GABA upon T. gondii infection and variability in the secreted levels of GABA was observed among the donors (Figure 1F). Monocytes challenged with T. gondii also exhibited an increase in GABA secretion (Figure S2A). Representative strains from the three predominant T. gondii genotypes (I, II and III) induced GABA secretion in infected DC (Figure S3). We conclude that upon Toxoplasma-infection, mouse and human myeloid DC exhibit elevated levels of GABA secretion.
In an effort to ascertain which GABAA receptor subunits are expressed in mouse DC, we screened the 19 subunits expression profiles in mouse DC and astrocytes. We detected GABAAR α3, α5, β1, β3, and ρ1 subunit transcripts in DC, whilst 12 different subunits were detected in primary astrocytes (Table 1, Table S1 for primer sequences). We decided to quantify differential gene transcription in mouse DC following T. gondii infection using α3, β3 and ρ1, the most strongly expressed subunits in non-infected mouse DC (Table 1). The transcript level analysis, using template from infected DC and non-infected DC, showed an up-regulation for the α3 and ρ1 transcripts after 2 h infection and down-regulation by 8 h. A down-regulation was observed for the β3 subunit at both time-points (Figure 2A). In addition, immunocytochemical stainings indicated expression of the β3 subunit in DC in CM (Figure 2B) and in Toxoplasma-infected DC (Figure 2C). In DC suspensions challenged with T. gondii, similar staining patterns were observed in infected and non-infected DC (Figure S4).
We next examined functional expression of the GABAA channels in DC using the whole-cell patch-clamp technique. We recorded currents from human and mouse DC infected for 12 h with T. gondii (Figure 2D, E, F, H) and non-infected DC (Figure 2G). At a negative holding potential (- 80 mV) in symmetrical chloride solutions, 1 µM GABA application to the cells resulted in an inward current that ranged widely in magnitude. In mouse and human DC the peak-current value ranged from −9 pA to −9.9 nA (n = 7) and −38 pA to −7.7 nA (n = 7), respectively. The currents reversed at positive holding potential (Figure 2F, +40 mV) and were inhibited by the GABAA competitive antagonist SR-95531 (Figure 2H). We conclude that human and mouse myeloid DC express functional GABAA receptors and that GABA can induce membrane currents in Toxoplasma-infected DC.
To investigate the effects of Toxoplasma infection on the GABAergic system, expression levels of the GABA transporter GAT4 and the GABA synthesizing enzymes, GAD65 and GAD67, were assessed in DC. A rapid induction of GAT4 transcription was observed shortly after infection (Figure 3A). In contrast, expression of GAD65 was detected in both non-infected and infected DC at similar levels (Figure 3A), whereas GAD67 expression was not detectable in either group (data not shown). Moreover, addition of GAD inhibitor (SC) and GAT4 inhibitor (SNAP) to infected DC nearly abolished or significantly reduced, respectively, the secreted levels of GABA in the supernatant (Figure 3B). Inhibitor treatments did not significantly affect intracellular parasite replication in vitro (Figure S5) or the GABA signal detected in complete medium containing extracellular parasites (Figure S6). We next assessed the impact of the GABAergic inhibitors on the transmigration of infected DC. Both GAT4 (SNAP) and GAD inhibition (SC) had a significant inhibitory effect on the transmigration of infected DC, and transmigration was significantly restored following incubation of the inhibitor-treated cells in supernatant from infected DC cultures (Figure 3C). In contrast, monocytes did not exhibit this migratory phenotype observed in human and murine DC [14] but GABAergic inhibition significantly reduced the transmigration of non-infected monocytes (Figure S2B). Furthermore, the transmigration phenotype of DC was either fully (SNAP, GAT4 inhibitor) or partially (SC, GAD inhibitor) restored after addition of exogenous GABA (Figure 3D). In line with this, GABAA receptor antagonist, and to a lesser extent GABAB receptor antagonist, significantly reduced transmigration of infected DC (Figure 3E). Notably, GABAA and GABAB receptor agonists did not enhance transmigration of non-infected or infected DC. Thus, GABA per se was not sufficient to induce transmigration of non-infected DC but could restore transmigration in infected DC impaired in GABA production or transportation. Altogether, these data implicate GABA synthesis, transportation and receptor activity in parasite-induced transmigration of DC in vitro.
To determine whether the GABAergic system also affected DC motility and chemotaxis, infected and non-infected DC were allowed to migrate along a concentration gradient of CCL19 in a chemotaxis chamber system. Non-infected DC exhibited a low level of random directional motility in absence or presence of chemokine and LPS-stimulation of non-infected DC resulted in a distinct directional migration towards CCL19 (Figure 4A). In contrast, Toxoplasma-infected DC exhibited a dramatically enhanced random directional motility in absence of chemokine (Figure 4B), with a significant increase in velocity compared to non-infected DC (Figure 4C). Interestingly, directionality towards CCL19 was observed for infected DC, similar to that observed upon LPS maturation (Figure 4B). It is also notable that Toxoplasma-infected DC ((−) chemokine, Fig. 4 B) outranged LPS-matured non-infected DC ((+) chemokine, Figure 4A) in migrated distances and velocity (Figure 4A, B, C). We conclude that Toxoplasma-infected DC exhibit a hypermigratory phenotype in vitro and that hypermotile Toxoplasma-infected DC maintain the ability to chemotax in vitro.
Next, we determined whether targeting the GABAergic system affected the migratory and chemotactic responsiveness of non-infected DC (Figure 5A, B) and Toxoplasma-infected DC (Figure 5D, E) in vitro. Overall, inhibition of GABA synthesis (SC, GAD inhibitor) or GABA transport (SNAP, GAT4 inhibitor) led to a significant decrease in the velocity and the accumulated distance covered by DC (Figure 5C, F). Interestingly, the ability to respond with directionality towards CCL19 was not abolished by inhibiting GABA transport or synthesis but, as a consequence of the reduction in velocity, the overall chemotactic response was diminished (Figure 5). No significant influence of a GABA gradient on the directionality of DC motility was observed for non-infected and infected DC (data not shown). In summary, present data show that inhibition of the GABAergic signaling system significantly reduces the velocity of infected DC in vitro and thereby the magnitude of the chemotactic response in vitro.
We next assessed the relative expression of the CCL19 ligand CCR7 on human and mouse DC by flow cytometry. First, the chemotactic responses observed with mouse DC were confirmed using human monocyte-derived DC (Figure 6A). Additionally, monitoring of infected and uninfected DC in suspensions challenged with T. gondii showed that the chemotactic response occurred preferentially in the infected (RFP+) DC population (Figure 6A, central panel). In line with this result, DC challenged with T. gondii or treated with LPS exhibited a relatively higher expression of CCR7 compared to DC in complete medium (Figure 6B). The analyses of DC populations challenged with T. gondii showed that upregulation of CCR7 occurred essentially in infected (RFP+) DC (Figure 6B, central panel). An upregulation of CCR7 was consistently observed in infected DC from 7 different human donors (Figure 6C). For mouse DC, a small but significant upregulation of CCR7 was observed in infected DC (Figure 6E). In the presence of GABAergic inhibitors (SC, SNAP), overall non-significant effects on CCR7 expression were observed (Figure 6D and E). We also assessed the effects of GABAergic inhibitors (SC, SNAP) on the expression of co-stimulatory molecules and maturation. Overall, no distinct or modest effects were observed by GABAergic inhibition (Figure S7). Altogether, we conclude that upon T. gondii-infection, DC exhibit a relative up-regulation of CCR7 that is consistent with the observed chemotactic responses in vitro.
Previously we have demonstrated that the adoptive transfer of T. gondii-infected DC leads to rapid dissemination of parasites as well as exacerbation of infection compared to infection with free parasites [14]. To assess whether GABAergic inhibition of T. gondii-infected DC had an impact on the aforementioned in vivo dissemination, mice were inoculated i.p. with freshly egressed luciferase-expressing tachyzoites or with tachyzoite-infected DC. Photonic emissions were measured by BLI daily for 5 days [30]. Infected DC were pretreated with a combination of inhibitors against GAD and GAT4 shown to have prolonged (24 h) inhibition on transmigration of DC in vitro. Interestingly, GABAergic inhibition of infected DC resulted in a significant reduction (∼2.8 fold) in total parasite photonic counts compared to non-treated infected DC by day 4 post infection (P = 0.0003, GLM ANOVA, Figure 7A and B). Furthermore, the photonic counts from the combination treated group were equivalent to levels observed during free tachyzoite infection (P>0.05, GLM ANOVA, Figure 7A and B). Analyses of adoptively transferred uninfected DC pretreated with GABAergic inhibitors showed similar numbers of treated and non-treated DC in the spleen and the peritoneum (Figure S8).
To determine the presence of parasites in different organs, photonic emissions were assessed ex vivo in the spleen, MLN and brain (Figure 7C). Special assessment of the brain showed significant differences in photonic emissions on days 1–2 with important variability between mice (Figure 7D). To quantify parasitic loads in target organs, plaquing assays were performed. Overall, higher parasitic loads were observed in mice challenged with non-treated infected DC compared to mice challenged with infected DC treated with GABAergic inhibitors (Figure 7E). Altogether, and in line with observations in vitro, this indicates that treatment of infected DC with GAD- and GAT4-inhibitor (SC, SNAP) results in a significant reduction in the dissemination of T. gondii, and subsequently a reduction of the parasitic loads during the course of infection in mice.
In the present study, we report that GABAergic signaling is closely linked to a hypermigratory phenotype in DC, which is induced by T. gondii infection [14]. Furthermore, we demonstrate that mouse and human myeloid DC possess functional GABAA receptors and are capable of producing and secreting GABA. Interestingly, challenge of DC with T. gondii consistently resulted in a significant increase in the levels of extracellular GABA over time in mouse DC and in DC derived from different human donors. The secretion of GABA was not related to DC activation or maturation following exposure to LPS, parasite lysates, supernatants from infected DC or uptake of heat-inactivated parasites, but linked to the live infection by T. gondii. Our data indicates that DC secrete GABA as a consequence of infection by the parasite and that non-infected by-stander DC only provide a minor contribution to the total secreted amounts. Also, the absence of a distinct modulation on the secreted levels of the GABA-precursor glutamate is indicative of a selective effect on GABA synthesis and secretion.
We found that a determined subset of GABAA receptors subunit genes was transcribed in mouse DC (α3, α5, β1, β3 and ρ1) in contrast to the broader expression in astrocytes. This finding is consistent with the concept that most GABAA receptor pentamers are composed of at least two α- and two β-type subunits while the final subunit type may vary [31]. This also strongly suggests that genetic control plays a major role in the choice of transcribed subunit variants. In fact, such differential expression has been implicated in the changes in responsiveness and function of GABAA receptors [32], [33]. Moreover, factors that affect GABAergic signaling, e.g. infection, may simultaneously induce up- and down-regulation of specific GABAA subunits through epigenetic mechanisms [34].
Here, we demonstrate for the first time that GABA evokes GABAA receptor-mediated currents in T. gondii-infected DC and in non-infected DC. While transcript levels of the α3, β3 and ρ1 subunits were modulated upon infection, functional patch-clamping data indicates that GABAA receptors are constitutively expressed in DC. In line with this, immunocytochemical analyses indicated expression of the GABAA receptor β3 subunit in infected and non-infected DC. As individual receptor subunits do not necessarily reflect the number of functional receptor pentamers or combinations of pentamers that are expressed in a particular cell, it remains unknown how individual subunits relate to the receptor function in DC. Thus, further studies are needed to characterize and quantify the precise receptor subset composition of immune cells, and whether sensitivity to GABA is modulated upon infection.
Recently, human monocytes were shown to express the GABAA β2 subunit and functional GABAA receptors were described in a human myelomonocytic cell line (α4, β2, γ1 and δ subunits) [35]. In contrast to DC, monocytes do not exhibit enhanced transmigration upon T. gondii-infection in vitro [36]. Here, we report that monocytes respond with GABA secretion upon T. gondii infection. Altogether, this raises the questions if receptor activation through secreted GABA, e.g. an autocrine loop, is needed for migratory activation or if different subsets of receptors are expressed in DC compared to monocytes. Whether these intriguing phenotypic differences between monocytes and DC depend on GABAA receptor expression levels, functional receptor subunit composition or capacity to rapidly secrete GABA awaits further investigation.
Induction of GABA secretion in infected DC was confirmed in strains from the three predominant genotypes of T. gondii, but it is unlikely that the maintenance and expression of GABA receptors and the GABAergic system in DC are exclusively a result of evolutionary pressure from T. gondii. Thus, additional functions for the GABAergic system in DC are likely to be discovered.
Mounting evidence indicates that T. gondii modulates the host's pathways for cell migration to facilitate its dissemination and establishment of a chronic infection. Our studies demonstrate that targeting the GABAergic machinery of host DC, i.e. GABA biosynthesis, transport, or ligand channel activation in vitro resulted in impaired ability to transmigrate, most prominently using inhibitors targeting host cell GAD and GAT4. Furthermore, the motility of treated infected DC was reduced to levels comparable to non-infected DC. In contrast, targeting the GABAergic system did not abrogate the ability of DC to respond with directionality in a CCL19 chemokine gradient but significantly reduced the speed and migrated distances of infected DC, thus reducing the chemotactic response in vitro. Further, in absence of a chemokine gradient, the motility and velocity of DC was reduced. This suggests that GABA and the GABAergic system may primarily interfere with the mechanisms of cell motility. In fact, GABA has been shown to promote the metastasis of cancer cells [37] and increase the velocity of human sperm motility via activation of GABAA and GABAB receptors [38]. Also, our in vitro chemotaxis data using GABA as a chemoattractant suggests that GABA alone is not sufficient to mitigate directional DC migration, irrespective of infection status or LPS maturation.
Notably, during T. gondii infections CCR7 and CCL19 are up-regulated [39], [40] and soluble factors can trigger upregulation of CCR7 in DC [41]. We have previously shown that Toxoplasma-induced hypermotility of DC in vitro occurs in absence of chemotactic cues and does not depend on CCR7-, CCR5- or MyD88-activation [14]. Here, we report that hypermotile T. gondii-infected human and mouse DC respond to chemotactic cues (CCL19) in vitro and significantly up-regulate CCR7. In neurons, reports indicate that GABAergic chemokinetic signaling cooperates with other chemotactic cues for embryonic cell migration [42]. In line with this, GABAergic inhibition has been shown to down-modulate the chemotactic responses of monocytes and neutrophils [35], [43]. Hypothetically, we propose that the two mechanisms tested here, GABA/GABAA receptor-mediated hypermotility and CCR7-mediated chemotaxis, could be acting simultaneously or even synergistically in Toxoplasma-infected DC, thereby enhancing DC motility and potentiating the dissemination of parasites infecting DC. Also, possible synergistic effects with activation of multiple chemokine receptors or downstream signaling [14], [40], [44] need to be tested in vivo in future research, especially in the context of the remarkably efficient passage of the parasite across restrictive biological barriers.
Present data are consistent with the notion that T. gondii is able to subvert the regulation of host cell motility and exploits the host's natural pathways of cellular migration for parasite dissemination [4]. In line with the above, GABAergic inhibition in T. gondii-infected DC adoptively transferred to mice abolished the disseminatory advantage provided by ‘shuttling’ DC [14] and resulted in a reduced parasite burden over the course of infection in mice. Although the overall trend was that GABAergic inhibition yielded generally lower parasitic loads in organs, it is particularly interesting that significantly lower parasitic loads were observed in the brain. Recent reports have elucidated that the processes leading to DC migration into the brain parenchyma in the context of toxoplasmic encephalitis are complex and include signaling through multiple chemokine receptors [40]. The present study does not address the passage of the parasite across the blood brain barrier and whether the infected DC directly transport parasites into the CNS. Alternative but not mutually exclusive hypotheses are possible for the observed differences in parasitic loads in the brain. First, GABAergic inhibition of DC, similar to pertussis toxin treatment [14], [40], may reduce the number of DC that reach the CNS microvasculature. Second, intracellular localization, e.g. in migratory DC, may offer a safe intracellular niche per se for targeted delivery to organs [4], [18]. Third, differential transfer of the parasite to other immune cells, e.g. NK, T cells, during infection may modulate the dissemination and passage of T. gondii and of infected leukocytes to the CNS [45]–[47]. The relative contribution of these processes to the passage of T. gondii across the blood brain barrier remains unknown. Nonetheless, adoptive transfers with infected DC led to higher parasitic loads in the CNS in models of toxoplasmosis and neosporosis [14], [16], [18], while adoptive transfer of infected macrophages and lymphocytes did not [36]. Different results have also been reported on the contribution of CD11b+/monocytic cells [16], [48]. Additionally, these processes need to be tested in the context of oral natural infections in future research.
On the technical note, bioluminescence imaging of the brain ex vivo failed to detect significant differences between groups treated with GABAergic inhibitors and untreated groups beyond day 2 post-infection. This could be due to the overall low parasitic loads in the brain and to the relatively high variations between individual mice, as demonstrated by plaquing assays. It also indicates that plaquing assays remain the superior method in detecting and quantifying viable parasitic loads in the CNS [30].
The effects of GABA on the immune system have not been extensively studied. In human blood, GABAA receptor subunits haven been detected in CD4+ and CD8+ T cells, B cells and certain monocytes [35], [49]. In rodents, functional GABAA receptors have been described in T cells [50] and peritoneal macrophages [51]. Activation of GABAA receptors has been shown to inhibit T cell proliferation [50], [52], [53] and autoimmune inflammation [51], [54]. Thus, it is conceivable that GABA release by T. gondii-infected DC may modulate DC activation to prevent T-cell proliferation during the early phase of infection.
The ‘inflammatory reflex’ is a concept gaining ground in the field of immunology in demonstrating that neurotransmitters can interact and influence immune cell function [55]. This is an interesting perspective given the psychobehavioural impact of chronic Toxoplasma infection in humans and rodents [56]. Also similar to DC, microglia respond with a migratory phenotype upon T. gondii infection [47] and their migratory behavior in the CNS could be modulated by ambient GABA. In fact GABA has been shown to suppress the reactivity of microglia [57], leading to attenuation of IL-6 and IL-12 responses [58]. Furthermore, GABA levels in the CNS range from submicromolar outside of synapses to millimolar concentrations in active synapses [59], [60] whilst GABA is present in peripheral tissues at around 100 nM [61]. Here, the current responses to GABA by DC exhibited characteristics of neuronal synaptic and extrasynaptic GABA-activated currents [62]. The synaptic-like currents responded rapidly and then decayed whereas the extrasynaptic-like currents activated with a delay and maintained low current amplitudes for an extended period. The finding that 1 µM GABA concentrations gave rise to robust transient phasic and tonic currents confirms that physiological GABA concentrations may be sufficient to activate both types of currents in DC. This also raises the prospect of GABA acting as a chemoattractant to mitigate ‘homing’ of the infected leukocytes to the CNS and as a possible modulator of microglia-mediated parasite dissemination in the CNS [47]. Human cord blood-derived hematopoietic and progenitor cells have been reported to migrate towards a GABA gradient [63]. In contrast, our in vitro chemotaxis data using GABA as a chemoattractant suggests that GABA alone is not sufficient to mitigate directional migration of T. gondii-infected DC in vitro.
A recently proposed model envisages a complex interplay between ambient GABA, GABAA receptor activation, and chloride transport as regulators of interneuron migration [64]. It is conceivable that the DC GABAergic system may be working in a similar manner in relation to the hypermigratory phenotype exhibited by T. gondii-infected DC. Our model comprises: 1) DC invasion by Toxoplasma, resulting in increased GABA production; 2) GABA is shuttled out of the cell by GABA transporters, leading to an autocrine effect in activating GABAA receptors; 3) chloride ion efflux by GABAA receptor channels and subsequent calcium influx maintain DC in a depolarizing migratory state. The effector mechanisms leading to increased GABA production in DC as a consequence of T. gondii-infection await further investigation. It has been shown that T. gondii-infection can lead to extensive transcriptional regulation of host genes in DC [65] and a modulated transcription of GAT4 and of GABAA α3, β3 and ρ1 subunits was observed in infected DC. We found no evidence of significant production of GABA by extracellular parasites but a modulation of the DC biosynthesis of GABA by intracellular parasites cannot be excluded.
In summary, we provide substantial evidence that the DC GABAergic system plays a significant role in the maintenance of the T. gondii-induced hypermigratory phenotype observed in infected DC. To the best of our knowledge, this constitutes the first report showing that the GABAergic system can be utilized by an intracellular pathogen to modulate host cell motility and potentiate systemic dissemination. It remains to be seen whether other pathogens also utilize the GABAergic system to facilitate the establishment of an infection. Further investigation of the specific molecules and pathways involved will enable a greater understanding of the diverse roles that the GABAergic system may play outside the CNS.
All protocols involving animals were approved by the Regional Animal Research Ethical Board, Stockholm, Sweden, following proceedings described in EU legislation (Council Directive 86/609/EEC). The Regional Ethics Committee, Stockholm, Sweden, approved protocols involving human cells. All donors received written and oral information upon donation of blood at the Karolinska University Hospital. Written consent was obtained for utilization of white blood cells for research purposes. The ethics committees approved this consent procedure.
Tachyzoites from the green fluorescence protein (GFP) and luciferase-expressing T. gondii line PTGluc (type II, cloned from ME49/PTG-GFPS65T) [30], RH-LDMluc (Type I, cloned from RH-GFPS65T) [30], CTGluc (type III) [66] and PRU-RFP [67] were maintained by serial 2-days passage in human foreskin fibroblast (HFF) monolayers. HFFs were propagated in Dulbecco's modified Eagle's medium (DMEM; Invitrogen) with 10% fetal bovine serum (FBS), gentamicin (20 µg/ml, Gibco), glutamine (2 mM, Gibco) and HEPES (0.01 M, Gibco) referred to as complete medium (CM).
C57BL/6 mice (6–10 weeks old) were maintained and bred at the animal facility of the department of Microbiology, Tumor and Cell Biology, Karolinska Institutet (Stockholm, Sweden). For bioluminescence assays, male BALB/c mice (6–8 weeks old) were purchased from Charles River (Sulzfeld, Germany) and maintained under pathogen-free conditions.
Mouse bone marrow-derived DC were generated as described [68]. Briefly, cells from bone marrow of C57BL/6 mice were grown in CM containing 20% supernatant from the GM-CSF-secreting cell line X63 or 10 ng/ml recombinant mouse GM-CSF (Peprotech). Loosely adherent cells were harvested on day 6. To generate human monocyte-derived DC, buffy coats obtained from healthy blood donors at the Karolinska University Hospital Blood Center were treated with 1 ml RosetteSep (StemCell Technologies) per 15 ml of buffy coat, followed by centrifugation on Lymphoprep (Axis.Shield PoC AS) gradients. The population, defined as monocytes, exhibited CD14+ (DakoCytomation) and <1% CD3+/19+ (BD Biosciences) as evaluated by flow cytometry (FACSCalibur, BD Biosciences). DC were generated as described previously [69]. Briefly, purified cells were cultured 7 days in CM supplemented with 100 ng/ml GM-CSF (Peprotech) and 12.5 ng/ml IL-4 (R&D Systems). DC were typified by expression of CD1a, CD11b, CD14 (DakoCytomation), CD80, CD83, CD86, HLA-DR, CD11a, CD18 (BD Biosciences). Primary astrocytes were generated from cortices from 1–3 day-old C57BL/6 mice as previously described [47].
Human monocyte-derived DC were cultured on poly-L-lysine-coated glass coverslips (Sigma) and challenged with freshly egressed tachyzoites (PTGluc, MOI 1) for 16 h. After fixation (3% paraformaldehyde; Sigma) and blockade (5% FBS; Gibco), the cells were stained with mouse-anti human β3 (NeuroMab; clone N87/25; UC Davis/NIH NeuroMab Facility; 1∶250). Anti-mouse Alexa Fluor-conjugate (Invitrogen) was used as secondary antibody. Slides were mounted using VectaShield with DAPI (Vector Laboratories) and assessed by epifluorescence microscopy (Leica DMRB).
DC (3×106 cells) were collected and centrifuged for 2 min at 100 g. The supernatant was removed, the pellet washed with the extracellular solution, and centrifuged for 2 min at 100 g. The pellet was then resuspended using 100 µl of the extracellular solution. Nanion's Port-a-Patch chip technology (Nanion, Germany) was used to voltage-clamp the cells. A cell suspension of 5 µl was dispensed into the extracellular chamber containing the recording chip of 2–3.5 MΩ resistance. The whole-cell configuration was established and currents were recorded at holding potential of −80 mV. GABA (1 µM and 1 mM) or GABA plus 100 µM SR95531 (GABAA antagonist, Sigma-Aldrich) were prepared with the extracellular recording solution and perfused into the extracellular chamber at a rate of 1 ml/min. Extracellular solution in mM: 145 NaCl , 5 KCl, 1 MgCl2, 1.8 CaCl2 , 10 TES (pH 7.3) and 297 mOsm/kg. Internal recording solution in mM: 50 CsCl , 10 NaCl, 60 Cs-Fluoride, 20 EGTA, 10 TES (pH 7.3) and 284 mOsm/kg. Extracellular recording solution in mM: 80 NaCl, 3 KCl, 10 MgCl2, 35 CaCl2, 10 HEPES (pH 7.3) and 296 mOsm/kg. All patch-clamp recordings were performed at room temperature (20–22°C). Patch-clamp recordings were done using an Axopatch 200B amplifier, filtered at 2 kHz, digitized on-line at 10 kHz using an analogue-to-digital converter and analyzed with pClamp software (Molecular Devices, USA).
Total RNA was extracted using TRIzol reagent (Invitrogen). First-strand cDNA was synthesized using Superscript II Reverse Transcriptase (Invitrogen). All primers were initially screened for signal detection using cDNA and conventional PCR. Following sequence confirmation, gene transcription levels were monitored. Using the SYBR Green Master Mix (Applied Biosystems) PCR reactions were carried out on the Applied Biosystems 7500 Fast Real-Time PCR System using the following program: 10 min holding at 95°C, 40 cycles of 95°C for 15 s, 60°C for 1 min. A dissociation stage (95°C for 15 s, 60°C for 1 min, 95°C for 15 s) was added to generate a melting curve for data analysis. Glyceraldehyde 3-phosphate dehydrogenase (GAPDH) was used as the reference gene. The data was analyzed using Sequence Detection Software v.1.3.1 (Applied Biosystems) and fold changes in expression were calculated using the comparative ΔCT method against the non-infected DC control. Primers against the 19 GABAAR subunits (Table S1) were designed using OligoPerfect Designer software (Invitrogen) and Primer-BLAST (http://www.ncbi.nlm.nih.gov/tools/primer-blast/), and validated using OligoAnalyzer (http://eu.idtdna.com/analyzer/Applications/OligoAnalyzer). The primers were purchased from Invitrogen or Qiagen.
DC were plated at a density of 1×106 cells per well and incubated with freshly egressed T. gondii tachyzoites (MOI 1) or with soluble extracts and reagents in a total volume of 700 µl CM for 24 h. Freshly egressed tachyzoites were heat-inactivated at 56°C for 30 min. Parasite lysates were obtained by harvesting supernatants after sonicating 1×106 tachyzoites for three 30 s pulses (Soniprep 150). Reagents were purchased from Sigma-Aldrich unless stated otherwise, and added at the following final concentrations: Semicarbazide (SC, 50 µM); (S)-SNAP-5114 (SNAP; 50 µM); LPS (100 ng/ml). The cells were centrifuged at 4000 g for 2 min, and the medium supernatant collected and stored at -70°C until analysis. Samples were extracted, derivatized, and incubated with antiserum according to manufactures protocol (GABA or Glutamate Research ELISA kits, Labor Diagnostica Nord, Nordhorn, Germany). GABA and glutamate concentrations were quantitatively determined by ELISA, monitored at 450 nm (Multiskan EX, Labsystem Oy, Finland).
DC and monocytes were plated at a density of 1×106 cells per well and incubated with freshly egressed T. gondii tachyzoites (at indicated MOI) or with soluble extracts and reagents in CM for 6 h. Cells were then gently transferred into triplicate transwell filters (8 µm pore size; BD) at a density of 2×105 per well, and incubated as indicated. Reagents were added at the following final concentrations: Muscimol (300 µM); Bicuculline (50 µM); Baclofen (500 µM); CGP35348 (500 µM); Semicarbazide (50 µM); SNAP 5114 (50 µM); LPS (100 ng/ml); GABA (0.5 µM). When indicated, parasite viability and replication was assessed by plaquing assays on HFF monolayers and vacuole counts in DC as previously described [47]. DC integrity was monitored by propidium iodide staining. In the GAD/GAT4 migration restoration experiment, GABA supernatant was obtained from infected DC which had been incubated for 24 h in CM to enrich for DC derived GABA or commercially available exogenous GABA was added to the transwell medium following DC transfer. Migrated DC were quantified using a neubauer hemocytometer.
Non-infected DC or Toxoplasma-infected DC (PTGluc/MOI 1) were incubated ± LPS (200 ng/ml) as well as ± SNAP5114 (50 µM) or Semicarbazide (50 µM) for 24 h. Following incubation, DC were mixed with Collagen I solution (3 mg/ml, Gibco), 7.5% NaHCO3 (Invitrogen) and 10× Minimum Essential Medium (MEM; Invitrogen). Approximately 7.5×104 cells were loaded into μ-slide 3D chemotaxis chamber slides (Ibidi) and placed at 37°C, 5% CO2 to allow gel formation. Then, to establish a gradient, 1.25 µg/ml CCL19 (R&D systems) or GABA (0.5 µM or 5 µM) were added to one chamber reservoir whilst the other reservoir was filled with CM. Control experiments used CM in both reservoirs. Cell migration was monitored using a Zeiss AxioImager Z1 microscope and AxioVision software (version 4.7.2). Images were taken every 60 s for 60 min. Cell tracking and chemotaxis analysis were performed using ImageJ (http://imagej.nih.gov/ij/) with Manual Tracking (Cordelières, Institute Curie) and Chemotaxis Tool (Ibidi) plugins.
CCR7 expression of human monocyte-derived DC was studied using FITC-labeled anti-CCR7 and mouse IgG2a isotype control antibodies (R&D Systems). Murine bone marrow-derived DC were stained with CCL19-Fc (eBiosciences), CD 40, CD80, CD86, MHC II, CD18 antibodies (BD Biosciences) as indicated by the manufacturer. Data were generated using a CyAn ADP (Beckman Coulter) or a FACSCalibur (BD Bioscience) flow cytometer. Fluorescence-activated sorting of Toxoplasma-infected cells was performed at the Center for Cell Analysis, Karolinska institutet on a FACSAria (BD Bioscience) system. Dead cells were gated out by SYTOX Blue stain (Invitrogen) and pre-sorting infection rates were 43–63%. Data analysis was done with FACSDiva software, version 6.1.3 (BD Bioscience).
DC were challenged with freshly egressed PTGluc tachyzoites for 6 h at MOI 1. Extracellular parasites were removed following three washes at 80 g. Following infection and resuspension in PBS, 5×104 colony forming units (cfu) were adoptively transferred into male BALB/c mice. Total numbers of cfu injected into animals was confirmed by plaquing assays [70]. SNAP 5114 (50 µM) and Semicarbazide (50 µM) was added upon DC infection for 5 h. When indicated, such groups were treated with an additional 50 µM combination therapy of SNAP 5114 and Semicarbazide for 1 h, and added to PBS DC suspension prior to injection.
DC were stained with 5 µM 5-,6-(4-chloromethyl)benzoyl-amino-tetramethylrhodamine (CMTMR; Molecular Probes) or 5 µM carboxyfluorescein diacetate, succinimidyl ester (CFSE; Invitrogen) as indicated by the manufacturer. Stained cells (±50 µM SC+50 µM SNAP for 6 h) were then injected i.p. into BALB/c mice. After 12 h, the spleens were harvested, homogenized and cells were analyzed by flow cytometry (FACScalibur). An intraperitoneal lavage was performed and cells were similarly analyzed by flow cytometry.
BLI was performed as described [30]. Briefly, BALB/c mice inoculated i.p. with freshly egressed PTGluc tachyzoites, or with PTGluc-infected DC ± Semicarbazide and SNAP 5114 were injected i.p. with 1.5 mg D-luciferin potassium salt (Caliper Life Sciences, Hopkinton, MA, USA) and anaesthetized with 2.3% isoflurane prior to BLI. Ten min after injection of D-luciferin, biophotonic images were acquired at a binning of 8 (medium) for 180 s with an In Vivo Imaging System Spectrum CT (Caliper Life Sciences). For ex vivo imaging of organs, mice were injected i.p. with 1.5 mg D-luciferin and euthanized after 10–15 min. Organs were extracted as assessed as above. Analysis of images and assessment of photons emitted from a region of interest (ROI) was performed with Living Image software (version 4.2; Caliper Life Sciences).
Plaquing assays were performed as described [14]. Briefly, organs were extracted and homogenized under conditions that did not affect parasite viability. The number of parasites was determined by plaque formation on HFF monolayers.
All statistics were performed using Minitab version 15 (Minitab Inc, PA, USA).
Table S1 shows the 19 subunit primer pair sequences used to screen DC and astrocyte cDNA for GABAAR subunit transcripts and primer sequences for GAD 65, GAD 67, GAT4 and GAPDH cDNA. Figure S1 shows the ELISA-determined glutamate levels from non-infected and Toxoplasma-infected DC supernatant. Figure S2 shows GABA secretion and transmigration by monocytes upon challenge with T. gondii. Figure S3 shows GABA secretion of DC challenged with T. gondii type I, II and III strains. Figure S4 shows immunocytochemistry of infected and non-infected DC. Figure S5 shows parasite replication in DC in the presence of GABAergic inhibitors. Figure S6 shows the effects of GABAergic inhibitors on tachyzoite-infected DC and extracellular tachyzoites. Figure S7 shows effects of GABAergic inhibition on activation and maturation markers of DC. Figure S8 shows flow cytometric analyses of DC in vivo after GABAergic inhibition.
|
10.1371/journal.ppat.1002365 | Single Molecule Analysis of Replicated DNA Reveals the Usage of Multiple KSHV Genome Regions for Latent Replication | Kaposi's sarcoma associated herpesvirus (KSHV), an etiologic agent of Kaposi's sarcoma, Body Cavity Based Lymphoma and Multicentric Castleman's Disease, establishes lifelong latency in infected cells. The KSHV genome tethers to the host chromosome with the help of a latency associated nuclear antigen (LANA). Additionally, LANA supports replication of the latent origins within the terminal repeats by recruiting cellular factors. Our previous studies identified and characterized another latent origin, which supported the replication of plasmids ex-vivo without LANA expression in trans. Therefore identification of an additional origin site prompted us to analyze the entire KSHV genome for replication initiation sites using single molecule analysis of replicated DNA (SMARD). Our results showed that replication of DNA can initiate throughout the KSHV genome and the usage of these regions is not conserved in two different KSHV strains investigated. SMARD also showed that the utilization of multiple replication initiation sites occurs across large regions of the genome rather than a specified sequence. The replication origin of the terminal repeats showed only a slight preference for their usage indicating that LANA dependent origin at the terminal repeats (TR) plays only a limited role in genome duplication. Furthermore, we performed chromatin immunoprecipitation for ORC2 and MCM3, which are part of the pre-replication initiation complex to determine the genomic sites where these proteins accumulate, to provide further characterization of potential replication initiation sites on the KSHV genome. The ChIP data confirmed accumulation of these pre-RC proteins at multiple genomic sites in a cell cycle dependent manner. Our data also show that both the frequency and the sites of replication initiation vary within the two KSHV genomes studied here, suggesting that initiation of replication is likely to be affected by the genomic context rather than the DNA sequences.
| Kaposi's sarcoma associated herpesvirus (KSHV) establishes lifelong infection in the infected host and induces lymphoproliferative diseases, body cavity based lymphomas and sarcomas in immune compromised individuals. Herpesviruses including KSHV uses host cellular replication machinery for the replication of their genome. Here, for the first time we show that KSHV not only uses the host cellular machinery for its replication but also uses a similar mechanism for replication initiation at replication zones. KSHV was able to initiate replication throughout the genome thus the entire genome may act as a replication initiation zones. These data propose that replication initiations are not determined by the specificity of sequences but the genetic context of the genome and so suggest that epigenetic modification may play an important role in initiating DNA replication. Broadly, these results shed light on the evolutionary trends of large oncogenic dsDNA virus replication, which is similar to the replication of cellular DNA and therefore provides a strategy for the viruses to escape the host immune surveillance.
| Kaposi's sarcoma associated herpesvirus, also referred to as human herpesvirus 8 (HHV8), belongs to the gammaherpesvirus family and is associated with multiple lymphoproliferative diseases including Body Cavity Based Lymphomas (BCBLs) and Multicentric Castleman's Disease (MCDs) [1], [2], [3]. KSHV, like other herpesviruses establishes lifelong infection in the infected hosts and maintains the viral genome as extra-chromosomal episomes in a latent state [4], [5], [6]. The virus encodes a limited number of genes for persistence without being recognized by the host immune surveillance [7], [8], [9]. Latency Associated Nuclear Antigen (LANA) is one of the proteins expressed in all latently infected cells [5], [10], [11]. LANA is considered an oncogenic protein because of its role in modulating cellular pathways required to induce/promote tumorigenesis [12], [13], [14], [15]. LANA has also been shown to degrade the tumor suppressors, p53, pRb and von Hippel Lindau (VHL) by recruiting ubiquitin ligases [13], [16], [17], [18]. LANA has also been shown to upregulate the proteins important for immortalization of infected cells including upregulation of hTERT [12], [19]. Along with its role in modulation of various cellular and viral pathways, LANA is critical for maintaining the viral genome in infected cells [5], [6], [20]. LANA docks onto the host chromatin through the amino terminal chromatin-binding domain (CBD) and tethers the viral genome to the host chromosome by binding to the DNA binding domain of the carboxyl terminus within the terminal repeats [5], [6], [21], [22].
The KSHV genome has multiple reiterated copies of the terminal repeats (TR), which are proposed to be the region required for circularization of the genome. Each terminal repeat unit is a 801 bp long high GC content DNA element and was shown to contain the latent origin, or replication initiation site similar to EBV [23], [24], [25], [26]. Each TR unit has two LANA binding sites (high affinity site termed as LBS1 and lower affinity one termed LBS2) [24]. A 31 bp long sequence upstream of the LANA binding sequences is mapped as a replicator element (RE), which is important for replication initiation [24], [26]. Plasmids containing the RE element along with the LANA binding sequences is replication sufficient in a LANA dependent manner [26]. Comparison of the functional replication origins of KSHV with EBV showed that these two viruses differ in sequence homology but retain significant structural similarities [24]. For example, the terminal repeats of KSHV has two LANA binding sites, high and low affinities (LBS1 and LBS2) similar to the high and low affinity EBNA1 binding sites in the dyad symmetry element of EBV [24]. This suggests that both LANA and EBNA1 may share similar functions in terms of recruitment of cellular proteins [24]. A single copy of the TR with both the LANA binding site (LBS1/2) and RE is able to support the transient replication of a plasmid but requires at least two copies of the TR for stable episomal maintenance [23], [27]. Similar to KSHV, EBV oriP can also replicate with 2 copies of the EBNA1 binding site but requires an additional EBNA1 binding site within the family of repeats for stable maintenance [27], [28], [29]. Both, EBNA1 and LANA have been shown to stimulate DNA replication by directly recruiting the host cellular replication machinery [30], [31], [32], [33], [34]. Origin recognition complexes (ORCs) are the essential proteins which form the pre-replication complex (pre-RC) by recruiting the minichromosome maintenance proteins (MCMs) [35]. Both, LANA and EBNA1 have been shown to recruit ORC and MCM proteins at the replication origins and are important for latent replication [30], [32], [33], [34]. We have shown that licensing of the replication factors at these origins occurs once per cell cycle, which is similar to the cellular replication origins [26]. We have also shown that the initiation of DNA replication at TR is regulated by Geminin, which blocks replication during the same cell division [26]. Work from our lab and others clearly demonstrated that each TR unit has a replicator element [24], [26]. However, our recent work demonstrated the presence of an additional replication site at the left end of the KSHV genome [36]. This replication site does not require expression of LANA in trans, therefore we will refer to this as an autonomous replication origin (oriA throughout the manuscript) [36]. We identified oriA by screening the 33 kb left end of the genome by subcloning this cis-element into a plasmid and analyzing its ability to support replication [36]. Cellular replication proteins accumulated at the oriA in a cell cycle dependent manner to form pre-RC [36]. The presence of additional replicator sequences encouraged us to analyze the entire KSHV genome for potential replicator elements.
We used the single molecule analysis of the replicated DNA (SMARD) technology to identify the replication zones of individual KSHV episomes. SMARD is a powerful method for detecting replication initiation sites on the individual molecules using fluorescence microscopy [37]. This technology has been used for understanding the replication initiation and the dynamics of replication fork movement of Epstein Barr Virus [38]. SMARD has also been used for characterizing a number of replication origins in cellular DNA [39], [40].
In this study, we performed an extensive analysis of replication initiation sites and the replication dynamics of KSHV genome in two human body cavity based lymphoma cell lines (BCBL-1 and JSC-1). By stretching the DNA molecules extracted from latently infected cells, we were able to collect sufficient numbers of KSHV genomes representing the various stages of the replication fork movement. SMARD allowed us to determine the replication initiation site, progression of the replication fork, and termination of replication on the KSHV genome. Since this technique determines the replication fork progression in a steady state, it can also be used to calculate the duplication time needed to replicate specific regions of the KSHV genome. Our data now shows that replication initiation events can occur throughout the KSHV genome, which is distinctly different from earlier conclusions that replication initiates from a specific site within the terminal repeats. Our data also shows that both, the frequency and the sites of replication initiation vary within the two-studied KSHV genomes (BCBL-1 and JSC-1) suggesting that initiation of replication is likely affected by the genomic context rather than the DNA sequences. Detection of protein components of the pre-RC on the genome by chromatin immunoprecipitation assay shows accumulation of ORC2 and MCM3 on various regions of the KSHV genome suggesting potential replicator sites. These data suggests that replication initiates at various regions of the KSHV genome, and is primarily controlled by the genomic context rather than specific sequences.
In order to delineate the DNA replication initiation sites of the KSHV genome, we used BCBL-1 and JSC-1 PEL cells which maintain the latent viral genome. These two cell lines were labeled with nucleotide analogs, IdU (5′-iodo-2′-deoxyuridine) and CldU (5′-chloro-2′-deoxyuridine) for fluorescent visualization of DNA replication across the genome using a previously established technique, single molecule analysis of the replicated DNA (SMARD) [37], [38]. In this technique, the replicating DNA is labeled in a way, which helps us to determine the position, direction and the abundance of regions of replication forks on replicated molecules. This in turn allows us to determine the replication initiation sites, progression of replication forks and replication termination sites on the genome analyzed. In this labeling procedure, asynchronously growing cells are sequentially labeled with IdU and CldU for the complete duplication of the KSHV genome (4 h). This allows the replicating molecules to incorporate the halogenated nucleotides throughout the entire length of the genome [38]. These cells are switched from the first label (IdU) to the second label (CldU) after 4 h, thus transitioning the incorporation of nucleotides, which are detected by halogenated nucleotide specific antibodies. The transition from red to green marks the position of the replication fork at the time of switch from the first to the second nucleotide labels.
Growth kinetics was performed to determine the concentrations of the halogenated nucleotides used for labeling the PEL cells. Both, BCBL-1 and JSC-1 cells were cultured with 10, 25 and 50 µM of IdU and CldU for 12 h and the cell viability was determined. These halogenated nucleotides did not show any significant change in the growth pattern up to 25 µM (Supplemental Figure S1A and B). Therefore, we used 25 µM, final concentrations, of these analogs for labeling the cells in our subsequent experiments. The labeled cells were also analyzed for their cell cycle profiles, to determine that the cells were growing normally, by flow cytometric analysis after staining with propidium iodide (PI). These cells did not show any difference in cell cycle profile as compared to the untreated cells suggesting that these analogs did not affect the cell cycle at the concentrations used for labeling the cells (Supplemental Figure S1, C and D). We also determined whether an addition of nucleotide analogs triggers lytic reactivation by analyzing the levels of the immediate early gene, RTA (replication and transcriptional activator). RTA, which is essential for initiating the lytic DNA replication, was undetected by the western blot in IdU treated BCBL-1 and JSC-1 cells (Supplemental Figure S1E).
Labeled cells were washed to remove any unincorporated nucleotides and were resuspended then placed in agarose plugs. These cells were lysed as mentioned in the materials and method section followed by digestion to linearize the KSHV genome. The cells are digested in agarose plugs to avoid any DNA breakage, as we required full-length episomes for analysis. The linearized KSHV genome was partially separated from genomic fragments by pulsed field gel electrophoresis using the CHEF DRII system as described in methods section. The band of linear KSHV genomes in the gel was determined by a Southern blot using specific probes shown in the supplemental information (Supplemental Figure S2). Linear KSHV genome showed a band at approx 165 kb, which was determined based on pulsed field DNA markers (NEB). Agarose from the gel was excised for DNA extraction. The DNA was extracted after melting the agarose slice-containing the regions of interest and digesting with gelase. An aliquot of DNA solution was stretched between the positively charged (3-aminopropyl-tri-ethoxysilane-coate) glass slides and non-silanized cover slip through capillary action. Staining the DNA with an intercalating dye, YOYO-1, aided in visualizing stretched DNA by fluorescence microscopy (Supplemental Figure S3). Since molecules stretched by capillary action may vary in their orientation and size, three biotinylated probes were used for fluorescence in situ hybridization (FISH) followed by detection with Alexaflour 350-conjugated Avidin (shown in blue throughout the manuscript) to distinguish KSHV molecules from other DNA molecules. Use of different sizes probes produced a distinctive blue “bar code” which helped in determining the orientation of the molecules. The halogenated nucleotides are detected using specific monoclonal antibodies and secondary antibodies conjugated with Alexaflour 488 (shown in green throughout the manuscript; CldU) and Alexaflour 594 (shown in red throughout the manuscript; IdU).
Since the DNA molecules were substituted throughout their length with nucleotide analogs, they were easily detected even in the presence of substantial background signals. Figure 1 shows an optical field with halogenated nucleotide substituted DNA (linear red and green signals). The signals, which are not on a line, may have been due to the background signals or broken pieces of DNA. The KSHV DNA was distinguished from the non-KSHV DNA by the presence of FISH (blue) signals. In order to determine the average length of linear KSHV episomes, fully substituted molecules either with the first label, IdU (detected in red) or the second label, CldU (detected in green) were aligned with the schematic of the PmeI linearized genome. These molecules marked by the presence of probe signals, P10 at the left side and P6 at the right side (depicted in the schematic Figure 2B) of the PmeI linearized genome, showed an average length of 66 µm corresponding to about 2.5 kb/µm (Figure 2A). The molecules that substituted fully during first or the second labeling periods allowed us to calculate fork rates.
In order to identify the replication initiation sites on the BCBL-1 genome, photomicrographic images of the molecules that were fully substituted and contained both nucleotide analogs were aligned using the FISH signals as a template. The alignment was done in a non-subjective manner but based on increasing incorporation of IdU (red signals) in randomly selected representative stretched DNA molecules from the pool of collected molecules. Alignment of molecules based on the increasing signals effectively removes any potential bias in the arrangement of molecules. Therefore, if a replication fork moves in a single direction within the examined region, the molecules will show an increase in the length of red signals from one end. A replication initiation site is characterized by a region of red signal (IdU sites) flanked by the green signals (CldU sites) on both sides. Additionally, if the replication fork moves in both directions, the molecules will have increasing red signal on both sides of the replication initiation sites. In contrast to replication initiation, the termination sites are marked by a patch of green stain flanked by the red signals on both sides. A schematic of replication initiation sites and termination are shown in Figure 3. Replication initiation sites are marked by black vertical black arrows (Figure 3B).
We collected images of 151 PmeI linearized KSHV episomes which incorporated the halogenated nucleotides along their entire length (41 fully substituted with IdU, 39 fully substituted with CldU and 71 stained with both IdU and CldU). The results of these experiments are shown in Figure 3 and supplemental Figure S4. The transition site from red to green in the molecules with both the halogenated nucleotides defines the position of replication fork at the time of transition from the first to the second label. The red staining portion of these molecules that varied in different PmeI linearized molecules suggested that replication initiated at different positions in these molecules. Schematics of the molecules with red and green staining patterns were drawn and center (vertical black arrow in diagrams and horizontal yellow arrows show direction of the replication forks) of the red stained patches were marked as the site of replication initiation assuming (as is usually the case) latent replication proceeds in a bi-directional manner (Figure 3A and B). In order to define the progression of replication throughout the genome, the replication profiles of these molecules were generated (Figure 3C). The replication profile was generated by dividing the KSHV genome map into 5 kb intervals and determining the percentage of molecules stained with red in each interval. The replication profiles are important in determining the region, which replicated first, termed as RRF (region replicating first) and regions, which replicated last, RRL. RRF are proposed to contain the replication initiation sites used most frequently for the duplication of the genome. However, in BCBL-1 episomes, approximately 60% of the molecules showed replication initiation within intervals 1 through 27. Intervals 28–31 showed slightly lower frequency of replication initiation (Figure 3C). Interestingly, the terminal repeats region (spanning intervals 22–25), containing multiple LANA dependent replication origins, did not show any preferential sites for replication initiation. This suggested that the KSHV genome could initiate replication throughout the genome. We also detected molecules with replication forks moving in only one direction (Figure 3A and B, molecules 56–71), which could be due to unidirectional replication in some molecules.
The progression of replication forks throughout the KSHV genome was described by the profiles of replication fork abundance, which was obtained by dividing the KSHV genome map into 5 kb intervals and then indicating the percentage of molecules with red-to-green transitions within each interval. As mentioned above, the transitions sites indicate the position and the direction of replication forks at the time of switching of the first to the second label [38]. Accumulation of transition sites from red to green at intervals 1, 3 and 9 indicated that the replication fork was not moving freely within those intervals which could suggest that the observed region may have a site where forks pause but only for a short duration.
Profiles of replication fork abundance were also used to determine the prevalent direction of replication fork movement throughout the KSHV genome (Figure 3D). In the BCBL-1 genome, the replication forks moved in both directions at similar frequencies in most of the intervals indicating bi-directional replication. However, intervals 27, 28, 29 and 32 showed higher number of molecules with forks moving towards the right end of the genome suggesting that the replication fork may have paused for a short time in that region (Figure 3D). Peaks with molecules having the replication fork movement in only one direction may indicate the replication fork pausing sites (Figure 3D). It could also be possible that the left end of the fork is masked with the probe (p10) or deleted during DNA processing in those molecules with single fork. Single replication fork could also be due to lytic replication, which utilizes a rolling circle strategy for replication. KSHV infected cells primarily maintains latent viral genome but a small proportion of the cells may undergo lytic reactivation spontaneously thus leading to the generation of the molecules with a single fork. Nonetheless these molecules did not appear to significantly affect the calculation of the replication fork progression.
Earlier studies mapped the replication initiation site to the terminal repeat region and to the autonomously replicating element region (oriA) [28], [36]. However, these studies were performed only on selected regions of the genome without taking the complete genome into consideration. SMARD detects replication initiation sites throughout the genome by analyzing the immunostaining patterns of the DNA molecules. Multiple initiation events should produce multiple red staining patches each surrounded by green signals. The immunostaining pattern of these molecules (shown in Figure 3B) revealed multiple replication initiation sites. For example, early initiation events took place within the TR region of molecules 9, 14 and 15 detected by red stained region flanked by the green signals (Figure 3B). However, shorter red regions were also detected on the same molecules indicating the occurrence of replication initiation events at later time points. We indicated that multiple replication initiations were present if replication initiation events on particular molecules were not synchronous (different sizes of red staining patches), we referred the primary to the first initiation event and secondary to the subsequent initiation events. The secondary initiation event in the TR region is shown by a smaller red staining signal in molecule 12 (Figure 3B). In molecules 1, 2, 35 and 37 multiple initiation events were recorded with secondary initiation events in region of oriA (Figure 3B). Secondary initiation events were also detected within regions of genes transcribed during latent infection. Molecule 28 has a secondary initiation event in the vIRF transcripts region. This suggests that initiation events were not limited to any particular region of the genome but that the entire KSHV genome constitutes a large initiation zone for replication.
In previous experiments we showed replication initiation sites and the progression of replication fork qualitatively. Here we wanted to determine the speed of replication forks in different regions of the KSHV genome using the equation described earlier [38]. SMARD data is used to determine the average time required for duplication of any portion of the genome (Td: Figure 4A). By using the time required for the duplication of a known length or segment, duplication speed (Sd) can also be calculated (Figure 4B). These conclusions are based on the data from a large pool of molecules and are performed on large genomic segments. Therefore the quantitation is not significantly affected by the resolution of the technique. We calculated the duplication speed of the left and the right side of the KSHV genome using the fully substituted red and green molecules (Supplemental Figure S4). We found that these regions replicated at a different speed with the left side proceeding at 0.91 kb/min and 0.62 kb/min in the right side of the molecules (Figure 4C). The differences in the speed could be due to the complexity (high GC content due to multiple TR units) on the right side of the genome, which may impede the movement of replication forks.
As mentioned previously, some molecules in the aligned image showed multiple red stained patches flanked by green patches indicating the occurrence of multiple initiation events (Figure 3B). Two replication initiation sites were detected in 10 out of 71 imaged molecules suggesting that multiple replication events significantly occurred significantly during replication. These initiation events are distinctly separated in terms of distance and time of their activation. However, replication initiation events that are located near each other may have not been detected as individual initiation sites due to the resolution limit (approx. 5000 bp) of this technique. The terminal repeat region, which consists of approx. 25 copies of terminal repeat units may be initiating replication at each unit since the individual units are shown to have functional replication elements [24]. It would be interesting to perform SMARD on the TR region cloned into a replication deficient vector (cosmid) to determine whether replication initiated at all the TR units or is restricted to any specific one. Based on the SMARD data performed on the entire genome we confirmed the occurrence of multiple replication events on the BCBL-1 genome.
Initiation of DNA replication requires assembly of pre-replication complexes (pre-RCs) on DNA [41], [42]. Assembly of the pre-RC begins with the binding of origin recognition complexes (ORCs) to the chromatin [41], [42], [43]. ORCs consist of a stable core complex including subunits ORC2-ORC5 associated with less stably bound components ORC1 and ORC6 [42], [44], [45]. These ORCs recruit two other proteins, Cdc6 and Cdt1, which stabilizes the binding of ORC and allows Cdt1 to load the replicative helicase, mini-chromosome maintenance (MCMs) proteins (MCM2-7) [46], [47], [48]. Association of ORCs, Cdc6, Cdt1 and MCM2-7 completes the process of pre-RC assembly, also referred to as licensing of the replication origin. Importantly, pre-RC formation and replication competency is restricted to the G1 phase of the cell cycle [46], [49]. Some replication origins are activated at the beginning of S-phase by the recruitment of protein kinases, Cdc7 and Cdk2 [46], [49]. Therefore, chromatin immunoprecipitated from G1/S cells with Pre-RC complex proteins brings down the active replication initiation sites. These pre-RC components are disassembled after the cell completes the cell division. We used ORC2 and MCM3 (components of pre-RC) antibodies to immunoprecipitate chromatin from G1/S cells and compared their association with G2/M phase cells. In order to get pure population of cells in G1/S and G2/M phases, we used the centrifugal elutriation on asynchronously growing cells. The cell cycle analysis of these cells by propidium iodide showed that over 80% cells were in G1/S phase (Supplemental Figure S5). Similarly, the G2/M phase cells had very high proportion of cells in the fraction used for ChIP assay (Supplemental Figure S5).
Immunoprecipitated chromatin from the G1/S and G2/M phase cells was analyzed for relative amplification of immunoprecipitated DNA in a semi-quantitative real-time PCR assay. We analyzed the entire BCBL-1 genome by using PCR primers at approximately 1.5 kb intervals. Comparison of immunoprecipitating copies of these genomic regions in G1/S and G2/M phase cells showed very specific association of both ORC2 and MCM3 in G1/S cells (Figure 5). Importantly, there were multiple sites on the genome, which were enriched with ORC2 and MCM3 chromatin immunoprecipitation suggesting the formation of pre-RC at those sites (Figure 5). This supports our SMARD data of multiple replication initiation sites on the genome. Regions around ORF6, ORF31, ORF50, ORF56, ORF71-71 and ORF75 showed relatively higher copies of genome precipitating with ORC2 in G1/S cells. The enrichment of these fragments was also seen with MCM3 antibodies suggesting the accumulation of pre-RCs at multiple sites on the KSHV genome. Chromatin immunoprecipitating the TR region (multiple reiterated copies of terminal repeats), represented by a primer set S45, showed specific enrichment of ORC2 and MCM3 in G1/S phase cells suggesting cell cycle specific accumulation of pre-RCs and origin firing at the TR.
We were interested in analyzing the KSHV genome replication in another PEL cell line. We selected JSC-1 cells because this cell line maintains higher copies of the viral genome in latently infected cells. In order to determine whether the difference in copy number was due to the number of replication initiation events or the speed of the replication fork movement? We performed SMARD similar to that done for the BCBL-1 genome and recovered the images of 72 PmeI linearized KSHV episomes substituted along their entire length with halogenated nucleotides (35 red/green, 19 red and 18 green). The molecules with red and green were aligned with increasing red staining signals as shown in Figure 6. These two KSHV strains, BCBL-1 and JSC-1 showed a slight difference in the replication initiation patterns. In BCBL-1 KSHV genome, replication initiation events occurred uniformly throughout the entire length but in JSC-1 KSHV the right end of the genome showed higher replication initiation events (Figure 6C). JSC-1 KSHV also showed multiple initiation events but was slightly lower compared to BCBL-1 (2 molecules in JSC-1 compared to 10 molecules in BCBL-1 images). These multiple initiation events were mostly asynchronous as determined by different sizes of the regions with red staining (Figure 6 A and B). The replication profile of JSC-1 KSHV also showed the difference in termination sites. However, the genome replicated uniformly without significant pausing (Figure 6D). In summary, these data showed that replication initiation sites are not confined to any specific region of the KSHV genome and thus led us to conclude that the utilization of replication initiation sites can change in different molecules as well as different strains of KSHV. This result also suggests that replication initiation may also be controlled by epigenetic modification and will be an important area of study.
Since the replication initiation sites of KSHV in BCBL-1 and JSC-1 cells were slightly different, we wanted to determine if the speed of DNA replication was also different. Images of the molecules shown in Figure 3 and the supplemental Figure S6 were used for calculating the Sd for the left and right portion of the JSC-1 genome. We found that DNA duplication proceeded at different speed in the left verses right portion of the genome with 1.49 kb/min being highest in the left side and 0.66 kb/min in the right side (Figure 7A). It was interesting to note that the right portion of the genome, which has terminal repeat region, replicated with almost similar speed, 0.66 and 0.62 kb/min in both BCBL-1 and JSC-1 KSHV genome, respectively. However, the left end of the genome duplicated much faster in JSC-1 KSHV (1.49 kb/min) as compared to the BCBL-1 KSHV (0.91 kb/min). This suggests that replication progresses with different kinetics within different regions of the genome and could be due to genomic complexity or due to the presence of different pausing sites. Additionally, duplication may also be affected by the presence of multiple initiation events within the region replicating faster.
The duplication speed, which reflects the average number of replication forks for the replication of specific region, was determined for the entire KSHV genome. Although the duplication speeds of the left portion of the BCBL-1 and JSC-1 virus were different the overall speed for the duplication of the entire KSHV genome was similar. Duplication speed of BCBL-1 KSHV was 1.08 kb/min and JSC-1 duplicated with a speed of 1.06 kb/min (Figure 7B). On average the duplication speed of KSHV in both PEL cells were 1.0 kb/min, which was slightly slower compared to the duplication speed of Epstein Bar Virus (EBV) (1.5 kb/min) [38].
The KSHV genome persists in the infected cells in a latent state by expressing a small number of viral genes [1], [2]. Furthermore, LANA is a predominantly expressed protein in all KSHV infected cells [2], [50], [51]. LANA has been shown to tether the viral genome to the host chromosome by binding to the histone proteins [5], [21]. Further, LANA has been shown to be critical in maintaining the viral genome copies in these cells as a recombinant KSHV (BAC36) deleted for the LANA gene could not establish latent infection and was lost from the cells within two weeks [20]. Additionally, LANA depletion by shRNA reduced the viral genome copies to basal level in PEL cells [52]. These studies suggested a role for LANA in maintaining the viral genome in infected cells. Since LANA colocalized with the KSHV genome on the host chromosome, it was suggested that LANA binds to the viral genome directly [5]. Studies focusing on the identification of LANA binding sites identified the LANA binding sites along with the replication elements within the terminal repeat region [22], [28]. Further studies, focused on the identification of replication initiation sites, identified the LANA dependent origin within the terminal repeat region [23], [24], [26], [53]. In our more recent study, we analyzed the left end of the KSHV genome and identified a replication origin (oriA), which did not require LANA expression in trans [36]. This prompted us to analyze the entire KSHV genome for replication initiation events using a recently developed and powerful technique, SMARD, that looks at the replication events on individual DNA molecules [38].
In this study, we determined how DNA replication initiated and progressed on KSHV episomes of two latently infected PEL cell lines (BCBL-1 and JSC-1). We further determined the abundance and the dynamics of replication fork movement in these cells. Replication can initiate at various regions of the genome including the terminal repeat region. There were slight differences in the direction of replication fork movement and the duplication speed of the left end of KSHV genome in BCBL-1 and JSC-1 cells (Figure 3 and 6). However, the frequency of replication initiation and termination events varied across the genome and between the two KSHV strains. These differences were detected even though the genomic sequences of these two strains are identical. SMARD performed on two different strains of EBV also showed a difference in the replication initiation and termination sites [38]. Differences in replication profiles were proposed to be due to the differences in genetic makeup of these two strains (MutuI has deletion of internal repeat 1 and Raji has two short deletions) [38]. However, analysis of various regions by SMARD and 2D gel analysis confirmed that these deletions did not account for all the differences in the replication pattern [38].
This study also characterized the pausing sites of the replication forks outside the TR region. Accumulation of replication forks is clearly detected in both the strains of KSHV within the genomic regions outside the TR and at the left end of the genome. Detection of lower numbers of pausing sites within the left end of JSC-1 could contribute to a faster replication fork movement within that region. Speed of replication within the right side of the genome did not vary significantly suggesting the presence of similar epigenetic and genetic organization in that region within these two KSHV strains. Since transcription may have an essential role in controlling replication fork movement and pausing, it may suggest that both strains have similar transcriptional patterns within the right end of the genome. Another common feature between BCBL-1 and JSC1 was that the time required for the replication of the genome was similar. This suggests that replication forks proceeded uniformly throughout the genome to complete the replication.
The presence of single replication forks in some molecules may indicate a replication pausing zone at one end of the bi-directional replication fork or the fork may have got masked in the FISH probe signal at the left end of the genome. Detection of single replication fork may also indicate that some molecules were undergoing unidirectional replication as proposed for the rolling circle mechanism. Herpesviruses replicate by a rolling circle mechanism during lytic cycle using lytic origins of replication [54]. Therefore, it could be possible that a small proportion of the cells BCBL-1 and JSC-1 may have undergone lytic reactivation and produced these single replication fork-containing molecules. It would be interesting to determine the lytic replication mechanism of KSHV using SMARD but that is beyond the scope of this study.
Chromatin immunoprecipitation data with ORC2 and MCM3 identified multiple sites where the pre-RCs were accumulated suggesting the formation of multiple active replication initiation sites. However, it is important to note that pre-RCs sites on the genome and replication initiation by SMARD showed very limited correlation. This could possibly be due to the fact that the SMARD data is from individual genome molecules as compared to ChIP data, which is a cumulative data from million copies of the KSHV genome. The genomic region between primer sets S16-S16 (Figure 5) showed enrichment of pre-RCs, which also correlated with replication initiation sites determined by SMARD (interval 5 of Figure 3D). Importantly, epigenetic modifications may also have a critical role in determining the replication initiation sites. Studies have shown that histone acetylation can influence the timing of replication origin firings [55]. A recent study showed comprehensive epigenetic modifications of histone H3 for acetylation, K4 and K9 tri-methylation and K27 tri-methylation [56]. Histone H3 acetylation and tri-methylation at K4 indicates a transcriptionally active chromatin. Chromatin immunoprecipitation with acetylated histone H3 and tri-methylated Lysine 4 (K4) in KSHV PEL cells, BCBL-1 has shown various regions with active chromatin [56]. Also studies have demonstrated that regions with active transcription shows reduced or no replication as both of these activities do not occur at the same time in mammalian cells [57]. Presence of multiple silent chromatin regions in the BCBL-1 genome could possibly explain the occurrence of multiple replication initiation events, and the pausing of replication forks on the latently replicating genomes. Each of the transcriptionally silenced regions may potentially serve as the replication initiation sites and thus were detected by the SMARD assay. Additionally, the replication fork stalling sites could be the regions where replication forks traversed through the RNA polymerase bound DNA site as the RNA polymerase may remove or inactivate the pre-replication complex deposited on the DNA for replication. However, understanding the coupled events of replication and transcription may be very complex mechanistically and would be beyond the scope this study. A recent study on bacterial genome replication understanding the movement of replisome through a protein nucleic acid block showed that head on collision of the replisome with RNA polymerase results in replication fork arrest but co-directional collision has little or no effect on the fork progression [58]. Therefore, the presence of transcriptionally active sites on the genome may be the reason for replication fork pausing if the replisome collides head on with the transcription machinery.
In terms of the replication elements, the terminal repeat elements are well characterized for the replication of plasmids in the presence of LANA [23], [28]. SMARD also detected replication initiation sites in the TR region but showed that this was not the preferred site for genome duplication. The infrequent usage of TR as the replication site does not seem to be due to the lack of pre-RCs accumulation in this region. Various studies have shown the recruitment of ORCs and MCMs in the TR region suggesting functional replication by the TR region [33], [34]. However, the role of TR mediated replication in the context of entire genome was not determined. Therefore, it would be interesting to generate a recombinant KSHV lacking the replication origin site within the TR element to evaluate its potential for replication and establishment of latent infection. There are technical difficulties to generate a recombinant KSHV deleted for RE element (element required for replication initiation in the TR region) because each TR copy has one RE element and there are 25–30 reiterated TR copies [24]. Therefore, it will require multiple rounds of recombination to remove the RE elements from each TR copy. Alternatively, a recombinant KSHV can be generated by excising all the TR copies and replacing it with RE deleted TRs in the BAC KSHV (BAC36) that will lack the TR mediated replication. Hence, this recombinant KSHV will be important in evaluating the role of primary and secondary replication initiation sites in the establishment of latent infection.
This study suggests that the KSHV genome replicates not only by using the cellular replication machinery but also follows a mechanism similar to that used for the cellular DNA replication. DNA replication in mammalian cells initiates in replication zones, which are similar to the replication zone of the entire KSHV genome [59]. This shows a broad impact of our study since KSHV can be used as a model system to investigate the unresolved molecular mechanism of cellular DNA replication. This study also highlights the divergent evolutionary trends of human viruses where large dsDNA viruses replicate using a similar mechanism to the host cellular DNA without requiring viral proteins which could be detected by the host immune system. Use of a similar replication mechanism by these viruses may provide an advantage in establishing life long infections without being detected by the host immune surveillance even in healthy individuals.
KSHV infected body cavity based lymphoma cells (BCBL-1 and JSC-1) were cultured in RPMI supplemented with 10% fetal bovine serum, 2 mM L-glutamine and penicillin-streptomycin (5 U/ml and 5 µg/ml, respectively). These cells were grown at low density (3×105–6×105) in order to keep the virus primarily in the latent state. Cells growing at approximately 5×105 cells per ml were labeled for 4 hrs each with 25 µM 5′-iodo-2′-deoxyuridine (IdU; first label) and 25 µM 5′-chloro-2′-deoxyuridine (CldU; second label). IdU was directly added into the media with actively growing cells followed by low speed centrifugation (600 xg) at RT to collect the cells at the end of first labeling period. These cells were washed with warm PBS and resuspended in pre-warmed media with a second label, CldU at the density of 5×105 cells/ml. Cells were collected, washed and mixed with molten InCert agarose (Lonza Inc. Rockland, ME) and poured into the molds for making plugs. These plugs contained approximately 106 cells/plug.
Plugs containing BCBL-1 and JSC-1 cells were treated with cell lysis buffer (0.5 M EDTA+1% Sarcosine and Proteinase K) at 50°C for 96 h by changing the lysis buffer at every 24 h. These plugs were thoroughly washed with TE and digested to linearize the KSHV genome; we chose PmeI restriction enzyme, which cleaves the viral genome once. Before digestion with PmeI, the plugs were incubated with digestion buffer and digested with 50 U of enzyme.
Digested plugs were loaded onto 0.6% LMP agarose (Seplaque, Lonza Inc., Rockland, ME) followed by resolving them on a pulse field gel electrophoresis using CHEF-DRII (Bio-Rad Laboratories, Hercules, CA). The region of linear KSHV genome banding on the gel was identified using KSHV specific 32P labeled probe after transferring part of the gel onto a nylon membrane. The linear KSHV genome was extracted from the gel by cutting a small slice of agarose containing KSHV DNA and treating with gelase (Epicenter Biotechnologies Inc., Madison, WI) as per manufacturer's recommendations.
We used a capillary method to stretch DNA molecules on the glass slides, described previously [37], [38]. Isolated and enriched linear KSHV genome, after gelase treatment, was gently deposited at the interface between silanized microscope slides and non-silanized coverslips. The stretching was determined under a fluorescent microscope by adding DNA staining dye, YOYO-1 to the DNA mix before stretching. DNA was diluted to a point where there were no overlapping DNA molecules. The coverslips were removed gently followed by dipping the slides in methanol. DNA molecules were denatured with 0.1 M NaOH and fixed with Glutaraldehyde.
Hybridization was performed as described earlier [38] using probes labeled with biotin-16-dUTP (Roche Inc. Indianapolis, IA). Three different sizes probes, p10 (10 kb region between 36883-47193), p15 (15 Kb region between 85820-100784) and p6 (6 kb region between 26937-33194 nt) corresponding to KSHV genome (accession number NC_93331) were used in this study. The above-mentioned regions of the KSHV genome were cloned into plasmid/cosmid vectors and sequenced to ensure the presence of desired region. A purified band excised from the vectors were labeled with biotin using bio-nick labeling kit (Invitrogen, Inc., Carlsbad, CA) according to the manufacturer's recommendations. The length of the probe varied from 100–500 bp as determined by agarose gel electrophoresis. The hybridization signals were detected by NeutrAvidin conjugated to Alexa Flour 350 (Molecular probes Inc.Eugene, OR) as described earlier [38]. Briefly, five layers of Alexa Flour 350 conjugated with NeutrAvidin and Biotinylated anti-Avidin antibodies (Vector Laboratories Inc., Burlingame, CA) were deposited on the microscopic slides after washing with PBS containing 0.03% Igepal CA-630 (Sigma Aldrich, St. Louis, MO) after each step. The purpose of the probe was to identify KSHV genome molecules from the pool of DNA molecules, including cellular DNA. By using asymmetric probes, we aligned the molecules in the same orientation, post-image capture. Since the cells were labeled for sufficient time (4 h each label) to replicate the entire KSHV genome, we could easily detect the KSHV molecules even in the presence of substantial hybridization background (blue dots in hybridization probes panel of Figure 1). The hybridization background does not affect the SMARD, therefore digitally removed in further Figures (Figure 2, 3, 6 and supplemental figures 4 and 6).
Incorporated IdU and CldU were detected by immunostaining of the stretched DNA along with the detection of biotinylated hybridization probes. The labeled IdU and CldU were stained with mouse anti-IdU and rat anti-CldU as primary antibodies, respectively. Alexa Flour 594 conjugated goat anti-mouse and Alexaflour 488 chicken anti-rat were used as secondary antibodies to detect IdU and CldU, respectively. As described previously [38], antibodies for IdU and CldU did not cross-react and also did not stain any unlabeled DNA. The slides were mounted and stored in the dark. The images were captured using an automated fluorescent microscope (Axio Observer, Zeiss Inc., Thornwood, NY) with 63X objective. The slides were scanned to find the signals for probe (Blue color) to ensure the intactness of the KSHV molecules. DNA molecules with either red, green or both colors with the hybridization signals were imaged. The lengths of the DNA molecules were measured after capturing the images using an inbuilt feature of image capture and analysis software (Axiovision, Zeiss In., Thornwood, NY). Molecules with all three probes signal and appropriate size were subjected to analysis. Individual molecules were cropped and aligned based on the content of increasing red signal from top to bottom (molecule number 1 through 72 in Figure 3 and 1 through 35 in Figure 6).
As described previously [38], we used exponentially growing asynchronous cells to label the replicating DNA. Additionally, we used labeling periods required to fully replicate the KSHV genome to ensure that the replicating molecules were duplicated as the duplication time could vary in different molecules. We examined the molecules, which were fully substituted during these labeling periods and the molecules that have just started to replicate at the time the pulse incorporates throughout DNA molecule. However, the molecules, which had a replication fork in the middle of the molecule at the time of pulsing, will have two labels. Another set of cells, which starts replication at the time of second pulse, will have first label incorporated throughout the molecule and small patch of second label at the site of replication origin. The site of transition from the first label to the second marks the site of the replication fork. By analyzing these fully substituted molecules, we eliminated the bias of looking at replication forks at any particular region of the molecule. Therefore, our analyses faithfully represent the distribution of the replication fork in a steady state population of replicating molecules [38].
The use of long labeling periods to substitute the region of interests with halogenated nucleotides is advantageous because it provides an internal control, which cannot be achieved by short labeling periods. Since the molecules are fully immunostained, they are easy to align with the KSHV genomic map. This also allows us to detect the unevenly stretched molecules, which were discarded from further analysis. Since we look for contiguous staining in fully substituted molecules, any breakage detected by loss of signal can be easily detected. Another shortcoming, which we encountered during stretching of the DNA molecules, is overlapping signals of the DNA molecules. Since the molecules are fully substituted, these overlaps are easily detected so these are also discarded from the analysis. It is also important to mention that a hybridization signal reduces the affinities of IdU and CldU antibodies to the corresponding regions of the DNA molecules (Figure 1). This causes a significant loss of information in regions to which the probe hybridizes but ensures that the immunostained molecules we have analyzed are indeed the molecules of interest. Lastly, the dual labeling scheme and detection of both nucleotides allowed us to believe that these labeled cells replicated normally without the introduction of any bias in image collection. In an ideal condition, when the above-mentioned criteria are met, the number of molecules fully substituted with first label, IdU are expected to be similar to the number of CldU substituted molecules. These controls represent the steady state replication of KSHV DNA molecules.
G1/S and G2/M fraction of the asynchronously growing BCBL-1 and JSC-1 cells were fractionated by centrifugal elutriation. Cells from these fractions were cross-linked with 1% formaldehyde by rocking for 10 min at RT followed by adding 125 mM glycine to stop the cross-linking reaction. Nuclei from these cells were isolated followed by sonication of the chromatin to an average length of 600 bp. Cell debris was removed by centrifugation of the chromatin at high speed for 15 min at 4°C. Samples were pre-cleared with salmon sperm DNA/ProteinA Sepharose slurry for 30 min at 4°C with rotation. Supernatant were collected after brief centrifugation. 10% of the total supernatant was saved for input control. The remainder, 90% was divided into three fractions. i) Control antibody [Sigma Aldrich St. Loius, MO]. ii) Anti-ORC2 [Santa Cruz Biotechnology, Inc., Santa Cruz, CA] and iii) anti-MCM3 antibody [Abcam Inc., Cambridge CA]. Immune complex was precipitated using salmon sperm DNA/ProteinA/ProteinG slurry. Beads were then washed consecutively and the complex was eluted by using the elution buffer [1% SDS/0.1 M NaHCO3] and reverse cross-linked by adding 0.3 M NaCl at 65°C for 4–5 h. The eluted DNA was purified by treatment with proteinase K at 45°C for 2 h, phenol extraction and ethanol precipitation. Chromatin saved for input was also reversed cross-linked to extract DNA for real-time PCR analysis. Purified DNA from input, control antibody, ORC2 and MCM3 were dissolved in 100 µl sterile water. Primers spanning the entire KSHV genome at every 1.5 kb were used in the real-time PCR assay. List of primers and their sequences are available upon request. DNA from input, control IgG, ORC2 and MCM3 samples were used from each primer set. Relative amounts of DNA in chromatin bound to ORC2 and MCM3 proteins were calculated by subtracting the amplification with control antibody.
Centrifugal elutriation was performed essentially as previously described [60]. Briefly, BCBL1 and JSC-1 cells grown in RPMI 1640 at 5×105 cells/ml were separated into different cell cycle phases, with the flow rates of 14, 16, 18, 20, 24, 29 and 36 ml/min by a modified Beckman JE 5.0 rotor (Beckman Coulter Inc., Brea, CA) The relative DNA content of the different fractions was determined by flow cytometry (Becton Dickinson Inc., Franklin Lakes, NJ) by the fluorescence intensity of propidium iodide stained cells. The fractions of G1 (16 ml/min) and G2 (29 ml/min) were performed for ChIP assay as described previously [61].
|
10.1371/journal.ppat.1000633 | The HIV Envelope but Not VSV Glycoprotein Is Capable of Mediating HIV Latent Infection of Resting CD4 T Cells | HIV fusion and entry into CD4 T cells are mediated by two receptors, CD4 and CXCR4. This receptor requirement can be abrogated by pseudotyping the virion with the vesicular stomatitis virus glycoprotein (VSV-G) that mediates viral entry through endocytosis. The VSV-G-pseudotyped HIV is highly infectious for transformed cells, although the virus circumvents the viral receptors and the actin cortex. In HIV infection, gp120 binding to the receptors also transduces signals. Recently, we demonstrated a unique requirement for CXCR4 signaling in HIV latent infection of blood resting CD4 T cells. Thus, we performed parallel studies in which the VSV-G-pseudotyped HIV was used to infect both transformed and resting T cells in the absence of coreceptor signaling. Our results indicate that in transformed T cells, the VSV-G-pseudotyping results in lower viral DNA synthesis but a higher rate of nuclear migration. However, in resting CD4 T cells, only the HIV envelope-mediated entry, but not the VSV-G-mediated endocytosis, can lead to viral DNA synthesis and nuclear migration. The viral particles entering through the endocytotic pathway were destroyed within 1–2 days. These results indicate that the VSV-G-mediated endocytotic pathway, although active in transformed cells, is defective and is not a pathway that can establish HIV latent infection of primary resting T cells. Our results highlight the importance of the genuine HIV envelope and its signaling capacity in the latent infection of blood resting T cells. These results also call for caution on the endocytotic entry model of HIV-1, and on data interpretation where the VSV-G-pseudotyped HIV was used for identifying HIV restriction factors in resting T cells.
| While receptor-mediated viral endocytosis or fusion with the cell membrane can be achieved through multiple surface molecules, the repetitious selection of two chemokine receptors, CCR5 or CXCR4, as the main HIV entry coreceptor implies an urgent viral need to exploit the chemotactic process in the immune system. Cytoskeletal rearrangement and cell migration are the primary consequences of chemotactic signaling. Nevertheless, previously published data demonstrated that depriving the virus of its signaling ability conferred higher infectivity through VSV-G-mediated endocytotic entry in transformed cells. We revisited the issue of chemokine coreceptor signaling and the role of cortical actin in HIV-1 latent infection of resting CD4 T cells, in which the virus can establish latency with a potential for productive replication upon T cell activation. Our results confirmed that only the genuine HIV-1 envelope protein, but not VSV-G, is capable of mediating latent infection of resting CD4 T cells. These findings highlight the importance of the HIV envelope and its signaling capacity in HIV infection of its natural target cells.
| Binding of the HIV envelope to its receptors, CD4 and the chemokine coreceptor, CCR5 or CXCR4, triggers sequential fusion and entry events [1],[2],[3],[4]. Fusion is believed to occur directly at the plasma membrane [5],[6],[7],[8],[9], but fusion in endosomes has also been proposed recently [10]. It has been known that HIV can enter cells through endocytosis, but the virion particles entering through this pathway appear to be trapped and subsequently destroyed [11],[12]. This endosomal degradation can be rescued either by blocking the acidification of the endosomal compartments [11],[12] or by pseudotyping the HIV virion with the vesicular stomatitis virus glycoprotein (VSV-G) [13],[14]. The VSV-G-pseudotyped HIV escapes from endosomes and is highly infectious, giving the virus 20- to 130-fold higher infectivity [15],[16]. The ease of producing high-titer virus through VSV-G pseudotyping has made the method very popular for manufacturing viral stock used for gene delivery, drug screening, and the identification of cellular genes and factors involved in HIV replication [13],[17],[18],[19].
Nevertheless, the VSV-G-pseudotyped viruses are not identical to the genuine HIV particles. For example, the HIV Nef protein, a critical factor involved in viral pathogenesis [20], no longer plays an important role in the infection by the VSV-G-pseudotyped virus [15]. Nef has been known to enhance viral infectivity by a factor of 4 to 40 [21],[22]. This positive effect of Nef on viral infectivity appears to be at an early step post entry, such as uncoating or reverse transcription [23],[24],[25]. Nef itself does not directly affect reverse transcription, since Nef-defective virions display normal levels of endogenous reverse transcriptase activity [25]. It is likely that this early activity of Nef is connected to cortical actin in some way. For example, when cells were treated with actin inhibitors, the effect of Nef on viral replication was lost [26]. This is also consistent with the fact that the VSV-G-pseudotyped virus circumvents the cortical actin; thus, the impact of Nef on viral infectivity is forfeited most likely because of the lack of interaction with the actin cortex [15].
The VSV-G-pseudotyped HIV also does not engage CD4 and CCR5 or CXCR4, and is deprived of the ability to transduce signals through these receptors [27],[28],[29]. These intracellular signaling cascades, particularly those transduced from the chemokine coreceptors, have been suggested to be unnecessary for viral fusion, entry, or the subsequent steps of viral replication in transformed cell lines [30],[31],[32],[33],[34],[35],[36],[37],[38],[39]. However, recently, several reports have suggested a requirement for CD4 receptor signaling to mediate viral fusion and entry [40],[41],[42],[43]. We have also observed an absolute requirement for CXCR4 signaling in HIV-1 latent infection of resting CD4 T cells [44] and demonstrated that HIV-1 relies on viral envelope and the Gαi-dependent signaling from CXCR4 to activate a cellular actin-depolymerizing factor, cofilin, to increase the cortical actin dynamics for viral intracellular migration [44]. Given that the VSV-G-pseudotyped HIV infects cells in the absence of receptor signaling, we performed parallel studies in which the VSV-G-pseudotyped HIV was used to infect both transformed and resting CD4 T cells to understand possible alternative pathways that the VSV-G-pseudotyped HIV-1 may employ to establish latent infection of resting CD4 T cells. Surprisingly, the VSV-G-pseudotyped HIV-1 exhibited a highly diminished ability to initiate viral DNA synthesis and nuclear migration in resting T cells, which is in striking contrast to the high efficiency of VSV-G to mediate HIV infection of transformed cells. The viral particles entering through the endocytotic pathway were destroyed within 1–2 days in resting T cells. These results indicate that the VSV-G-mediated endocytotic pathway, although active in transformed T cells, is defective and not a pathway that can establish HIV latent infection of primary CD4 T cells. These results highlight the importance of the genuine HIV envelope and its signaling capacity in the latent infection of primary resting T cells.
We compared the infectivity of HIV-1 carrying either the HIV envelope (Wt) or the VSV glycoprotein (VSV-G). Both viruses were produced in parallel using the same cell culture and transfection conditions (Figure 1A). Following harvesting of viral particles, an equal p24 level of both viruses was used to infect a transformed T cell line, CEM-SS. Viral replication was monitored by p24 release. As shown in Figure 1B, we observed faster and stronger replication of the VSV-G- pseudotyped virus, which reached a level approximately 30 fold higher (at 48 hours) than the wild-type HIV-1. Our result was consistent with previous reports showing that the VSV-G-pseudotyped virus was 20 to 130 times more infectious than wild-type HIV-1 [15],[16]. It is likely that without the limitation of HIV receptors, much more productive viral entry may occur through the VSV-G-mediated endocytosis, resulting in a much higher level of viral replication. In addition, the faster replication kinetics of the VSV-G-pseudotyped virus is likely a result of faster entry and nuclear migration.
We also compared viral early processes after entry by following viral DNA synthesis and nuclear migration. We infected cells using an equal TCID50 dosage instead of an equal p24. Although more Wt particles were used (based on p24), infection with an equal TCID50 ensured that the productive viral processes such as viral DNA synthesis and nuclear migration would occur at comparable levels within the viral population in each case. The TCID50 of both viruses was measured on a Rev-dependent indicator cell, Rev-CEM, as previously described [45]. As shown in Figure 2, at 2 hours post infection, viral DNA synthesis was measured, and the VSV-G-pseudotyped HIV synthesized only approximately 20% of viral DNA in comparison with the wild-type virus (Figure 2A), probably either because fewer of the VSV-G-pseudotyped particles enter the cells or because these particles are less efficient at mediating viral DNA synthesis. We also followed viral nuclear migration at later time points using viral 2-LTR circles as a surrogate. As shown in Figure 2B, the VSV-G-pseudotyped virus produced slightly more 2-LTR circles than the wild-type HIV-1, and the relative ratio of 2-LTR circle to total viral DNA was approximately 7 times higher in the VSV-G-mediated infection (Figure 2C). These data suggest that in transformed T cells, the VSV-G-mediated endocytotic entry is much more efficient in delivering viral DNA into the nucleus. Even with a lower amount of viral DNA synthesized initially, a higher percentage of these DNA molecules entered the nucleus. On the other hand, in the wild-type infection, even with more viral DNA synthesis, a lower percentage of viral DNA molecules can enter the nucleus. These results are consistent with a model [46] in which the cortical actin plays an important role in viral reverse transcription [47], but the actin cortex also serves as a natural barrier for viral intracellular migration [44].
In contrast to the VSV-G-mediated endocytotic entry, the HIV envelope-mediated fusion and entry requires specific interaction with CD4 and the chemokine coreceptor, CCR5 or CXCR4. These receptors not only mediate fusion but also transduce signals upon gp120 binding [27],[48],[49]. In particular, signals transduced from the chemokine coreceptor CXCR4 have recently been shown to be essential for HIV-1 latent infection of resting CD4 T cells [44]. Thus, we examined the ability of VSV-G-pseudotyped HIV-1 to establish latent infection of resting CD4 T cells in the absence of HIV coreceptor signaling. Resting CD4 T cells were purified from the peripheral blood of healthy donors by negative depletion (Figure 3A). Cells were rested overnight, and then infected with an equal p24 level of the VSV-G-pseudotyped HIV-1 or the wild-type HIV-1. Following infection, cells were washed and incubated for 5 days in the absence of T cell activation. During this incubation, productive viral replication does not occur. However, viral replication remains inducible upon T cell activation [44]. As a control, cells were also pre-activated for 1 day with antibody stimulation of the CD3/CD28 receptors (Figure 3B) and then identically infected. As shown in Figure 3C, in CD3/CD28 pre-activated T cells, productive viral replication occurred, and the VSV-G-pseudotyped viral replication was approximately 10-fold greater (48 h.p.i) than that of the wild-type virus. This is similar to the VSV-G-pseudotyped viral replication in transformed T cells (Figure 1). However, in latently infected resting CD4 T cells, when cells were activated at day 5 post infection, only the wild-type virus but not the VSV-G-pseudotyped HIV-1 was induced to replicate (Figure 3D). This was strikingly different from the 10 to 30-fold higher replication capacity of the VSV-G- pseudotyped virus in pre-activated and transformed T cells (Figure 1 and Figure 3C). These results were repeated using CD4 T cells from another donor with AZT added to limit viral replication to a single cycle (Figure 3E and 3F). Reproducibly, only the wild-type virus but not the VSV-G-pseudotyped virus was able to replicate following activation of resting T cells at day 5 (Figure 3F), even though the VSV-G-pseudotyped virus replicated to a 30-fold higher level (48h.p.i) in pre-activated T cells (Figure 3E).
We followed the steps for viral infection of resting T cells. Using a sensitive Nef-luciferase-based entry assay [50], we detected Wt viral entry into both resting and activated T cells (Figure 4A), although the entry into resting T cells was significantly lower. However, we could not detect viral entry into both resting and activated T cells in the VSV-G-pseudotyped virus infection, although we detected the entry of the VSV-G-pseudotyped virus into CEM-SS cells using the identical infection condition (Figure 4A). Since the VSV-G-pseudotyped HIV-1 can replicate in activated T cells (Figure 4C and 4E), these results suggested that the Nef-luciferase-based entry assay may not have the sensitivity to measure the VSV-G-mediated entry in primary T cells, either resting or activated. Thus, we used an alternative method to detect viral entry by measuring intracellular p24 following infection. Cells were infected for 2 hours, trypsinized, washed, and then lysed for intracellular p24. As shown in Figure 4B, we observed a comparable level of intracellular p24 in resting T cells infected with the wild-type or the VSV-G-pseudotyped HIV-1. We also observed a higher level of intracellular p24 in HIV-1(VSV-G)-infected active T cells (Figure 4B). These results suggested that entry of virion particles was similar in resting T cells infected with the wild-type HIV-1 or HIV-1(VSV-G).
We then followed the course of viral DNA synthesis and nuclear migration in resting CD4 T cells after infection. Unstimulated resting CD4 T cells from another donor were infected with an equal TCID50 dose of the VSV-G-pseudotyped HIV-1 or the wild-type virus. After washing away free viruses at 2 hours post infection, cells were continuously incubated without activation for 5 days, and then activated at day 5 with CD3/CD28 stimulation to initiate viral replication (Figure 4C). When viral DNA synthesis was analyzed, we did not observe viral DNA synthesis above the initial background (Figure 4D) in cells infected with the VSV-G-pseudotyped virus at any time point post infection, whereas in cells infected with the wild-type virus, we observed the typical course of viral DNA synthesis in which viral DNA synthesis proceeds slowly and usually peaks at day 2, recedes at day 3, and then increases again following T cell activation [44]. When viral 2-LTR circles were measured, in cells infected with the VSV-G-pseudotyped virus, we also could not detect 2-LTR circles at any time point, even after CD3/CD28 stimulation at day 5, whereas in cells infected with the wild-type virus, 2-LTR circles were detected and the copy number increased with time (Figure 4E). We repeated these experiments using resting CD4 T cells from another donor. This time, resting cells were infected with an equal p24 level of both viruses. We observed similar results. As shown in Figure 4F, in cells infected with the VSV-G-pseudotyped virus, the initial viral DNA detected (0.1 to 0.5 day in Figure 4F) diminished with time, and no 1-LTR circles can be detected at any time point post infection, whereas in the wild-type infected cells, the syntheses of both viral DNA and 1-LTR circles were obvious (Figure 4F and 4G). Based on these results, we concluded that in resting CD4 T cells, only the HIV envelope-mediated entry but not the VSV-G-mediated endocytosis can lead to viral DNA synthesis and nuclear migration, which are a prerequisite for the establishment of HIV latent infection of resting CD4 T cells [44].
We also measured the decay kinetics of the VSV-G-pseudotyped HIV-1 in resting CD4 T cells. Unstimulated resting CD4 T cells were infected with an equal p24 level of both viruses. Following infection for 2 hours, cell-free viruses were washed away. Infected cells were then activated immediately or activated at day 1, 3, or 5 post infection. As a control, resting cells were also pre-activated with CD3/CD28 for 1 hour and then identically infected. As shown in Figure 5, in CD3/CD28 pre-activated CD4 T cells, both the VSV-G-pseudotyped HIV-1 and the wild-type virus replicated after infection (Figure 5A1 and 5B1). In resting T cells, when cells were activated immediately after infection and washing (2 hours post infection), viral replication was also initiated from both the VSV-G- pseudotyped HIV-1 and the wild-type virus (Figure 5A2 and 5B2). These results suggest that entry of the VSV-G-pseudotyped virus into resting T cells occurs, and viral replication can be rescued if cells are activated immediately. However, when resting cells were left unactivated, after 1 day, the replication of the VSV-G- pseudotyped virus following activation was greatly diminished (Figure 5A3), and no viral replication could be initiated after 3 days (Figure 5A4 and 5A5). This was in great contrast to the wild-type HIV infection of resting T cells, in which the capacity of HIV to replicate following activation increased with time (Figure 5B1 to 5B5). The highest viral replication occurred after 5 days of incubation. These data, in combination with the results in Figure 4, suggest that in resting CD4 T cells, the VSV-G-mediated endocytotic entry does not lead to a productive pathway, and the viral particles are trapped in cells and subsequently destroyed within 1–2 days. Our data are also consistent with a previous study demonstrating that the VSV-G-pseudotyped HIV-1 has a half-life of only 1–2 days in resting CD4 T cells [51]. The increasing ability of the wild-type HIV-1 to replicate following incubation has also been observed previously [52],[53],[54]. Although HIV does not directly replicate in resting CD4 T cells, the viral envelope-mediated entry establishes an active process that enhances the ability of HIV to replicate following T cell activation. This capacity has been attributed to the synthesis of Nef, which lowers the threshold required for the activation of resting CD4 T cells [52],[53],[54],[55],[56]. Certainly, our data confirmed these previous findings and further indicated that only the genuine HIV envelope protein but not the VSV-G can deliver the virus into the nucleus, where the subsequent action of Nef can occur.
In contrast to resting T cells, in transformed cell lines, the VSV-G-mediated entry is very efficient in mediating HIV infection. This fact has prompted a major argument that HIV may predominately fuse in the endosome rather than at the plasma membrane [10]. Microscopic imaging tracking the behaviors of the majority of MLV particles suggested that the HIV-1 envelope-pseudotyped virus entered cells predominantly through endocytosis [10]. Indeed, dynasore, a dynamin-dependent endosomal scission inhibitor, was shown to inhibit viral replication [10], supporting the model that the endosomal fusion is associated with a productive pathway. Nevertheless, this mode of entry is in conflict with numerous previous observations suggesting that genuine HIV envelope-associated endocytotic entry, although occurring at a significant scale, does not naturally lead to productive infection [5],[6],[7],[9],[11],[12]. For example, inhibition of the endosomal/lysosomal functionality can spare HIV from degradation and enhance viral replication [11],[12], demonstrating that the endosomal viruses are normally directed for degradation. In addition, the rate of CD4 or CCR5 endocytosis does not appear to affect viral entry or replication [6],[8],[9], supporting direct viral fusion at the plasma membrane. Nevertheless, the endocytosis entry as proposed [10] is an attractive alternative pathway. If proven biologically, it would require significant remodeling of the role of the cortical actin in viral entry and early post-entry steps. The involvement of the cortical actin in early endocytosis is largely limited to membrane scission of clathrin-coated pits [57]. This process does not involve direct contact between the cortical actin and the viral particles. If there is any viral contact with actin, it would be in the cytoplasm following endosomal fusion. This interaction may also affect reverse transcription and nuclear migration, but such effects would occur at different levels. The issue of entry is so critical in the understanding of the role of the cortical actin in HIV biology that we felt compelled to revisit some of the key biological evidence - in particular, the inhibition of HIV replication by the dynamin-dependent endosomal fusion inhibitor, dynasore. Dynamins are a group of fundamental proteins involved in multiple cellular processes such as vesicle transport, cytokinesis, organelle division and cell signaling (for a review, see [58]). To minimize possible cytotoxicity from prolonged inhibition of fundamental cellular proteins, we treated cells only briefly with dynasore during viral infection. Viruses that failed to enter were subsequently washed away along with the drug. We also used the Rev-dependent indicator cell, Rev-CEM [45], to measure dynasore inhibition, instead of simply using p24 ELISA, which by itself is not capable of distinguishing between HIV-specific inhibition and general drug cytotoxicity. Additional advantages of using Rev-CEM are its high specificity and the ability to distinguish subpopulations of cells by flow cytometry so that non-specific cytotoxicity can be excluded [59]. As shown in Figure 6A, we first tested dynasore in the inhibition of the VSV-G-pseudotyped HIV-1 replication and observed dosage-dependent inhibition of viral replication. At 80 µM, dynasore moderately decreased the GFP+ population from 16.1% to 11.5%; at 8 µM, dynasore also slightly decreased the GFP+ population; at 0.8 µM, dynasore minimally affected viral infection. However, when dynasore was used on identically treated cells that were infected with HIV-1, we did not observe similar dosage-dependent inhibition. Even at 80 µM, dynasore minimally affected HIV-1 infection (Figure 6B). These results demonstrate a clear distinction between the VSV-G-mediated endocytotic entry and the HIV-1-envelope-mediated entry in mediating productive viral replication.
In this report, we demonstrated a fundamental difference between the HIV-1 envelope and VSV-G in mediating HIV-1 latent infection of primary resting CD4 T cells, namely that only the HIV-1 envelope but not VSV-G is capable of supporting HIV latent infection of resting T cells. The block to the VSV-G-pseudotyped virus in resting T cells was most obvious at post-entry steps such as viral DNA synthesis and nuclear migration. The virion particles trapped in cells were subsequently destroyed within 1–2 days in resting T cells. These results demonstrated the importance of the genuine HIV envelope in mediating latent infection of resting T cells.
Previously, we demonstrated a critical function of the HIV-1 envelope in mediating CXCR4 signaling and promoting the cortical actin dynamics necessary for HIV latent infection of resting T cells [44]. We also proposed a dual function of F-actin in which the actin cortex serves as an anchorage for reverse transcription and as a vehicle for the delivery of the preintegration complex across the cortical actin through actin treadmilling [44],[46]. At least four HIV proteins in the preintegration complex are known to interact with actin; the viral nuclear capsid [60],[61],[62],[63], the large subunit of the reverse transcriptase [64], the integrase [65], and Nef [66]. We have also shown that blocking actin polymerization with Jasplakinolid (120 nM) or Latrunculin A (2.5 µM) inhibits viral DNA synthesis or HIV latent infection. Conversely, triggering actin polymerization through cofilin shRNA knockdown enhances viral DNA synthesis [44]. These previous results and other studies [47] are consistent with the findings in this study, in which the VSV-G-mediated entry that bypasses the cortical actin led to less viral DNA synthesis in transformed cells (Figure 2). The VSV-G-pseudotyping also resulted in a lack of the slow viral DNA synthesis that is normally seen in HIV-1 latent infection of resting T cells (Figure 4D and 4F). Viral nuclear DNA was also completely missing in the VSV-G-mediated entry in resting cells (Figure 4E and 4G). These results suggest that the VSV-G-pseudotyped particles may be delivered to a different cytoplasmic location and enter the nucleus by a different route, one that is normally highly effective in transformed or metabolically active cells but defective in resting T cells.
Our results are consistent with a recent independent study demonstrating that only the CXCR4-tropic HIV-1 envelope but not VSV-G can support lentiviral vectors to deliver genes into resting CD4 T cells [67]. In this study, Agosto and co-authors also found that viral DNA synthesis was greatly diminished in resting CD4 T cells infected with the VSV-G-pseudotyped lentiviral particles. Nevertheless, the limitation on viral infection was specifically attributed to the lack of viral entry and fusion in the VSV-G-mediated infection of resting T cells [67]. Our results suggested that the restriction was likely at an unknown post-entry step such as endosomal fusion, uncoating, or reverse transcription. The discrepancy in conclusions arises from different explanations of the data acquired from entry and fusion assays. Both Agosto and co-authors [67] and we observed an absolute lack of entry signals in HIV-1(VSV-G)-infected resting T cells, using two different entry assays. However, these assays, particularly the BlaM-Vpr-based fusion assay [68] may not be appropriate for the measurement of VSV-G-mediated fusion in resting T cells. It is possible that if the VSV-G-pseudotyped virus is trapped in a compartment, or is going through a degradation process with a half-life of only 1 day [51], the BlaM substrate which takes about 12–18 hours to load may not be able to access or sufficiently react with the enzyme. Given this lack of mechanistic clarity of how these enzyme-tagged particles are delivered through VSV-G in resting T cells, we did not feel confident that conclusions can be drawn based on a fusion assay. Thus, we drew our conclusions relying on multiple results. Firstly, we detected a comparable intracellular p24 level in resting T cells infected with Wt or HIV-1(VSV-G) (Figure 4B). Secondly, the VSV-G-pseudotyped virus can be partially rescued if resting T cells were activated within 1 day of infection, indicating some viral entry into the cells (Figure 5A2). Thirdly, low levels of viral DNA were also detected at early time (2 hours, Figure 4F), indicating again that there were some levels of entry. Given that the VSV-G-pseudotyped viruses are 20 to 130-fold more infectious than the wild-type HIV-1, these initial viral activities should give rise to a measurable level of viral replication, but they did not.
The failure of the VSV-G-mediated entry to establish latent infection of resting T cells is not currently understood. It is possible that the cellular environment in resting T cells may not permit viral fusion in endosomes. Alternatively, successful endosomal fusion may occur, but the quick delivery of viral particles into the cytosol may be detrimental [69], likely due to the possible restrictive environment of resting cells [17],[70] or a lack of cytosolic factors for uncoating [71] DNA synthesis, or nuclear localization. Our attempts to rescue the VSV-G-pseudotyped virus by changing the intracellular PH were not successful (data not shown). Pre-stimulation of the CD4 and CXCR4 receptors with gp120 or antibodies also could not rescue the VSV-G-pseudotyped virus in resting T cells (data not shown), although these pre-stimulations enhanced the wild-type HIV replication several fold following T cell activation [44]. These results are consistent with the fact that the positive benefits of viral receptor signaling are only associated with gp120-mediated entry but not with the VSV-G-mediated endocytosis that circumvents the cortical actin.
The high efficiency of VSV-G to mediate endosomal escape and HIV replication in transformed cells has led to the misconception that the VSV-G-pseudotyped HIV should be as effective as the wild-type HIV for latent infection of resting T cells [72],[73]. Several previous studies have also used the VSV-G-pseudotyped virus to identify restriction factors in resting T cells [17],[74]. Our results suggest that these data need to be interpreted cautiously. Apparently, the VSV-G-mediated entry does not experience the same intracellular environment as HIV does, and cannot lead to the establishment of latent infection in resting T cells. Thus, those previously identified cytoplasmic restriction factors may or may not directly affect HIV infection. Interestingly, a recent imaging study demonstrated a direct dependence of active viral nuclear migration on F-actin, since actin inhibitors diminished the nuclear concentration of the preintegration complex (PIC) (Dr. Thomas Hope, personal communication). This study raises the possibility that PIC may be associated with F-actin up to the nucleus [75],[76]. Given that viruses usually use F-actin for short-distance travel, and the cytoplasmic space between the cortical actin and the nucleus is relatively thin in T cells, it is possible that the cytosolic exposure of PIC in T cells is minimal.
All protocols involving human subjects were reviewed and approved by the GMU IRB. Informed written consents from the human subjects were obtained in this study.
Plasmid pNL4-3 was kindly provided by Dr. Malcolm Martin [77]. The env mutant, pNL4-3(KFS), was kindly provided by Dr. Eric Freed [78]. pHCMV-G that expresses the vesicular stomatitis virus glycoprotein has been described previously [79]. pNLΔΨEnv was constructed by inserting the env gene of HIV-1NL4-3 into the lentiviral vector pNL-RRE-SA [80]. The packaging signal was further deleted by cutting with KasI plus BssHII and re-ligating.
HIV-1NL4-3 was generated by transfection of plasmid pNL4-3 into HEK293T cells using lipofectamine 2000 (Invitrogen) as described previously [80]. The VSV-G-pseudotyped virus, HIV-1(VSV-G), was produced by cotransfection of HEK393T cells (3×106) with 10 µg of pHCMV-G and 10 µg of plasmid pNL4-3(KFS). The HIV-1 envelope-typed virus, HIV-1(Env), was produced by cotransfection of HEK293T cells with 10 µg of pNLΔΨEnv and 10 µg of pNL4-3(KFS). Viral supernatant was harvested at 48 hours post cotransfection, centrifuged for 15 minutes at 500×g to remove cellular debris, filtered through a 0.45 µm filter, treated with Benzonase (Novagen) (250 U/ml) at 37°C for 15 minutes, and then stored at -80°C. Levels of p24 in viral supernatant were measured using the Perkin Elmer Alliance p24 antigen ELISA Kit (Perkin Elmer). Plates were kinetically read using an ELx808 automatic microplate reader (Bio-Tek Instruments) at 630 nm. Viral titer (TCID50) was determined on the Rev-dependent GFP indicator cell, Rev-CEM [45],[81]. CEM-SS cells from Dr. Peter L. Nara [82] were obtained through the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH. All cells were cultured in RPMI 1640 medium supplemented with 10% heat-inactivated fetal bovine serum (Invitrogen), penicillin (50 U/ml), and streptomycin (50 mg/ml).
Peripheral blood mononuclear cells (PBMC) were obtained from healthy donors at the Student Health Center, George Mason University (GMU), Fairfax, VA. Resting CD4 T cells were purified by two rounds of negative selection as previously described [54]. Purified cells were cultured in RPMI 1604 medium supplemented with 10% heat-inactivated fetal bovine serum (Invitrogen), penicillin (50 U/ml), and streptomycin (50 µg/ml) overnight before infection or treatment. For activation of resting CD4 T cells with PHA (3 µg/ml) (Sigma) plus IL-2 (100 U/ml) (Roche Applied Science), cells were cultured in the presence of these agents for 12 hours. For infection, CD4 T cells were incubated with the virus for 2 hours and then washed twice with medium to remove unbound virus. Infected cells were resuspended in fresh RPMI 1604 medium supplemented with 10% heat-inactivated fetal bovine serum at a density of 106 per ml and incubated for 5 days without stimulation. Cells were activated at day 5 with anti-CD3/CD28 magnetic beads at 4 beads per cell. For the viral replication assay, 10% of infected cells were taken at days 1, 3, 5, 6, 7, 8, and 9 post infection. Cells were pelleted and the supernatant was saved for p24 ELISA.
Monoclonal antibodies against human CD3 (clone UCHT1) and CD28 (clone CD28.2) were purchased from BD Pharmingen (BD Biosciences). For conjugation, antibodies were conjugated with 4×108 Dynal beads (Invitrogen) for 30 minutes at room temperature. Free antibodies were washed away with PBS-0.5% BSA. The conjugated magnetic beads were resuspended in 1 ml of PBS-0.5% BSA. For stimulation of resting CD4 T cells, antibody-conjugated beads were washed twice and then added to cell culture and rocked for 5 minutes.
Resting CD4 T cells or CD3/CD28-stimulated cells (106) were used for the analysis. Before staining, magnetic beads were removed by incubating with DNase I releasing buffer as recommended by the manufacturer. Cells were suspended in 1 ml of 0.03% saponin in PBS and then incubated in 20 µM 7-amino-actinomycin D (Sigma) for 30 minutes at room temperature in the dark. Cells were kept on ice for at least 5 minutes, pyronin Y (Sigma) was added to a final concentration of 5 µM, and the cells were then incubated for 10 minutes on ice. Stained cells were directly analyzed by flow cytometry on a FACS (Becton Dickinson FACSCalibur).
Plasmid pCDNA3-Nef-Luc was kindly provided by Dr. Robert Davey [50]. Viruses containing Nef-luciferase was produced as described previously [50]. Briefly, 293T cells cultured in a 10 cm petri dish were cotransfected with 10 µg pNL4-3 plus 10 µg of pcDNA3-Nef-luc, or with 10 µg pNL4-3(KFS) plus 7.5 µg pcDNA3-Nef-luc plus 2.5 µg pHCMV-G, using lipofectamine 2000 (Invitrogen) as recommended by the manufacturer. Viruses were harvested at 48 hours post cotransfection and filtered through a 0.45 µM filter. For entry assays, cells (1×106) were infected with 200 ng of Nef-luciferase containing viruses at 37°C for 2 hours, and then washed three times with medium. Cells were resuspended in 0.1 ml of luciferase assay buffer (Promega) and luciferase activity was measured in live cells using a GloMax-Multi Detection System (Promega).
Total cellular DNA was purified using the Wizard SV Genomic DNA Purification System as recommended by the manufacturer (Promega). The detection of viral late DNA and 1-LTR-circles by PCR was performed as described previously [83]. Briefly, for viral late DNA, forward primer: 5′ GGTTAGACCAGATCTGAGCCTG 3′ and reverse primer: 5′ TTAATACCGACGCTCTCGCACC 3′ were used. PCR was carried out in 1×Ambion PCR buffer, 125 µM dNTP, 50 pmol each primer, 1 U SuperTaq Plus (Ambion) with 30 cycles at 94°C for 20 seconds, 68°C for 40 seconds. For detection of 1-LTR circle, primers LTR-nef2 (5′ TGGGTTTTCCAGTCACACCTCAG 3′) and LTR-gag (5′ GATTAACTGCGAATCGTTCTAGC 3′) were used. The reaction was carried out in 1×Ambion PCR buffer, 1.5 nM Mg2+, 125 µM dNTP, 50 pmol each primer, 1 U SuperTaq Plus (Ambion) with 35 cycles at 94°C for 20 seconds, 68°C for 90 seconds. Real-time PCR quantification of viral late DNA and 2-LTR circles was also performed as described previously [44],[84], using 300 nM primers and 200 nM probes. The DNA standard used for both late DNA and 2-LTR circle quantification was constructed using a plasmid containing a complete 2 LTR region (pLTR-2C); the plasmid was cloned by amplification of infected cells with 5′-TGGGTTTTCCAGTCACACCTCAG-3′ and 5′-GATTAACTGCGAATCGTTCTAGC-3′. Measurement was run in triplicate ranging from 1 to 106 copies of pLTR-2C mixed with DNA from uninfected cells.
FITC-phalloidin staining of F-actin has been described previously [44]. Stained cells were imaged using a Zeiss Laser Scanning Microscope, LSM 510 META, with a 40 NA 1.3 or 60 NA 1.4 oil DIC Plan-Neofluar objective. Images were processed and analyzed by LSM 510 META software.
Dynasore monohydrate (Sigma) was dissolved in DMSO. Following dynasore treatment, infection, and washing, cells were incubated for 48 hours, and then 500 µl cells were removed and stained with 2 µg/ml propidium iodide solution (Fluka) for 5 minutes at room temperature. Following incubation, cells were analyzed using the FACSCalibur (BD Biosciences). Data analysis was performed using CellQuest (BD Biosciences).
|
10.1371/journal.pmed.1002305 | Impact evaluation of different cash-based intervention modalities on child and maternal nutritional status in Sindh Province, Pakistan, at 6 mo and at 1 y: A cluster randomised controlled trial | Cash-based interventions (CBIs), offer an interesting opportunity to prevent increases in wasting in humanitarian aid settings. However, questions remain as to the impact of CBIs on nutritional status and, therefore, how to incorporate them into emergency programmes to maximise their success in terms of improved nutritional outcomes. This study evaluated the effects of three different CBI modalities on nutritional outcomes in children under 5 y of age at 6 mo and at 1 y.
We conducted a four-arm parallel longitudinal cluster randomised controlled trial in 114 villages in Dadu District, Pakistan. The study included poor and very poor households (n = 2,496) with one or more children aged 6–48 mo (n = 3,584) at baseline. All four arms had equal access to an Action Against Hunger–supported programme. The three intervention arms were as follows: standard cash (SC), a cash transfer of 1,500 Pakistani rupees (PKR) (approximately US$14; 1 PKR = US$0.009543); double cash (DC), a cash transfer of 3,000 PKR; or a fresh food voucher (FFV) of 1,500 PKR; the cash or voucher amount was given every month over six consecutive months. The control group (CG) received no specific cash-related interventions. The median total household income for the study sample was 8,075 PKR (approximately US$77) at baseline. We hypothesized that, compared to the CG in each case, FFVs would be more effective than SC, and that DC would be more effective than SC—both at 6 mo and at 1 y—for reducing the risk of child wasting. Primary outcomes of interest were prevalence of being wasted (weight-for-height z-score [WHZ] < −2) and mean WHZ at 6 mo and at 1 y.
The odds of a child being wasted were significantly lower in the DC arm after 6 mo (odds ratio [OR] = 0.52; 95% CI 0.29, 0.92; p = 0.02) compared to the CG. Mean WHZ significantly improved in both the FFV and DC arms at 6 mo (FFV: z-score = 0.16; 95% CI 0.05, 0.26; p = 0.004; DC: z-score = 0.11; 95% CI 0.00, 0.21; p = 0.05) compared to the CG. Significant differences on the primary outcome were seen only at 6 mo. All three intervention groups showed similar significantly lower odds of being stunted (height-for-age z-score [HAZ] < −2) at 6 mo (DC: OR = 0.39; 95% CI 0.24, 0.64; p < 0.001; FFV: OR = 0.41; 95% CI 0.25, 0.67; p < 0.001; SC: OR = 0.36; 95% CI 0.22, 0.59; p < 0.001) and at 1 y (DC: OR = 0.53; 95% CI 0.35, 0.82; p = 0.004; FFV: OR = 0.48; 95% CI 0.31, 0.73; p = 0.001; SC: OR = 0.54; 95% CI 0.36, 0.81; p = 0.003) compared to the CG. Significant improvements in height-for-age outcomes were also seen for severe stunting (HAZ < −3) and mean HAZ. An unintended outcome was observed in the FFV arm: a negative intervention effect on mean haemoglobin (Hb) status (−2.6 g/l; 95% CI −4.5, −0.8; p = 0.005). Limitations of this study included the inability to mask participants or data collectors to the different interventions, the potentially restrictive nature of the FFVs, not being able to measure a threshold effect for the two different cash amounts or compare the different quantities of food consumed, and data collection challenges given the difficult environment in which this study was set.
In this setting, the amount of cash given was important. The larger cash transfer had the greatest effect on wasting, but only at 6 mo. Impacts at both 6 mo and at 1 y were seen for height-based growth variables regardless of the intervention modality, indicating a trend toward nutrition resilience. Purchasing restrictions applied to food-based voucher transfers could have unintended effects, and their use needs to be carefully planned to avoid this.
ISRCTN registry ISRCTN10761532
| Cash-based interventions (CBIs) are being increasingly used during humanitarian emergencies as an alternative to food-based interventions to prevent wasting.
However, due to the lack of evidence available on the impact of CBIs in these settings, it is unclear what are the best ways to implement them.
There is a lack of evidence to show that CBIs are effective in reducing the risk of wasting in young children in humanitarian aid settings.
There is limited, and sometimes confusing, evidence about the impact of different CBI modalities on nutritional status, and no evidence of a longer-term impact.
Our study evaluated the effects of three different CBI modalities on nutritional outcomes in children under five years of age at six months and at one year.
We conducted a four-arm parallel longitudinal cluster randomised controlled trial in 114 villages (2,496 households with 3,584 children) in Dadu District, Pakistan.
The interventions included two different-sized unconditional grants (standard cash [SC] and double cash [DC]), a fresh food voucher (FFV), and a control group (CG).
We saw a 48% decrease in the odds of a child being wasted in the DC arm and an improvement in ponderal growth in the FFV (+16 weight-for-height z-score [WHZ]) and DC (+11WHZ) arms at six months only.
We saw a negative impact on haemoglobin status in children in the FFV arm (−2.6 g/l).
All three CBIs resulted in children having a decreased odds of being stunted and an improvement in linear growth at both six months and one year.
The amount of cash given in this setting was important in terms of reducing the odds of wasting, but the effects were only seen at the six month time point.
Regardless of which CBI was received, height growth deteriorated less in the intervention groups than in the CG at six months and at one year, indicating improved nutrition resilience in these children.
Caution is needed when applying restrictions to food vouchers in order to secure a diverse food basket that provides adequate macro- and micronutrients.
| The current global estimate of wasting prevalence is 7.4%, affecting approximately 50 million children under the age of 5 y annually [1]. The World Health Assembly (WHA) 2025 target to reduce and maintain childhood wasting at 5% is unlikely to be met [1]. Globally, attention to child and maternal undernutrition is very high, with agreed targets and impetus through, e.g., the Scaling Up Nutrition (SUN) Movement and the Zero Hunger Initiative, as well as WHA nutrition targets and indicators in the recently framed Sustainable Development Goals. In addition, there is considerable attention being paid to food systems and healthy diets as a potentially sustainable means of preventing high levels of stunting, wasting, and micronutrient malnutrition [2]. According to the 2016 Global Nutrition Report, the overall trend is one of reduction in the prevalence of child undernutrition, though the rate of progress between regions is uneven [3], with the most progress occurring in Asia and the least in sub-Saharan Africa. Asia, however, has the largest numbers of wasted and stunted children [4]. Pakistan presents a particular challenge as the nutritional status of children has shown very little progress over the last 15 y and has, for some nutrition indicators, worsened [5]. This is especially so in Sindh Province, which has the highest prevalence of childhood wasting and stunting in Pakistan [6]. The most recently available population data in Sindh Province indicate that the prevalence of wasting and stunting is 15.4% and 48.0%, respectively, in children under 5 y of age [7]. Levels of anaemia and vitamin A deficiency in Sindh Province have both shown an increase since 2001 [5,6]. In 2011, 73% of children under 5 y of age in Sindh Province were anaemic (haemoglobin [Hb] level < 110 g/l) [6]. Taken together, these statistics indicate an ongoing and serious public health problem.
Previous efforts to improve child and maternal nutrition in Pakistan have been inconsistent, and the coordination needed to develop and implement a coherent nutrition strategy has been weak [5]. However, more recently there have been concerted efforts to develop strategies to tackle undernutrition. For example, the Pakistan Integrated Nutrition Strategy, involving government, bilateral agencies, non-governmental organisations, civil societies, and donors, has been developed, and Pakistan joined the SUN Movement in 2013. Furthermore, the national social safety net system, called the Benazir Income Support Programme (BISP), which uses wealth ranking to select the poorest in the population and provides them with a cash transfer, has started to include child nutrition indicators in its targeting.
Globally, it is believed that cash-based interventions (CBIs), which include cash and food vouchers, offer an alternative to food-based interventions for reducing the risk of wasting during seasonal periods of food insecurity referred to as the “lean season”. Since CBIs are increasingly being incorporated into emergency response programming, more information is needed on the impact of these interventions, particularly where nutrition objectives are established. Whilst there is greater evidence of the impact of CBIs in development aid settings on household dietary diversity and access to health care—through improving household income and protecting household assets [8–10]—there is mixed evidence about whether CBIs improve and protect child growth [9,11], with even less evidence from humanitarian aid settings [12]. As well as this, the available evidence on the effectiveness of different CBI designs in humanitarian aid settings is unclear and potentially conflicting. For example, studies comparing the use of cash versus food vouchers have shown different effects [13,14], while no evidence exists regarding whether different amounts of cash are associated with different levels of impact in preventing wasting in emergencies [15]. Furthermore, no evidence exists on the longer-term impacts of CBIs following an emergency response.
This study compares the nutritional status of children under 5 y of age from households that were allocated to receive either a monthly unconditional cash transfer (one of two amounts), a monthly fresh food voucher (FFV), or a standard package of interventions (the control group [CG]) over six consecutive months. A final round of data was collected 6 mo after the last intervention disbursement to determine any residual impact on nutritional status. We investigated the effect of the different interventions, which were delivered by Action Against Hunger working in Dadu District, Sindh Province, Pakistan, within the context of their Women and Children/Infants Improved Nutrition in Sindh (WINS) programme, funded by the European Union (EU), with further funding from the Directorate-General for European Civil Protection and Humanitarian Aid Operations (DG ECHO). Impact was assessed at two time points: immediately after the final disbursement (at 6 mo) and then 6 mo later (1 y after baseline).
The overarching aim of this study was to evaluate the impact of three CBI modalities on nutritional outcomes in children under 5 y of age from poor and very poor households in Dadu District, Sindh Province, Pakistan, in the context of the lean season. We hypothesized that, compared to the CG, FFVs would be more effective than cash of the same value, and that a higher amount of cash would be more effective than a lower amount of cash at both 6 mo and 1 y in terms of reducing the risk of wasting.
Ethical approval was obtained from the Pakistan National Bioethics Committee and the Western International Review Board. The trial was registered on 26 March 2015 with the ISRCTN registry (ISRCTN10761532). Participating households were enrolled at baseline after providing written informed consent from the household head or the participating mother, father, or primary carer.
Dadu District is largely agrarian, with the economy dependent on crop production, livestock keeping, and agriculture labour. The majority of the population is highly vulnerable to environmental shocks, especially the poorest households, and there is a lack of alternative income sources, which are further constrained by a lack of economic opportunities. Dadu District experiences frequent flooding and droughts, and extreme temperatures (above 45°C). The results from the most recent nutrition survey, conducted in November 2014 in Dadu District, estimated that 14.3% of children aged 6–59 mo were wasted. For our study, initiated at the start of the lean season (May/June) and including poorer households, we expected the baseline prevalence of wasting to be higher.
This was a longitudinal cluster randomised controlled trial, with four parallel arms, conducted among 114 villages, selected from the Action Against Hunger WINS programme database, in Dadu District, Pakistan. The trial design, setting, and characteristics of the study population have been previously described [16].
Households were selected from villages from three agricultural areas sharing similar livelihoods, geography, and access to the same elements of the standard WINS programme. Action Against Hunger provided the initial household lists, and these were further verified and updated by the study research team. Households defined as poor or very poor—using eligibility criteria decided upon by the research team with village participation, and based on ownership of cultivated land and number of goats—and with one or more children aged 6–48 mo were selected. The study was a closed cohort and followed all children in the same eligible households regardless of their baseline anthropometric status.
The study also involved a mixed-methods process evaluation to understand further how intervention implementation may have affected intervention impacts in this setting, and to quantify the causes of any impacts seen. Some of the process evaluation results are presented here, particularly those related to the impacts seen in the study. A further analysis is forthcoming focusing specifically on the pathways that were involved in the main impacts seen.
Three CBIs were implemented: two unconditional cash transfers—a “standard cash” (SC) amount of 1,500 Pakistani rupees (PKR) (approximately US$14) and a “double cash” (DC) amount of 3,000 PKR (approximately US$28)—and one FFV with a cash value of 1,500 PKR (approximately US$14), which could be exchanged for specified fresh foods (fruits, vegetables, milk, and meat) in nominated shops. Action Against Hunger ensured that all FFV villages had good access to these shops, by nominating shops in, or nearby, these villages. All villages were served by at least one nominated shop.
The cash and vouchers were disbursed at the same time every month for six consecutive months. The CG received no additional intervention beyond the basic WINS programme activities (described below) that were provided to all groups. A pure CG was not feasible given WINS programme coverage across Dadu District.
The SC amount was set to equal the amount disbursed by the BISP at the time of the baseline survey. At the time of the study, the purchasing power parity for Pakistan was 0.286 PKR = US$1. The cash and vouchers were disbursed at distribution points on a monthly basis either by mobile banks that travelled to a central location serving some of the participating villages or through central banks that served a number of villages. The FFVs were disbursed to participating households at the village level. All three interventions were delivered with verbal messaging from Action Against Hunger field staff, who were present at all distributions, that children should benefit from the transfers.
All villages had access to the WINS programme, which provided outpatient treatment for children 6–59 mo with severe acute malnutrition (SAM), micronutrient supplementation (children and pregnant and lactating women), and behaviour change communication (BCC). Key BCC messages on the causes of undernutrition, the benefits of exclusive breastfeeding, improved complementary feeding practices, food and water hygiene, handwashing, and sanitation were targeted at mothers. These messages were delivered monthly to all study participants in group sessions by the research mobilisers. Research mobilisers also facilitated data collection activities, such as locating households and setting up times to be available, but were not involved in the data collection itself.
Children identified as severely malnourished during the study period were referred to outpatient treatment. These children were still followed up, with consent from the parents, and were identified in the dataset as to whether or not they received supplementary rations. All parents gave consent, and receipt of supplementary rations was adjusted for in the analysis.
Two of the intervention arms (SC and FFV) were funded by the EU. The DC arm was funded by the DG ECHO. The interventions took place over six consecutive months (July to December 2015).
The primary outcomes were the prevalence of being wasted (weight-for-height z-score [WHZ] < −2) and mean WHZ at 6 mo and at 1 y amongst children less than 5 y.
Secondary outcomes in children were prevalence of SAM (WHZ < −3), mean mid-upper arm circumference (MUAC), prevalence of stunting (height-for-age z-score [HAZ] < −2), prevalence of severe stunting (HAZ < −3), mean HAZ, morbidity, mean Hb concentration, and prevalence of anaemia (Hb < 110 g/l) and severe anaemia (Hb<70 g/l) at 6 mo. Due to the longer-term nature of stunting, the stunting outcomes were also assessed at 1 y. Secondary outcomes for mothers were also assessed, including mean Hb concentration and MUAC. Population cutoffs of 120 g/l and 130 g/l for pregnant and non-pregnant women, respectively, were used to determine levels of anaemia. Body mass index (BMI) was assessed for non-pregnant mothers.
As Global Positioning System mapping is not permitted in Pakistan, the research team carried out a mapping exercise by hand to assess the size of each village and the potential number of eligible households. Only one small village (five households) declined to be included in the study. Because it was not possible to carry out a public randomisation, randomisation was done by the principal investigator (PI) using a random number table to generate the randomisation sequence and then drawing village names from a box. Block randomisation was done, allowing equal distribution of the villages to each arm for small (<40 households), medium (40–85 households), and large (>85 households) villages. The PI had no knowledge of the villages involved and was not involved in the intervention implementation or any data collection. Study participants were enrolled by the data collection team and were not aware which of the interventions they would be getting at enrolment. However, masking of participants was not possible due to the nature of the intervention. The data collection team was different to the cash and voucher disbursement team. The data collection team was responsible for the collection of data and sensitisation of the study recipients to the use of the cash and vouchers. The data collection team was accompanied by local research mobilisers who, as well as facilitating the data collectors in, e.g., locating households, were also responsible for delivering key BCC messages.
The target sample size (approximately 632 households per arm) was calculated to measure a detectable difference of prevalence of being wasted of 7% between the intervention groups and the CG post-intervention [15]. The sample size was also powered to detect a 0.19 WHZ difference between the intervention groups and the CG. This sample size was reached for the SC, FFV, and CG arms. However, for the DC arm the sample size was 600 due to the different funding amounts given for this arm, which did not allow for an equivalent number of households to be included compared to the other three arms. The target sample size was calculated using an estimated intraclass correlation coefficient (ICC) of 0.02 for prevalence of being wasted from an Action Against Hunger nutrition survey in Dadu District. The ICC for prevalence of being wasted for this current study was 0.01.
Quantitative data were collected at baseline and then after each cash and voucher disbursement (6 mo in total), with a final round of data collection 1 y after baseline. Data for Hb were collected (using the HemoCue Hb 201+ System) only at baseline and at 6 mo due to the costs involved. Data for the main impact analysis and findings reported here involved three periods: baseline (May to July 2015), 6 mo after baseline (December 2015), and 1 y after baseline (June/July 2016). Data collected from the months between baseline and 1 y were analysed to illustrate the changes in the prevalence and mean of weight-based indicators during this time. These monthly data will be analysed further in a mediation analysis to be published at a later stage. All questionnaires were translated and administered in the local language, Sindhi. Piloting and back-translation were carried out to ensure that the intended meaning of the questions was retained. Quantitative data were collected using android mobile phones with Open Data Kit software. In order to ensure the quality of the data collected, daily field supervision, meetings with the study coordinator, a mid-term refresher training session, and regular checking of the data were carried out. Data were sent to the ENN PI on a weekly basis for checking.
Qualitative data were collected using focus group discussions (FGDs), key informant interviews, and longitudinal in-depth interviews. Data were collected by a qualified qualitative researcher who conducted two rounds of in-depth interviews with 32 study mothers and 34 FGDs that included study mothers and fathers and other female and male non-participants. Qualitative data were collected using digital dictaphones, and the mp3 files of the recorded interviews were transcribed and translated into English in MS Word and then analysed using a thematic approach. For this analysis, the qualitative data have been used to help interpret the main findings.
Data entry and validation checks were conducted both by the research team and ENN. Analysis was conducted entirely by the PI and was supported and verified by a statistical adviser.
The ZSCORE06 command [17] in STATA (SE version 14; StataCorp) was used to calculate z-scores. WHZ data were coded as missing if WHZ > +5 or WHZ < −5; HAZ data were coded as missing if HAZ > +5 or HAZ < −6. A child’s data were excluded from the analysis if the child was deemed to be a different child to the child enrolled at baseline. Whilst checks were put in place to ensure that the same child was measured every month, in some cases these were not followed. We used as our criteria for exclusion a decrease in height or length of more than 1 cm (measurement error) or an increase of more than 15 cm (considered the maximum height a child could grow in 6 mo).
Proportions, means (standard deviations), and medians (interquartile ranges) are presented for key baseline variables for households, mothers, and children. All effect analyses are intention-to-treat. Results are presented as crude difference-in-differences estimates (DDEs), adjusted linear changes over the study period, and partially and fully adjusted effect sizes at 6 mo and at 1 y with 95% confidence intervals (CIs). In order to account for clustering at the distribution point level, multilevel mixed-effects regression models were used to generate odds ratios (ORs) for binary outcomes and regression coefficients (β) for continuous outcomes. These results at each time point compare the intervention arm to the CG. The intervention distribution point and household were included as random effects. Village size (small, medium, or large), child age at baseline, child sex, and baseline values of the outcome variables were included as fixed effects in all models. Baseline values of the outcome variables were added into the adjusted models to take into consideration any individual variation at baseline. The village size variable was included to account for different village sizes used for block randomisation; child age and sex and baseline values of the outcome variables were included to adjust for potential individual differences at baseline. Sensitivity analyses (with and without adjustment for other baseline characteristics) were carried out to assess whether adjusting for chance residual baseline imbalances significantly altered the results, such as access to the BISP, deworming, and socio-economic status. These baseline characteristics were those observed as dissimilar in terms of a difference in proportion, mean, or median across arms.
To measure the intervention effect, we included an intervention × time term in the model. Significance was defined as p < 0.05. All analyses were carried out in STATA software SE version 14.
The flow of clusters and participants through the trial is shown in Fig 1. Enrolment and baseline data collection started together at the end of May 2015 and continued until the beginning of August 2015. Thirteen eligible households refused to participate at the enrolment stage as permission was not given by the head of household. Twenty-seven households migrated away from their village after enrolment (CG = 11, DC = 4, FFV = 3, SC = 9) and were not replaced. These households had similar baseline characteristics between arms. There were a small number of children for whom outcome data were collected who were considered to have been different from the child enrolled at baseline, and these children were excluded from the analysis at 6 mo (n = 29) and at 1 y (n = 36). Overall, the number of households was slightly lower in the DC arm, which was known before randomisation but was not in the original research protocol. No evaluation clusters were lost to follow-up; response rates for households and children, respectively, within clusters were 95.6% and 98.3% at 6 mo and 95.0% and 96.8% at 1 y. The number of missing child data was slightly lower at 1 y compared to 6 mo for the CG only, as efforts were made to reduce loss to follow-up by offering a hygiene kit once the final data had been collected. Compared to the other arms, the CG had the lowest number of missing child data but, as the extent of missing data was small for all arms, we did not anticipate any effects on comparability between arms. All clusters received and utilised the correct intervention assigned during implementation.
Baseline characteristics of clusters and participants between the different intervention arms and the control arm were well balanced for mothers and their children, apart from the proportion of children who had received deworming treatment, which was lower in the CG (Table 1). There were a few potential imbalances at the household level and between villages due to the clustered nature of the study design. These include village size, ethnicity, access to safe water, and distance to nearest health service. In the CG, there was a higher proportion of households of Balochi ethnicity. In this arm, there also appeared to be differences in the socio-economic status and educational status of mothers and fathers (both lower) and a higher number of households participating in the BISP.
The proportions of children who were wasted at the different time points are shown in Fig 2. The trend across time was similar for each arm, increasing during the first month and then decreasing to 6 mo. Prevalence was higher again at 1 y, although it was lower than at the same time in the previous year (baseline) for all arms. The differences in prevalence between arms at each month were quite similar.
Assuming a linear change over months (i.e., a consistent rate of change from one month to the next), adjusting for stratification, clustering, and baseline variables, there were no observed significant differences in children being wasted over the 1-y study period (DC: OR = 0.99; 95% CI 0.96, 1.03; p = 0.69; FFV: OR = 1.02; 95% CI 0.99, 1.06; p = 0.21; SC: OR = 1.02; 95% CI 0.99, 1.05; p = 0.29). We were not expecting the interventions to have an immediate effect, but more an accumulative effect, which explains this lack of significance from month to month. From baseline to 6 mo and to 1 y, the largest (unadjusted) DDE in proportion of children wasted was in the DC arm at 6 mo (−3.3%; 95% CI −8.2%, 1.6%, p = 0.19), with a smaller reduction at 1 y (−1.6%; 95% CI −7.0%, 3.9%; p = 0.58). In both the FFV and SC arms, the DDEs in the proportion of children wasted were higher than in the CG at both time points (FFV—6 mo: 1.6%; 95% CI −3.3%, 6.4%; p = 0.53; 1 y: 1.8%; 95% CI −3.6%, 7.2%; p = 0.51; SC—6 mo: 0.8%; 95% CI −4.0%, 5.6%; p = 0.74; 1 y: 1.0%; 95% CI −4.3%, 6.3%, p = 0.71). None of these differences were statistically significant.
Changes in mean WHZ are shown in Fig 3 and show similar trends to the prevalence data in Fig 2. Again, assuming a linear change, adjusting for stratification, clustering, and baseline variables, there are no observed significant differences over months in children’s ponderal growth (DC: −0.003; 95% CI −0.013, 0.007; p = 0.62; FFV: −0.002; 95% CI −0.012, 0.008; p = 0.74; SC: −0.005; 95% CI −0.015, 0.005; p = 0.28).
Between baseline and 6 mo, the largest unadjusted DDE in mean WHZ was in the FFV arm (0.15; 95% CI −0.02, 0.31; p = 0.08), followed by the DC arm at 6 mo (0.09; 95% CI −0.07, 0.25; p = 0.28). These differences were not statistically significant. The SC arm at both time points and the DC and FFV arms at 1 y showed very little difference from the CG.
The remaining tables show the results from the regression models. The crude models in each case differ from the adjusted models in terms of the width of the CIs, which are narrower in the adjusted models, likely due to inclusion of baseline values of the outcome variables. In nearly all cases, sensitivity analyses resulted in very similar outputs.
The odds of children being wasted in the DC arm were 48% lower compared to the CG at 6 mo (Table 2), adjusted for baseline age, sex, and baseline WHZ. This difference between arms was statistically significant (p = 0.02). This intervention effect was seen at 6 mo only and was not observed at 1 y. There were no significant intervention effects for children being wasted in either the FFV or SC arm compared to the CG at either time point.
Children in the FFV arm showed the largest significant increase in mean WHZ in both the partially and fully adjusted models (+0.16 WHZ) at 6 mo, followed by children in the DC arm (+0.11), adjusted for baseline age, sex, and baseline WHZ. These intervention effects were not present at 1 y. The SC arm was no different from the CG for any of the primary outcomes at either time point.
There were no significant changes in the odds of being severely wasted (WHZ < −3) at 6 mo, adjusted for baseline age, sex, and baseline WHZ (Table 3). For mean MUAC, children were no different from the CG for any of the interventions at 6 mo, with ORs very close to zero.
For other secondary child anthropometric outcomes, all three intervention groups showed a significant decrease in the odds of being stunted and severely stunted and in mean HAZ at both 6 mo and 1 y compared to the CG (Table 3). At 6 mo, the odds of being stunted (HAZ < −2) were 61% (DC) and 64% (SC) lower for the two cash arms, followed by the FFV arm (59% lower odds). For severe stunting, children in the FFV arm had the lowest odds (62% lower), followed by DC (60% lower) and SC (53% lower). At 1 y, the odds of being stunted and severely stunted were similar for the three arms and still statistically significant. Regression coefficients were similar for all three intervention groups at both time points. Children in the FFV arm had the greatest improvement in mean HAZ at both time points (+0.27 and +0.30), followed by similar improvements in the two cash arms: DC (+0.24 and +0.19) and SC (+0.24 and +0.21).
There was no intervention effect on the odds of children being anaemic at 6 mo for any of the intervention arms (Table 4). However, for mean Hb status, children in the FFV arm had a significantly lower Hb level compared to the CG (−2.6 g/l).
There was no intervention effect for children having diarrhoea for any intervention arm (Table 5). The odds of having an acute respiratory infection (ARI) were 43% lower for children in the DC arm. The odds of having fever/malaria were similarly lower than in the CG for children in both the DC and SC arms (37% and 36%, respectively). The FFVs had no discernable effect on morbidity.
Mothers in the FFV arm saw a significant positive intervention effect on BMI at 6 mo (0.29 kg/m2) (Table 6). There was no effect on maternal BMI for the DC or SC arm. There were no effects on mean MUAC across all intervention arms at 6 mo. Mothers were twice as likely to be anaemic in the FFV arm. This negative effect was also seen in the FFV and SC arms for Hb status (−5.0 g/l and −4.2 g/l). There was no intervention effect for anaemia or Hb status for the DC arm at 6 mo.
Whilst households used the different cash amounts in very similar ways: 90% on food, 8% on medical supplies/services, and 2% on non-foods, more households in the SC and FFV arms thought that the 1,500 PKR amount was not enough to meet all their needs:
Those getting the 3,000 PKR amount were more content:
Households receiving either cash amount were happy that they were getting cash rather than the FFVs:
And many households receiving the voucher would have preferred the cash:
However, some households receiving the FFVs preferred them to cash:
However, there was some negative feedback from the FFV arm, especially concerning the vendors where recipient households were able to redeem their vouchers:
There was some concern that the vouchers could have been exchanged for cash or other “non-specified” food and non-food items:
Households receiving the larger amount of cash (DC) saw a significant reduction in the odds of their children being wasted at 6 mo. In addition, the DC intervention had positive and significant effects on stunting (HAZ). The FFVs also had positive effects on stunting, although the odds of being wasted for children in this intervention group was no different from that in the CG. No intervention effects for wasting were seen 6 mo after the last disbursement (at 1 y). Children in households receiving SC were no different from children in the CG for the wasting outcome.
All three interventions resulted in a reduction in odds of being stunted and severely stunted and saw positive effects on linear growth (mean HAZ). These effects remained 6 mo after the last disbursement (at 1 y). We did not see any effect from any of the interventions on severe wasting and child MUAC. The FFVs resulted in a reduction in mean Hb concentration, although this did not translate into an increased risk of being anaemic (potentially because the proportion of children already anaemic was very high; approximately 90%). We saw no effects on risk of having diarrhoea for any arm. However, there was a positive intervention effect of DC on ARIs and fever/malaria, and of SC on fever/malaria.
The intervention effects for mothers mirror those for their children for BMI and Hb status: a positive effect on BMI was found for the FFVs and, in the same arm, a negative effect for Hb status, as was the case with the SC arm. For FFVs, this lower Hb status also translated into a significant increase in the risk of being anaemic. As for children, no intervention effects were found for mother’s MUAC status.
The results for the DC arm support our hypothesis that larger amounts of cash combined with BCC can benefit child growth. The qualitative study suggests that households were happier with the larger amount, and this may in itself have conferred a more positive attitude toward the intervention, and potentially toward the uptake of BCC messages. It is interesting, however, that increasing the amount did not confer positive intervention effects on Hb status or anaemia prevalence. There is evidence that CBIs can improve anaemia status [19], although it is possible that in contexts where very high levels of anaemia already exist, non-food/dietary-based factors, e.g., intestinal worm infestation, may mask or undermine any positive impact of the CBIs, suggesting the need for additional interventions in tandem with an increased cash amount.
A surprising and unintended outcome was the significantly lower Hb levels in children and their mothers in the FFV arm. In addition, mothers in the FFV arm saw a significant increase in the prevalence of anaemia. We had thought that households had been allowed to exchange their FFVs—against protocol—for other items such as foods with low levels of iron (rice, oil, or sugar) or foods with detrimental effects on iron absorption (milk, eggs, or tea). The qualitative data, however, confirmed that the vendors rarely exchanged the vouchers for other, non-fresh-food items. Nor did vendors exchange the vouchers for cash. We had hypothesized that, compared to the CG, the FFVs, with a similar value to the SC transfers, would deliver a greater nutrition impact. This was true in terms of WHZ, whereby children grew more in the FFV arm than in the SC arm. However, it was thought that the FFVs would impact growth and micronutrient status through increasing dietary diversity. An analysis of dietary diversity at the mother and child levels (S1 Table) saw a significant improvement for all three arms, but this improvement was lowest in the FFV arm (highest in the DC arm). Regarding child dietary intake of specific foods, it is not obvious if this had an effect on Hb in the FFV arm as, in all significant cases, children in all the intervention arms had a higher intake of specific foods compared to the CG (S2 Table). Whilst there was a significant increase in consumption of animal protein compared to the CG, the type of meat was not differentiated. Qualitative evidence suggests that the only meat available for the FFVs was chicken, which is itself low in iron. The DC arm had higher intakes of both iron-rich foods and iron absorption inhibitors (e.g., milk and eggs), which may explain why increases in Hb were not seen here. Given that child mean WHZ improved over time in the FFV arm, and yet child dietary diversity was lower compared to the CG, a possible explanation is that there were differences in the amounts of food consumed (i.e., children’s caloric intakes could have been higher in the FFV arm). However, it is not possible to conclude anything about the quantities of food eaten from this study, as these data were not collected.
The intervention effect for the DC arm on child weight-based variables was only apparent at 6 mo. This suggests that where CBIs have an objective to reduce the risk of wasting, this can be effective, but when the causes of food insecurity and high morbidity are not removed, children remain vulnerable to wasting. The limited evidence of impact of CBIs on wasting in the literature is entirely focused on the short term [14], unlike for food-based interventions, where there is some evidence that the risk of being wasted remains 12 mo after recovery [20].
For height-based variables, the positive intervention effect was found at both 6 mo and 1 y in all three intervention groups compared to the CG. This is in itself an important finding, as stunting is a well-accepted marker of overall national development, and its reduction to 20% is a WHA target. Many governments, development partners, and global actors are actively supporting efforts to see acceleration in the rates of stunting reduction. High rates of child stunting also carry a mortality risk, and the more severe the stunting, the greater the risk. Moderately stunted children have a 2.3 times increased risk of death, and severely stunted children are 5.5 times more likely to die [21]. In this study, we saw similar and significant improvements in reduced odds of moderate and severe stunting across the intervention arms at 6 mo and 1 y. The finding that the odds of being stunted were significantly reduced in children at 6 mo is a potentially unexpected outcome from this study, given the short-term nature of the interventions. What is reassuring are the similar results at 1 y, indicating a real effect. Another cash-based longitudinal cluster randomised controlled trial set in Malawi also found a positive impact on linear growth over 1 y and attributed this reduction in prevalence of stunting to the intervention improving food security and dietary diversity [22], which we also saw in our study. The results from our study will be further examined in a forthcoming mediation analysis.
The higher amount of cash (DC) reduced the risk of ARIs, and both the SC and DC interventions reduced the risk of fever/malaria. The morbidity reduction effect was stronger in the DC arm and may in part be explained by the improvement in nutritional status of children in this arm. Expenditures on health and access to health services will be evaluated in further analyses. That there was no reduction of risk for any disease in the FFV arms suggests that FFVs are not as effective as cash at reducing morbidity risk, particularly for fever/malaria.
This is one of only two robust RCTs to our knowledge carried out in a humanitarian aid setting showing significant CBI intervention effects on child nutritional status. A similar positive intervention effect using a CBI has been seen in one other published study [23]. Our study provides evidence to inform policy and programmes and offers good practice-based evidence for all those concerned with reducing the risk of increased child undernutrition in emergencies and in severe lean season contexts through CBIs. That the BISP and other programmes in Pakistan have shown great interest in the study and its results provides impetus for future CBI programme design linked with nutrition policy objectives within Pakistan.
The strengths of this study include its randomised design, adherence to the implementation process, good retention rates, extensive process evaluation, and cost-effectiveness analysis (to be published elsewhere). Even though data were collected under difficult conditions, the rate of missing data was low. The theory of change for the relationship between CBIs and nutritional status is complex, making scale-up difficult unless some of the questions about how and why the intervention worked, or didn’t work, are understood. The process evaluation undertaken during this study will be used together with a mediation analysis to understand the “how” and “why” of the intervention effects in a future analysis. The recent High Level Panel on Humanitarian Cash Transfers agreed that cash can be effective in humanitarian aid settings but may, at some times and in some places, be inappropriate [24]. To ensure that CBIs are designed in the best possible way, it is important to ensure that there are functioning markets and to understand the causes of undernutrition within a setting. In Pakistan, a previous nutrition causal analysis identified that low income was a significant underlying cause of undernutrition [25], and the current study has shown that cash transfers can have a positive impact on this underlying cause.
There are a number of limitations to this study. First, masking of the interventions to both participants and data collectors was not possible in this setting and for this type of study. Precautions were taken at the start of the study to try to mask the different interventions to participants, e.g., through incorporating “buffer” zones and training data collectors to keep the information to themselves, but it soon became clear that participants were aware of the other interventions. This was especially so for the CG as the dropout rate for this arm increased more in subsequent months after baseline. However, to encourage continued participation, this group was given a hygiene kit after the last round of data collection. For all groups, the data collectors were trained to sensitise the participants to the study objectives and to ensure the same key messages were highlighted during data collection. The data collectors could not be masked about which arms were getting what intervention because part of the process evaluation was to ask questions about the use of the intervention. To ensure similarity between intervention arms, data collectors were rotated so they covered different groups. The disbursement of the cash and vouchers was done by different organisations, and the cash participants had further to travel to their distribution point, which may well have added to the opportunity costs to households and reduced the actual transfer value. Added to this, the FFV arm had more direct contact with Action Against Hunger field staff during voucher disbursement, which could have affected the results through greater exposure to key messages. Efforts were made throughout the study to engage with the Action Against Hunger field staff and to sensitise them to the study objectives. It is also possible that the vouchers themselves were too restricted. They were designed to purchase fresh fruit, vegetables, and fresh meat and were, therefore, dependent on what the vendors stocked, such as chicken being the only available meat. There were also many anecdotal reports regarding vendors overcharging for food items redeemed against the vouchers as a way to cover their own administration fees in recovering the voucher costs. In this respect, the actual transfer value given may have been lower than the face value.
With these data, it is not possible to calculate a “threshold” for the minimum amount of cash that would have had a significant effect. We know, however, that this threshold falls somewhere between the amounts in the SC and the DC interventions. We can also not say anything about the quantities of food bought or the quality of medical services accessed as these data were not collected. Finally, the Sindh Province context presented a number of difficulties affecting data collection. The baseline survey took longer than expected since recruitment of female data collectors was difficult and was a reason why the baseline data collection was extended. Added to this, temperatures reached 52°C, which not only affected the data collection team’s working ability but also had an effect on the HemoCue devices used to measure Hb. There are more missing data, therefore, for Hb in the two cash arms than in the FFV and CG arms for both children (Table 4) and their mothers (Table 6).
The results from this study are the first to our knowledge to be seen from a CBI programme in a humanitarian aid setting in Asia. Whilst these results are very compelling, the findings raise questions about the optimal approach when using FFVs in contexts of high anaemia prevalence and the need for future programme design to ensure such interventions enable access to the correct foods, in the correct amounts, and do not have restrictions attached to them. At the same time, understanding and mitigating the non-food causes of anaemia are warranted.
Unconditional cash transfers of at least 3,000 PKR, equivalent to approximately US$28 (twice as much as the SC amount based on the BISP), were more effective in improving weight-based growth immediately following the intervention in a population of poor and very poor households with young children. This effect was seen against a backdrop of very high wasting at baseline (>20%), an indicator of the deleterious effects of seasonal food shortages and high morbidity in this region. The type of intervention did not really matter for height-based variables as all three intervention groups had a significant improvement at both the 6 mo and 1 y time points compared to the CG, thus indicating movement toward greater nutrition resilience, whereby having a better nutritional status increases the capacity of a person or population to withstand shocks or stressors that might adversely affect the causes of undernutrition.
The FFVs had an unintended negative impact on Hb status, and this may have been due to the restrictive nature of the voucher—in this sense, unconditional cash transfers were better than vouchers, though mean WHZ did improve in the FFV arm.
|
10.1371/journal.ppat.1003408 | The Systemic Immune State of Super-shedder Mice Is Characterized by a Unique Neutrophil-dependent Blunting of TH1 Responses | Host-to-host transmission of a pathogen ensures its successful propagation and maintenance within a host population. A striking feature of disease transmission is the heterogeneity in host infectiousness. It has been proposed that within a host population, 20% of the infected hosts, termed super-shedders, are responsible for 80% of disease transmission. However, very little is known about the immune state of these super-shedders. In this study, we used the model organism Salmonella enterica serovar Typhimurium, an important cause of disease in humans and animal hosts, to study the immune state of super-shedders. Compared to moderate shedders, super-shedder mice had an active inflammatory response in both the gastrointestinal tract and the spleen but a dampened TH1 response specific to the secondary lymphoid organs. Spleens from super-shedder mice had higher numbers of neutrophils, and a dampened T cell response, characterized by higher levels of regulatory T cells (Tregs), fewer T-bet+ (TH1) T cells as well as blunted cytokine responsiveness. Administration of the cytokine granulocyte colony stimulating factor (G-CSF) and subsequent neutrophilia was sufficient to induce the super-shedder immune phenotype in moderate-shedder mice. Similar to super-shedders, these G-CSF-treated moderate-shedders had a dampened TH1 response with fewer T-bet+ T cells and a loss of cytokine responsiveness. Additionally, G-CSF treatment inhibited IL-2-mediated TH1 expansion. Finally, depletion of neutrophils led to an increase in the number of T-bet+ TH1 cells and restored their ability to respond to IL-2. Taken together, we demonstrate a novel role for neutrophils in blunting IL-2-mediated proliferation of the TH1 immune response in the spleens of mice that are colonized by high levels of S. Typhimurium in the gastrointestinal tract.
| Bacteria belonging to the genus Salmonella are capable of causing long-term chronic systemic infections in specific hosts where they are shed in the feces. These persistently infected individuals include typhoid carriers and they serve as a reservoir for disease transmission. Despite the importance of Salmonella as a human pathogen, relatively little is known about the host immune response to persistent bacterial infections in the context of transmission. We had shown previously in a mouse model of Salmonella infection that mice shedding high levels of Salmonella (>108 bacteria per gram of feces), known as super-shedders, transmit disease to naïve mice. We show here that these super-shedder mice have a unique immune state compared to mice that have lower levels of Salmonella in their gut. The super-shedder immune state is characterized by an active inflammatory immune response with elevated serum IL-6 and high levels of neutrophils in both the gastrointestinal tract and the systemic sites but a dampened adaptive CD4 T helper type1 (TH1) cell response specific to the spleen. Importantly, we show that the blunted adaptive response, as characterized by reduced TH1 cell frequencies and ability to respond to IL-2 and IL-6, is intimately linked to the levels of neutrophils present in the spleen. We go on to show the functional consequences of dampened cytokine responsiveness, as TH1 cells from moderate-shedders are unable to undergo IL-2-mediated expansion when neutrophilia is induced. Additionally, we show that neutrophil control of IL-2 mediated expansion of TH1 cells is independent of infection. In summary, we describe an immune phenotype associated with transmission of a pathogen and a single immune cell type, neutrophils, which control specific aspects of the super-shedder immune state.
| Host-adapted pathogens depend on their host for transmission and dissemination within a population. Recent epidemiological studies have uncovered heterogeneities in infection wherein a minority of the infected individuals (20%) are responsible for the majority of the infections (80%), described as the 80/20 rule [1]. In the case of pathogens transmitted via the fecal-oral route, these individuals are the ones that shed the highest numbers of bacteria. Recent studies on the transmission of Escherichia coli O157 within cattle herds demonstrated that over 95% of the infections were caused by between 8–10% of the most infectious individuals, or super-shedders [2]–[4]. Identification of these individuals is required for control of the infection [1], [5], [6]. However, little is known about what distinguishes them from other infected hosts.
Salmonella enterica serovar Typhi, the causative agent of typhoid fever in humans, is a human-adapted pathogen and establishes a persistent long-term infection in about 1–6% of the infected hosts [7], [8]. These individuals are known as typhoid carriers and periodically excrete large amounts of the bacilli in their feces, thereby offering both a reservoir for the pathogen and the opportunity of transmission to new hosts. However, they do not display any of the clinical signs characteristically associated with typhoid fever [7], [9]. While individuals with acute infections can transmit the pathogen for brief periods of time, for the purposes of this study, we will focus on transmission from persistently infected hosts who play a far larger role in transmission of host-adapted pathogens [1], [2], [4], [10].
Host immune responses to persistent microbial infections must balance between control of the pathogen and minimizing inflammatory damage to the host [11]. To this end, chronic viral infections often result in a contraction of the adaptive immune response, an example of which would be T cell exhaustion [12], [13]. An ineffective CD4 T cell response, characterized by anergy or apoptosis, has also been observed in persistent bacterial infections such as with Helicobacter pylori, Staphylococcus aureus and Salmonella enterica serovar Typhimurium [14]–[16]. Intriguingly, regulatory T cells have also been shown to lose their suppressive ability during the later stages of persistent S. Typhimurium infection [17].
Characterizing the host immune response in individuals that transmit disease is important for two reasons. First, understanding the mechanistic differences in this subset of hosts might explain the heterogeneity of infectiousness observed in relatively homogenous populations, such as with inbred herds of cattle. Second, such studies could lead to the development of biomarkers that are unique to the identification of individuals with the highest risk of transmitting the pathogen within a population.
Modeling the immune state using laboratory animals allows us to dissect the mechanisms behind host-pathogen interactions that lead to transmission. Our lab has established a mouse model of persistent S. Typhimurium infection wherein 30% of the infected mice, termed super-shedders, shed >108 Salmonella and rapidly transmit disease to naïve cage mates [18]. This variation in infectiousness is observed in inbred strains amongst cage mates and siblings, implying predisposition to super-shedder status might not be heritable. Super-shedder mice also develop colitis, displaying moderate to severe inflammation in the colon and ceca. Surprisingly, super-shedder mice do not display outward signs of illness such as ruffled fur, fever or malaise [18] suggesting that they can tolerate, and perhaps control the inflammation. How, then, does the interplay between pathogen and the murine immune system evolve to allow such high levels of gastrointestinal Salmonella in some mice, but not others?
The host immune response to Salmonella infections has been characterized primarily in mice that have increased susceptibility to intracellular pathogens due to the presence of a mutated Nramp1 gene [19]–[21]. In susceptible mice, wild type S. Typhimurium infection results in death within one week of infection. In contrast, these susceptible mice strains survive infection with attenuated S. Typhimurium strains (e.g, AroA− AroD− mutant). In this model the adaptive immune response to persistent Salmonella infection was found to be TH1 biased and was dependent upon expression of transcription factor T-bet [22].
In this study, we used a mouse strain (129x1/SvJ) that carries functional Nramp1 and a non-attenuated Salmonella strain to more closely mimic natural infection [23]. This persistent infection model was used to investigate host immune adaptations to gastrointestinal inflammation that are associated with survival. We hypothesized that the ability of the host to co-exist with large numbers of bacteria in the gastrointestinal tract requires either a dampened systemic inflammatory response or the ability to tolerate inflammation. To test this, we compared the host immune responses in super-shedder and moderate-shedder mice and characterized an immune state specific to super-shedders. Broadly, we found super-shedders have an activate innate inflammatory response in both the gastrointestinal tract and the systemic organs but a dampened adaptive T cell response specific to the systemic sites. Finally, we identify an unexpected host immune mechanism mediated by neutrophils that controls TH1 cell expansion in the super-shedder immune state.
To examine the host immune response to persistent Salmonella infection, we first enumerated the bacterial burden in gastrointestinal and systemic tissues of mice infected with S. Typhimurium for 30 days. Bacterial loads in the spleen and the mesenteric lymph nodes (MLN) were tightly clustered across mice, while those in the gastrointestinal sites showed the large variation characteristic of super-shedders and moderate-shedders ([18], Figure 1A). It is known that super-shedder mice develop colitis that results in an influx of granulocytes to the colon [18]. We identified moderate and super-shedder mice as outlined in material and methods and asked if the neutrophilia extended to secondary lymphoid organs. We observed a notable increase in the frequency of Gr1+ (Ly6G+Ly6C+) granulocytes in the spleen, MLN, and blood (Figure 1B) of super-shedder mice compared to moderate-shedders. Super-shedder spleens were composed of 12±2.8% neutrophils; in contrast, on average, the moderate-shedder spleen contained only half that amount (5. 3±1.1% neutrophils). Histological analysis demonstrated that neutrophils constituted over 80% of splenic myeloid cells (Fig. 1C). Neutrophils were also the most numerous cell type in super-shedder blood, with 63.6±21.6% of the non-red blood cells identified as circulating neutrophils, while moderate-shedder blood contained only 25.3±3% neutrophils. When compared to uninfected mice, the levels of neutrophils in the systemic organs of moderate-shedders were significantly higher (Figure 1B). However, no significant difference in colonic neutrophil levels between uninfected and moderate-shedder mice was observed indicating that neutrophilia in the systemic sites was a stronger indicator of shedding status. Taken together, our data suggest that neutrophil levels in the systemic sites are positively correlated with gastrointestinal bacterial load.
Having shown that splenic neutrophilia varied with gastrointestinal and not splenic bacterial burden, we examined whether there were associated variations in adaptive immune responses. To exploit the variation in fecal shedding, we asked what aspects of the host immune response varied with fecal bacterial load.
Given the importance of CD4 T cells in the host defense against Salmonella infection, [24] we first focused on the frequencies of two CD4 T cell subsets: TH1 and Tregs. The host immune response against Salmonella infection requires the induction of a CD4 type 1 Helper T cell or TH1 immune response, involving CD4 T cells expressing the transcription factor T-bet. TH1 activity can in turn be controlled by regulatory T cells (Tregs) expressing the transcription factor FoxP3. In one representative experiment, we measured these subsets in the spleens and colons of 12 mice of which five were super-shedders. This was confirmed by the levels of colonic inflammation though 2 of the 5 mice were shedding between 107 and 108 cfu/gm. The infected mice clustered into two distinct groups with 5 of the 7 moderate-shedders clustered together with fewer Tregs and more TH1 cells while 4 out of the 5 super-shedders were in a cluster that contained fewer TH1 and more Tregs (pink box vs. blue box, Figure 2A). The percentage of T-bet+ CD4 T cells (TH1) cells in the spleen significantly negatively correlated with fecal bacterial burden (Spearman's correlation = −0.58) but was not significantly correlated to the splenic bacterial burden (data not shown). Additionally, this dichotomy between active and suppressive T cells was not observed in the colon (Figure S1A). Importantly, in uninfected mice there was a positive correlation between the frequencies of CD4 T cells expressing T-bet and those expressing FoxP3, indicating that the skewing in the populations of TH1 and Treg cells is dependent on infection and is not a result of an underlying natural variation in the uninfected mouse population (Figure S1B). We found very few Rorγt-expressing CD4 T cells in persistently infected mice (data not shown). Infection-induced variation was further evidenced by the relationship between TH1 and neutrophil percentages in the spleens of infected mice. All 5 super-shedder mice clustered together (blue box) with high levels of neutrophils and correspondingly lower levels of TH1 cells. All moderate-shedders clustered on the opposite end with fewer than 5% splenic neutrophils but higher frequencies of TH1 cells (Figure 2B). Correspondingly, splenic neutrophilia was significantly positively correlated with fecal bacterial load (spearman's correlation R = 0.9). Importantly, splenic bacterial load did not correlate with fecal bacterial load (Figure S2A).
CD4 T cell exhaustion is a hallmark of persistent viral infections [12]–[14] so we sought to determine if these TH1 cells maintained antigen-responsiveness. Splenocytes from infected mice were incubated with S. Typhimurium-infected macrophages and the level of intracellular IFNγ levels was measured. Salmonella-specific IFNγ+ Tbet+ CD4 T cells were first detected at 8 days post-infection and expanded continuously through 30 days post-infection (Figure S3A). While culling of flagellin-specific CD4 T cells has been previously reported during Salmonella infection [16], we saw a steady expansion of total memory effector CD4 T cells over a time course of infection (Figure S3B). Antigen specific IFNγ production was observed in CD4 T cells across all mice, regardless of shedder status (data not shown). These data show that while super-shedder spleens have lower numbers of TH1 T cells relative to moderate shedders, these cells still make IFNγ in response to Salmonella antigen. Finally, we asked if the dampened TH1 response also resulted in reduced antibody production. To determine this, we measured anti-Salmonella antibodies in the serum of persistently infected mice and found no correlation with shedding status (Figure S2B) indicating that the blunted TH1 response in super-shedders did not affect antibody production.
Having determined that the TH1 cells were antigen-responsive, we further investigated the ability of CD4 T cells to respond to IL-2, a cytokine that induces T cell proliferation. The high affinity IL-2 receptor is expressed on Tregs and memory effector CD4 T cells; upon binding IL-2 one of the first steps initiated in the intracellular signaling cascade is the phosphorylation of STAT5 protein. Thus, the mean fluorescence intensity (MFI) level of phosphorylated STAT5 was measured in the subset of CD4 T cells that responded to IL-2 (gated on pSTAT5+CD4+ T cells). This metric represented the degree of cytokine responsiveness and correlated negatively with the gastrointestinal Salmonella burden (Figure 2C). pSTAT5 response to ex vivo IL-2 stimulation has been previously established as an indicator of T cell expansion in vivo [25]. To determine if this reduction in IL-2 responsiveness in super-shedders coincided with reduced T cell proliferation we measured the expression of Ki-67, a marker of actively proliferating cells. Consistent with decreased IL-2 responsiveness, super-shedder mice had significantly fewer Ki-67+ CD4 T cells in the spleen (Figure 2D). That STAT5 phosphorylation correlated inversely with the levels of bacterial shedding suggests that a feature of the super-shedder immune response involves blunting of IL-2 responsiveness. We therefore evaluated the extent to which other alterations in persistent immune responses were linked to fecal shedding status.
Given the dampened IL-2 responsiveness of CD4 T cells, we asked if the ability to respond to other cytokines was also impaired. Previous work in a mouse model of septicemia showed that naïve splenic CD4 T cells have a dampened response to IL-6 compared to uninfected mice, indicated by a reduced ability to phosphorylate STAT1 in response to ex vivo IL-6 stimulation [26]. In the CD4 T cell compartment, the IL-6 pSTAT1 response is primarily restricted to naïve cells, as memory effector cells express very little IL-6 receptor (data not shown). Naïve CD4 T cell pSTAT1 responsiveness to IL-6 negatively correlated to fecal shedding levels (Figure 2E), reminiscent of the IL-2 response observed earlier. It is important to note that the IL-6 response is dampened in super-shedder mice only with respect to moderate or low shedders. Compared to uninfected mice, Salmonella infection induces increased IL-6 responsiveness in naïve CD4 T cells (Figure S4) but this responsiveness varied with the levels of fecal shedding. Furthermore, circulating IL-6 levels were significantly higher in super-shedder mice compared to moderate-shedders and uninfected mice (Figure 2F).
Taken together, these results reveal that across equivalently infected mice, those that developed as super-shedders are characterized by an activated innate inflammatory response with high levels of circulating IL-6 and neutrophils that is associated with a spleen-specific dampened CD4 T cell response. This dampened T cell response is characterized by a partial loss of cytokine responsiveness to IL-2 and IL-6 compared to moderate-shedder mice. Finally, in super-shedder spleens, the balance of CD4 T cell subsets supports a dampened CD4 T cell response, with fewer TH1 cells and more Tregs (Table 1).
Does the gastrointestinal Salmonella burden dictate the systemic immune profile, or does the immune response control the bacterial load? Since the numbers of bacteria in the gastrointestinal tract are correlated with specific changes in the neutrophil and CD4 T cell immune response in the spleen, we investigated whether altering the levels of S. Typhimurium in the gut was causal to changes in the splenic immune response. Previously it was shown that alterations of the microbiota in moderate-shedders via a single dose of streptomycin resulted in super-shedder levels of Salmonella in the gastrointestinal tract [18], [27]. Therefore, moderate-shedder mice were treated with an oral dose of streptomycin and their fecal shedding and splenic immune state were monitored (the Salmonella strain used - SL1344 - is resistant to streptomycin). Three days after streptomycin treatment, moderate-shedders shed >108 Salmonella per gram of feces, i.e. super-shedder levels. Their splenic bacterial burden remained unchanged, as compared to untreated moderate-shedder mice (Figure 3A). Moreover, within 3 days streptomycin-treated moderate-shedders developed increased levels of neutrophils in the colon and spleen, comparable to those seen in super-shedders (Figure 3B). This was accompanied by a decrease in splenic TH1 cells in 3 out of the 5 streptomycin-treated mice (Figure 3C), and by one week post-treatment, all of the streptomycin-treated moderate-shedders had fewer splenic TH1 cells (data not shown).
Notably, streptomycin-treated moderate-shedders had increased levels of circulating IL-6 and a concomitant decrease in the ability of CD4 T cells to respond to IL-6 as measured by pSTAT1 levels (Figure 3D, Figure 3E). However, the percentage of splenic Tregs did not change (Figure S7A). Therefore, many aspects of the splenic super-shedder immune phenotype are induced by raising gastro-intestinal levels of Salmonella, although the frequency of regulatory T cells in the spleen is independently regulated.
We next sought to identify which components of the host immune response control the dampened TH1 response observed in the spleens of super-shedder mice as compared to moderate shedders. Increased neutrophil numbers were seen in the spleen as early as four days post-infection, a time point at which Salmonella was undetectable outside the gastrointestinal tract (Figure S5). Since increased levels of neutrophils in the colons and spleens of super-shedder mice correlated with the dampened adaptive TH1 immune response, we proposed that neutrophils play a role either directly or indirectly in mediating the immune blunting.
We depleted neutrophils using the monoclonal antibody RB6, which targets cells expressing both Ly6C and Ly6G. Neutrophil depletion increased the levels of splenic TH1 cells from 10.2±5.8% in control mice to 24.1±8.8% in RB6-treated mice (Figure 4A). Similar results were obtained with a Ly6G-specific depletion antibody, IA8 (Figure S6A,B). An increase was observed in the frequency of pSTAT5+ CD4 T cells that responded to ex vivo IL-2 stimulation regardless of infection, indicating that neutrophils suppress CD4 T cell responsiveness to IL-2 (Figure 4B). Intriguingly, uninfected mice depleted of neutrophils also showed a similar increase in pSTAT5+ CD4 T cells, but without TH1 expansion. This indicates that the IL-2/pSTAT5 response may be an intermediate step to TH1 expansion and that TH1 biasing occurs only in the context of infection. The percentage of splenic Tregs in infected and uninfected mice did not statistically change upon neutrophil depletion, implying that the increase in IL-2 responsive CD4 cells was not due to Treg reduction (Figure 4C). In addition, RB6-treated mice had significantly higher bacterial burdens in the spleen compared to control mice (Figure S6C). This demonstrates that neutrophils are necessary to control splenic infection, and that the increased TH1 response is unable to compensate for neutrophil depletion. Surprisingly, there was no difference observed in fecal bacterial burden (Figure 4D) suggesting that there might be an organ specific function for neutrophils in persistent Salmonella infection. Taken together, the results indicate that high levels of neutrophils in the spleen are necessary for both dampened IL-2 responsiveness and a reduction in the levels of of TH1 cells.
Based on the finding that neutrophil depletion induced TH1 expansion, we investigated whether the strong systemic neutrophil induction seen in super-shedder mice was sufficient for limiting the TH1 response. Moderate-shedder and uninfected mice were injected with G-CSF for 3 days to induce granulopoiesis. G-CSF treatment increased splenic neutrophil levels to those observed in the spleen of super-shedder mice (Figure 5A). Importantly, G-CSF treatment of moderate-shedder mice led to a concomitant decrease in the frequency of splenic TH1 cells (Figure 5B). However, G-CSF treatment did not influence the frequency of Tregs in the spleens of moderate-shedder or uninfected mice (Figure S7B). Histological analysis of the spleen and bone marrow of G-CSF-injected mice revealed that the granulocytes induced were primarily neutrophils. Additionally, there was a significant increase in the levels of immature neutrophils in the bone marrow of G-CSF-treated moderate-shedders compared to untreated mice (Figure S8A–C).
Since G-CSF treatment caused a reduction in splenic TH1 cell frequencies, we assayed whether cytokine responsiveness of CD4 T cells was also blunted. Compared to untreated moderate-shedders, naïve CD4 T cells from G-CSF-treated moderate-shedders displayed a dampened pSTAT1 response to IL-6 stimulation similar to that observed in super-shedders (Figure 5C). This was finding was in alignment with the trend towards increased levels of serum IL-6 in G-CSF-treated moderate-shedders (Figure S9). G-CSF treatment also significantly dampened IL-2-mediated induction of pSTAT5 in CD4 T cells that responded to IL-2 (Figure 5D). Furthermore, uninfected mice treated with G-CSF demonstrated dampened responses to IL-2 and IL-6, indicating that neutrophil-mediated control of IL-2 induced pSTAT5 and IL-6 induced pSTAT1 responses are independent of infection (Figure 5C, 5D). These data suggested that neutrophils might suppress TH1 expansion via the IL-2/pSTAT5 pathway. Treatment with G-CSF did not increase fecal bacterial load (Figure 5E). Importantly, in ex vivo experiments, G-CSF induced STAT5 activation in granulocytes but not CD4 T cells (data not shown), indicating that G-CSF does not act directly on T cells.
Having observed that G-CSF mediated neutrophilia dampens IL-2 responsiveness across the CD4 T cell population; we next investigated whether TH1 and Treg CD4 T cell subsets differed in their IL-2 responsiveness with functional consequences. Both Tregs and TH1 cells activated by infection induce phosphorylation of STAT5 in response to IL-2 (Figure S7C). Previous studies have shown that IL-2 antibody complexed with IL-2 cytokine (hereafter referred to as IL-2 antibody complex) can induce expansion of Tregs in uninfected mice [28], [29]. When S. Typhimurium-infected mice were treated with IL-2 antibody complex, we observed expansion of both Tregs and TH1 cells (Figure 6B). This was accompanied by an increase in the number of pSTAT5+ total CD4 T cells both before (basal) and after ex vivo IL-2 stimulation (Figure 6A). Furthermore, IL-2 mediated TH1 expansion was significantly greater in moderate-shedders than super-shedders (p<0.05). These results indicate that high levels of gastrointestinal Salmonella burden and neutrophilia may be associated with an impairment in the ability of splenic TH1 cells to undergo IL-2 mediated proliferation.
Our findings that both gastrointestinal Salmonella and G-CSF-mediated neutrophilia are associated with dampened IL-2 responsiveness in TH1 cells, suggest that neutrophil levels influence the ability of TH1 cells to expand. To investigate this, moderate-shedders were treated with G-CSF for 3 days, then subsequently administered IL-2 antibody complex for another 2 days, to determine the effect of neutrophilia on T cells expansion. After IL-2 antibody complex treatment, moderate-shedders pretreated with G-SCF had lower levels of both basal and IL-2 responsive pSTAT5+ CD4 T cells compared with mice that were not administered G-CSF (Figure 6C). This loss of IL-2 responsiveness correlated with significantly fewer Ki-67+ CD4 T cells (Fig. 6D), indicating that G-CSF treatment inhibited the ability of CD4 T cells to proliferate in response to IL-2. However, this inhibition was specific to TH1 cells, as G-CSF did not affect Treg expansion in response to IL-2 antibody complex (Figure 6E). This result is consistent with our previous finding that only TH1 cells and not Tregs expanded upon neutrophil depletion (Figure 4A, 4C). Thus, CD4 T cells in G-CSF treated moderate-shedders recapitulate the super-shedder phenotype. Taken together, we show that treatment of moderate-shedders with G-CSF is sufficient to recapitulate specific aspects of the super-shedder immune response (Figure 7).
We have described here a unique immune phenotype in the spleen that is linked to the ability of an enteric bacterial pathogen to replicate to high numbers in the gastrointestinal tract and thus transmit to a new host. One interesting aspect of this phenotype is that the immune state in the spleen is associated with the levels of bacteria in a distal site (the gastrointestinal tract) rather than local bacterial burden. Previous work has shown that the gastrointestinal commensal microbiota can activate antigen-presenting cells, which go on to drive adaptive immunity in distal sites such as the lung [30]. However, to the best of our knowledge, this is the first report of a splenic immune phenotype specifically associated with gastrointestinal pathogen load and inflammation.
The super-shedder immune phenotype is composed of a highly inflammatory response in both gastroinestinal and systemic sites (evidenced by neutrophilia and serum IL-6) but a dampened adaptive T cell immune response specific to the systemic sites. The high levels of circulating IL-6 and moderate to severe colonic inflammation seem at odds with the absence of weight loss or malaise observed in these mice. The blunted splenic CD4 T cell response, dampened cytokine responsiveness and increased levels of regulatory T cells might be instrumental in the tolerance of the inflammatory environment that is the super-shedder gut. In persistent bacterial infections, this might provide an opportunity for the host to suppress the long-term inflammatory effects of the adaptive T cell response while still controlling pathogen load via neutrophil recruitment.
The molecular mechanism behind neutrophil recruitment in persistent Salmonella infection is yet to be identified. Gastrointestinal Salmonella is sufficient to induce granulopoiesis and systemic neutrophilia. Neutrophils were observed in the spleen at four days post infection, a time point when Salmonella was not detected in systemic sites but was present in the gut. Previous studies have demonstrated a role for IL-17 mediated neutrophil induction [31]. IL-17 is secreted by multiple cell types including Rorγt expressing CD4 T cells or TH17 cells which have also been shown to play a role in acute Salmonella infection [32], [33]. However, we found very few Rorγt+ CD4 T cells in the spleen of persistently infected mice (data not shown). Additionally we did not detect IL-17 in serum or spleen supernatant of these mice indicating that the TH17 or IL-17 response does not play a role in persistent Salmonella infection (data not shown).
Two interventions shifted moderate shedders towards the super-shedder immune phenotype – streptomycin induction of gastrointestinal bacterial expansion and G-CSF mediated neutrophilia. In both instances, the cytokine signaling profiles and TH1 levels recapitulated those of natural super-shedders. However, a concordant increase in Tregs was not observed. This suggests that the increased Treg levels observed in super-shedders are regulated by a mechanism independent of bacterial load or neutrophil levels. Indeed, the co-expansion of Tregs and TH1 cells is decoupled in super-shedder mice where IL-2 mediated TH1 but not Treg expansion is dampened. It was surprising to find that short-term treatment with G-CSF was sufficient to recapitulate this phenotype. The mechanism behind this and other aspects of the super-shedder immune state remain unknown, as do the bacterial and host effectors required to establish this immune phenotype.
The mechanism by which induction of granulopoiesis inhibits the TH1 response, both in T-bet expression and IL-2 mediated expansion is unclear. It will be important to determine whether this mechanism is mediated by cell-cell contact or through cytokine secretion. Anti-inflammatory neutrophil populations which secrete IL-10 in response to Gram-negative bacteria have been previously described [34]. However, we were unable to detect IL-10 in supernatants of cultured neutrophils isolated from moderate and super-shedder mice. Additionally, co-administration of IL-10 receptor-blocking antibody along with G-CSF did not prevent the dampening of the TH1 response of infected mice treated with G-CSF (data not shown). One of the notable findings of this study was that Tregs are not suppressed in a similar manner to TH1 cells. Previous work has shown that Tregs lose potency in the later stages of persistent Salmonella infection [17]. Our data indicated that Tregs from mice infected for 30 days were capable of IL-2-mediated expansion. However, the suppressive potency of these cells was not analyzed and it is possible that the expanded population is not capable of suppressing T cells.
While induction of granulopoiesis was sufficient to induce aspects of the super-shedder systemic immune network in moderate-shedder mice, no differences were seen in fecal shedding. We suspect that this was because the persistent Salmonella infection had already been established. The only factors shown to induce super-shedder status so far have been antibiotic-mediated ablation of the gut microbiota or neutralization of IFNγ cytokine [18], [23], [27].The immune processes pivotal in the establishment of the super-shedder state may take place at the onset of infection. It should also be noted that G-CSF induction only mimicked the super-shedder phenotype in the spleen. The colonic and cecal inflammation typically seen in super-shedders did not develop in G-CSF-treated moderate-shedders (data not shown). It is possible that either a longer G-CSF treatment or Salmonella-specific induction of host chemokine responses are required for sustained gastrointestinal neutrophilia associated with gastrointestinal pathology.
As noted, the systemic super-shedder immune phenotype is primarily one of an active inflammatory response and a dampened TH1 response. However, these TH1 cells are not anergic, as they retain the ability to recognize Salmonella antigen and secrete IFNγ in response. The serum levels of IFNγ were elevated by four days post-infection and remained consistently high during the first month of infection (Figure S3A). Since the cytokine levels were elevated before the T cell response was initiated, these data suggest that antigen-specific CD4 T cells may not be the primary source of serum IFNγ during infection. It is likely that the source of the early IFNγ is activated macrophages or monocytes. This is further supported by our observation that serum IFNγ did not correlate with fecal shedding and was not a part of the super-shedder immune phenotype. Salmonella-induced blunting of flagellin-specific T cells has been previously reported [16], however, while we did not investigate the clonality of the CD4 T cell population, we saw a steady expansion of antigen-specific, IFNγ producing, memory effector CD4 T cells (Figure S3).
This study describes a distinct splenic immune signature in the Salmonella carrier state responsible for transmission. These results might have important implications for identification of S. Typhi carriers. For instance, current serological methods do not distinguish between carriers and people who have cleared the infection [35], [36]. While Typhi infection induces a neutrophillic inflammatory response, clinical reports on typhoid patients have only described a mild transient granulocytosis [37]–[39] and neutrophils are typically not found in stool samples of S. Typhi patients [40]. It remains unknown if any variation in neutrophil frequency was observed in typhoid patients and if this correlates with carrier status. For example, serum IL-6 levels are elevated in patients with typhoid and have been correlated with prolonged fever [41]. However, in these studies and others, immune correlates with fecal shedding are difficult to obtain. A study on 13 long-term typhoid carriers actively shedding Typhi demonstrated a wide variance in antibody titers in accordance with our data in mice infected with S. Typhimurium ([9], Figure S2B). In other animal models, studies in S. Typhimurium-infected swine uncovered positive correlations between early fecal shedding of salmonella (within the first week of infection) with increased neutrophil numbers and high serum IFNγ [42]. Identifying immune correlates that are associated with active shedding of Salmonella by the host would help determine biomarkers for screening of Typhoid carriers.
Interestingly, there is evidence for an immune phenotype associated with human S. Typhi carriers. Transcriptional profiling of a cohort of acute, convalescing and recovered typhoid patients uncovered specific neutrophil and lymphocyte gene expression sets associated with each of those stages. Specifically, while 50% of the convalescing patients had a gene expression signature indistinguishable from healthy controls, 25% of the patients showed a distinct gene expression pattern in multiple cell types that were more similar to those of newly admitted patients despite being collected 9 months after treatment. This dataset supports the possibility of a long-term alteration in immune response in a subset of S. Typhi patients [43]. The results presented here are a step toward the definition and practical identification of immune states associated with high levels of Salmonella fecal shedding and transmission. That this is mediated by a novel neutrophil-dependent mechanism of IL-2- mediated TH1 blunting, suggests novel disease management approaches both for individuals and human communities where Salmonella is endemic.
All animal experiments were performed in accordance with Stanford University's Institutional Animal Care and Use Committee and NIH Guidelines for Euthanasia of Rodents Using Carbon Dioxide. All animal experiments were approved by Stanford University's Administrative Panel on Laboratory Animal Care (A-PLAC). Stanford University Animal Welfare Assurance Number: A3213-01. Protocol ID 12826. All animals were housed in a centralized research animal facility, fully staffed with trained personnel and accredited by the Association of Assessment and Accreditation of Laboratory Animal Care International (AAALAC). Mice were monitored daily; mice displaying signs of pain, distress (hunched posture, lethargy, ruffled fur) and weight loss were euthanized humanely.
Salmonella enterica serovar Typhimurium wild type strain SL1344 was used for all infections. This strain is resistant to streptomycin. The bacteria were grown, shaking, at 37°C overnight in Luria-Bertani (LB) broth containing 200 µg/mL streptomycin. Bacteria were spun down and washed with phosphate buffered Saline (PBS) before being resuspended into the desired concentration for infection. For macrophage infection, bacteria were diluted to the desired concentration and pipetted onto the cells.
Female 129x1/SvJ mice were obtained from The Jackson Laboratory and infected when they were 7–9 weeks old. Food and bedding were changed once a week by the Stanford Animal Facility and access to food and water was unlimited. For infections, food but not water was removed 12–16 hours ahead. Mice were infected orally, drinking 108 CFU in 20 µL PBS from a pipette tip. For the streptomycin treatment, the antibiotic was delivered orogastrically with 5 mg streptomycin (Sigma Aldrich, S6501) dissolved in 100 µL PBS. Where indicated, infected mice were identified as super-shedders (fecal shedding >108 cfu/gm) and moderate-shedders (fecal shedding <106 cfu/gm) at 15 days post infection.
To determine fecal bacterial load, fresh fecal pellets were collected by placing individual mice in sterile isolation chambers until 2–4 pellets were excreted. These were weighed and placed in 500 µL PBS. Pellets were resuspended via vortexing and CFUs were determined by plating log dilutions on LB agar plates containing 200 µg/mL streptomycin.
Fecal CFU was checked at 2–4 timepoints between 15 and 30 days post infection. Mice that consistently shed <106 cfu/gm at all times were identified as moderate-shedders while mice that shed >108 cfu/gm at all times were identified as super-shedders. Infected mice remained with their original cage mates even after determination of shedding status. Super-shedder status was also confirmed by verifying colonic and cecal inflammation after sacrifice of the animal.
Individual organs were collected, weighed and homogenized in 1 mL of PBS and log dilutions plated onto LB agar containing 200 µg/mL streptomycin.
Spleens from mice were mechanically dissociated into single cell suspensions in RPMI (Gibco, 11875) media with 10% fetal bovine serum (FBS) (Gibco, 26140) using glass slides. Spleens weighing less than 0.25 gm were suspended in a total volume of 10 mLs, those between 0.25–0.4 gm in a total volume of 15 mLs while spleens weighing more than 0.4 gm were resuspended in total of 30 mLs. Mesenteric lymph nodes were dissociated into single cell suspensions using a motorized pestle into 1 mL (Kontes, K749540-0000). All single cell suspensions were then filtered through a 70 µm cell strainer (BD Falcon, 352350). Blood was collected via cardiac puncture into a syringe containing 100 µL Heparin (BD, 366480). Samples were spun down for 10 minutes at 8000 rpm and serum removed and stored at −80°C. The cell pellet was then resuspended in 1 mL PBS. The blood sample was then treated with ACK buffer (0.15 M NH4Cl, 10 mM KHCO3, 0.1 mM Na2-EDTA, pH 7.4) to lyse red blood cells as described in [44]. Colons were isolated, opened, cleaned and washed four times with PBS. Colonic tissue was subsequently cut into 5–10 mm pieces and added to 10 mLs of RPMI containing 10% FBS, 0.5% Hepes (Sigma, H3375), 0.1% β Mercaptoethanol (Sigma, M3148) as well as the following enzymes at 1 mg/mL: Collagenase Type 1A (Sigma C9891), DNase I (Roche. 10104159001), Trypsin Inhibitor Type 1-S (Sigma, T6522). The tissue was gently agitated every 15 minutes and incubated for an hour at 37°C. The cells were then passed through a 70 µm filter and spun down at 1500 rpm. Leukocytes were then isolated from the colonic cells using CD45+ microbeads (Miltenyi Biotec, 130-052-301) as described in the product datasheet. All cells were then fixed with 1.6% paraformaldehyde (Electron Microscopy Sciences,15710) for 10 minutes at room temperature, washed twice with FACS buffer (PBS containing 0.5% BSA and 0.02% sodium azide). Cells were then either stored in methanol or permeabilized with saponin and stained as described in the flow cytometry section.
Cytokine stimulations were conducted as described in [26]. Briefly, single cell suspensions of splenocytes were recovered for 30 minutes at 37°C. Splenocytes were then left unstimulated or stimulated with 40 ng/mL IL-2 or IL-6 (BD Biosciences) for 15 minutes at 37°C. Splenocytes were then fixed as described previously, spun down and resuspended in cold methanol. Samples were stored at −80°C until staining for flow cytometry.
Samples stored in methanol were washed twice with FACS buffer. Cells were stained for 30 minutes with surface marker antibody cocktail comprised of Gr1-APC-Cy7, B220-PE-Texas Red, CD4-Alexa Fluor 700, CD11b-PerCP-Cy5.5, CD44 (-V500 or in-house conjugated to Pacific Orange), CD25-PE, Ki-67-Alexa Fluor 647 and CD62L-biotin (with streptavidin-quantum dot605 secondary) and phospho-Stat antibodies pStat5-Alexa Fluor 488 or pStat1-Alexa Fluor 488. All these antibodies were purchased from BD Biosciences. For measurement of transcription factors, FoxP3-PE and T-bet-Alexa Fluor 647/eFluor 660 were used (eBioscience). After staining, the cells were washed with FACS buffer and run on an LSR II flow cytometer (Becton Dickinson). Cells were acquired with DIVA software (BD Biosciences) and analyzed using FlowJo software (Tree Star). Cells were either measured as a percentage of total intact cells (determined by forward and side scatter measurements) or as a percentage of a specific cell type (e.g. total CD4 T cells). Alternatively, when measuring phospho-protein expression, median fluorescence intensity (MFI) of the cell populations was used. Staining for transcription factors was carried out using saponin permeabilization buffer (PBS containing 0.3% saponin, 0.5% BSA and 0.02% sodium azide). Cell populations were gated as described in Figure S10.
Mice were injected with 1 µg each of one of two different neutrophil depletion antibodies; anti-Ly6G (clone IA8, BioXcell, BE0075-1) and anti Gr1 (clone RB6-8C5, BioXcell, BE0075-1) or PBS controls. Antibodies or PBS were administered intraperitoneally every day for 3 days and mice sacrificed on the fourth day.
Serum was collected as described above and IFNγ (eBioscience, 88-8314-22) and IL-6 (BD Bioscience, 550950) levels were measured using sandwich ELISA kits.
ELISA plates were coated with Salmonella lysate for 1 hour at 37°c for 1 hour, then blocked with 3% Bovine Serum Albumin in PBS for 1 hour. Serum samples were then diluted 10 fold in wash buffer for a minimum of six serial dilutions. Samples were incubated for 2 hours at 37°c, washed, and incubated with Biotin conjugated anti- mouse IgG (Abcam, ab64255) for a further 2 hours at the same temperature. Plates were washed then incubated with Streptavidin conjugated Horse Radish Peroxidase (R&D,DY998) for 30 minutes, washed and incubated with TMB substrate (BD,555214) for 10 minutes, then stopped with 2N H2SO4. Titers were estimated by determining the lowest sample dilution with an optical density reading higher than undiluted serum from uninfected mice. All washes were carried out 5 times in between all steps using wash buffer consisting of 0.5% Tween in PBS.
Spleens were processed as described above and bone marrow was harvested by flushing a single tibia with 1 mL of RPMI containing 10% FBS. Cytologic specimens were prepared from single-cell suspensions of harvested bone marrow and spleen via concentration of the suspensions using a cytocentrifuge (Shandon Cytospin 4, Thermo Fisher Scientific, Waltham, MA). Slides were fixed and stained with modified Wright-Giemsa (Accustain, Sigma-Aldrich, St. Louis, MO). All slides were reviewed in a blinded fashion by a board-certified veterinary clinical pathologist. 1000-cell differential counts were performed with all myeloid and erythroid cells categorized as either proliferative or maturing (post-mitotic). For example, proliferative neutrophilic cells include myeloblasts, promyelocytes and myelocytes; maturing neutrophilic cells include metamyelocytes, band and segmented neutrophils. Neutrophils, eosinophils, macrophages, lymphocytes and plasma cells were further separated into individual categories.
Pegylated G-CSF (GenScript, Z00393-50) was resuspended in PBS at a final concentration of 1 mg/mL and injected intraperitoneally at a dosage of 1 ug/mouse. Control mice were injected with 100 ul of PBS intraperitoneally. Mice were injected for three days consecutively and sacrificed on the fourth.
Bone marrow-derived macrophages were prepared as previously described [45]. Five days after thawing, macrophages were plated at 2.5×105 cells per well in a 24 well dish (Corning). The cells were then infected at a multiplicity of infection of 5. Three hours after infection, supernatant was removed and 500 µL of the splenocyte single cell suspension added. After three hours of incubation, the cell suspension was collected. Cells were fixed, permeabilized with saponin and stained for intracellular IFNγ and transcription factors.
IL-2 antibody complex was prepared as described previously [29]. IL-2 mouse antibody (clone JES6-1, eBiosciences, 16-7022-81) was incubated with recombinant mouse IL-2 (eBiosciences, 14-8021-64) for 15 minutes before intraperitoneal injection into mice. Mice were injected for 2 days consecutively and sacrificed on the third. Control mice were injected with an equivalent volume of PBS.
Bubble plots in Figures 2 A,B and Supplementary Figure 1A were constructed using JMP software (SAS software, Cary, NC). Supplementary Figure 4C was visualized using MATLAB (MathWorks, Natick, MA). All other Figures were made using Prism (GraphPad, La Jolla, CA). All statistics were calculated using Prism and a two-tailed Mann-Whitney non-parametric test of significance was used unless otherwise mentioned. Spearman's correlations values were deemed significant using a two-tailed calculation based on the number of samples.
|
10.1371/journal.pntd.0001355 | Life Quality Impairment Caused by Hookworm-Related Cutaneous Larva Migrans in Resource-Poor Communities in Manaus, Brazil | Hookworm-related cutaneous larva migrans (CLM) is a common but neglected tropical skin disease caused by the migration of animal hookworm larvae in the epidermis. The disease causes intense pruritus and is associated with important morbidity. The extent to which CLM impairs skin disease-associated life quality has never been studied.
A modified version of the Dermatology Life Quality Index (mDLQI) was used to determine skin disease-associated life quality in 91 adult and child patients with CLM, living in resource-poor communities in Manaus, Brazil. Symptoms and signs were documented and skin disease-associated life quality was semi-quantitatively assessed using mDLQI scores. The assessment was repeated two and four weeks after treatment with ivermectin.
Ninety-one point five percent of the study participants showed a considerable reduction of skin disease-associated life quality at the time of diagnosis. The degree of impairment correlated with the intensity of infection (rho = 0.76, p<0.001), the number of body areas affected (rho = 0.30; p = 0.004), and the presence of lesions on visible areas of the skin (p = 0.002). Intense pruritus, sleep disturbance (due to itching) and the feeling of shame were the most frequent skin disease-associated life quality restrictions (reported by 93.4%, 73.6%, and 64.8% of the patients, respectively). No differences were observed in skin disease-associated life quality restriction between boys and girls or men and women. Two weeks after treatment with ivermectin, skin disease-associated life quality improved significantly. After four weeks, 73.3% of the patients considered their disease-associated life quality to have returned to normal.
CLM significantly impaired the skin disease-associated life quality in child and adult patients living in urban slums in North Brazil. After treatment with ivermectin, life quality normalised rapidly.
| Hookworm-related cutaneous larva migrans (CLM) is a parasitic skin disease common in developing countries with hot climates. In resource-poor settings, CLM is associated with considerable morbidity. The disease is caused by animal hookworm larvae that penetrate the skin and migrate aimlessly in the epidermis as they cannot penetrate the basal membrane. Particularly in the rainy season, the intensity of infection is high with up to 40 larval tracks in an affected individual. Tracks are very itchy and are surrounded by a significant inflammation of the skin. Bacterial superinfection is common and intensifies the inflammation. The psychosocial consequences caused by CLM have never been investigated. We showed that CLM causes skin disease-associated life quality impairment in 91 patients with CLM. Skin disease-associated life quality was significantly impaired. The degree of impairment correlated to the intensity of infection and the number of body areas affected. After treatment with ivermectin, life quality was rapidly restored.
| Hookworm-related cutaneous larva migrans (CLM) is a parasitic skin disease caused by the migration of animal hookworm larvae such as Ancylostoma braziliense, A. caninum or Uncinaria stenocephala in the epidermis. The infection occurs when third-stage larvae come into contact with human skin and penetrate into the epidermis. Since animal hookworm larvae cannot penetrate the basal membrane of the human host, they remain confined to the epidermis where they migrate for several weeks or months, and eventually die in situ [1]. CLM is frequent in impoverished rural and urban communities in countries with hot climates [2], [3], [4], [5], [6]. In these settings the prevalence of CLM can reach 4% in the general population and 15% in children <4 years. [6], [7], [8]. CLM belongs to the category of neglected tropical diseases [9], [10].
The main symptom of CLM is severe pruritus, which intensifies at night. The itching leads to sleep disturbance and day somnolence [6]. Scratching may cause extensive excoriations and subsequent bacterial superinfection of the lesions, typically by Streptococcus pyogenes or Staphylococcus aureus. Bacterial superinfection by group-A streptococci may induce the development of post-streptococcal glomerulonephritis [11].
A recent study on knowledge, attitudes and practice among mothers of children with CLM highlighted the psychosocial stress associated with this parasitic skin disease and its negative impact on family life (H.Lesshafft 2010, unpublished data). This prompted us to investigate the impairment of skin disease-associated life quality in patients with CLM in a semi-quantitative manner.
The study was carried out in Manaus, the capital of Amazonas State, North Brazil, from October 2008 to February 2009. Patients were actively recruited in resource-poor neighbourhoods, so called invasões. Patients were identified via word-of-mouth advertising through primary health care centres, neighbourhood organisations and community leaders. Twenty-three patients were recruited in Barrio da União and 28 in Nova Vitória; 40 patients came from five further resource-poor communities scattered in the city of Manaus. All communities were situated along small tributaries of the Amazon River (igarapés).
In these communities, most houses are built on stilts (palafitas) and made of wooden planks or recycled materials. Streets are unpaved, access to drinking water is precarious, sanitation is deficient and garbage is usually disposed in the adjacent igarapé or on the street. Dogs and cats stray around and feed on garbage found below and around the houses. In the rainy season, the communities are regularly inundated and animal faeces are widely dispersed.
Usually, households include two to six children. Blended family constellations, single mothers, adult illiteracy and unemployment are frequent. Alcoholism, psychological and physical violence and drug abuse are common.
The setting in which the study was carried out shares many social and economical characteristics with numerous other impoverished urban communities in South America. Most households in which the patients lived benefitted from the national Bolsa Familia and Bolsa Escola programs which support families with a monthly per capita income <140 Brazilian Reais (equivalent to 54 Euros at the time of study) with regular financial contributions.
The study is part of a larger research project on the epidemiology, morbidity, and control of CLM in North Brazil. Individuals aged ≥5 years with a diagnosis of CLM were eligible for the study. The investigation was performed as a prospective study with active case detection. Pregnant women and children <5 were excluded from the study because ivermectin treatment is contra-indicated in these groups. The study took place between October 2008 and July 2009.The diagnosis of CLM was made clinically. The whole skin was examined in a room where privacy was guaranteed and good lighting was available. The genital area was only inspected when the patient or his/her carer gave verbal consent. Children were always examined in the presence of their mothers. CLM was diagnosed when the characteristic elevated linear or serpingious track was visible and the lesion had moved forward during the preceding days [6], [12]. The number and the topographic localisation of each lesion was documented. Each track was defined as a single lesion, irrespective of the distance between the tracks. Tungiasis (jigger flea) and scabies, parasitic skin diseases also characterized by itching skin lesions, were excluded by careful clinical examination.
In order to determine the topographic distribution of the lesions and the affected area of the skin, the body surface was divided into right and left. As in previous studies each side was subdivided into 14 areas as follows: head, upper arm, forearm, hand, thorax, abdomen, back, buttock, genital/inguinal area, thigh, lower leg, ankle, back and sole of the foot [13]. Body areas were further classified into clearly visible areas (head, forearm, hand, lower leg, back and sole of the foot), partially visible areas (upper arm, thorax, abdomen, back, thigh) and non-visible areas (buttock, genital/inguinal area) according to local dress codes. Lesions were differentiated into papular, crusted-papular, and nodular [13]. The presence and dimensions of excoriations were documented. A simple lesion was defined as a track without bacterial superinfection, excoriations, or an significant inflammation presenting nodular lesion or an extended erythema. Bacterial bacterial superinfection was diagnosed when pustules, suppuration, or an abscess were present [6].
The severity of CLM was determined semi-quantitatively, using a severity score. This score combines the following variables: number of tracks (1–2 tracks = 1 point, 3–5 tracks = 2 points, 6–10 tracks = 3 points, >10 tracks = 4 points); presence/absence of secondary infection (0/2 points); signs of local inflammation (erythema, warmness or swelling = 1 point, pain = 2 points, nodular lesions = 3 points); presence of lymphadenopathy proximal to the lesion (0/1 point). Hence, the severity score can vary between 1 and 10 points.
Immediately after diagnosis patients were treated with ivermectin (200 µg/kg) in a single oral dose (Revectina; Solvay Farma Ltda, São Paulo, Brazil). Two and four weeks after treatment, the patients were re-examined and the mDLQI was determined again.
The Dermatology Life Quality Index (DLQI) was developed by Finlay and Khan in 1994 [14]. It is a validated instrument to assess skin-associated life quality impairment and it is the most frequently used tool to determine skin disease-associated life quality in patients with skin diseases of infectious and non-infectious origin [15], [16], [17].
The original DLQI questionnaire is available in English and in several other languages (www.dermatology.org.uk). In the present study, the Brazilian Portuguese translation was used. First, the wording was adapted to local culture and attitudes according to guidelines described by Cestari et al. [18]. Second, the questions were modified to focus on characteristic sequelae of parasitic skin diseases, and their impact on life quality in the setting of resource-poor communities in Brazil. Third, questions not applicable to children, such as the impact of skin disease on sexual life, were omitted in accordance with the original questionnaire for children [19]. This resulted in a modified dermatology life quality index (mDLQI) with eight items and a score varying between 0 and 24 points. The items were the following: pruritus, sleep disturbance, feeling of shame, need to adapt clothing in order to cover up skin lesions, problems faced at work or in school, impairment of leisure activities, impairment in personal relationships, teasing (only children), and problems concerning sexual relationships (only adults). The mDLQI has been validated by Worth et al. in patients with scabies living in a similar setting in northeast Brazil [20].
Since illiteracy was widespread, each statement was read out loud to the patient by one of the investigators (AS or HL) and its meaning explained in a standardized manner. The answers to each statement were weighted as follows: not at all = 0 points, a little = 1 point, quite a lot = 2 points, very much = 3 points [14]. The points for each statement were added up and formed the mDQLI for each patient. The mDLQI scores were categorised as shown in table 1.
The data were entered twice into a database using Epi Info software package Version 3.4.3 (CDC Atlanta, USA) and checked for errors which may have occurred during data entry. Data analysis was performed using SPSS for Windows (Version 16.0; SPSS Inc., Chicago, Illinois). Since data did not follow a normal distribution, the median and the interquartile range (IQR) were used as an indicator of central tendency and dispersion of the data, respectively. The Spearman rank correlation coefficient was calculated for correlations between mDLQI scores and other ordinal variables. The Mann-Whitney-U test was used to compare mDLQI scores between subgroups of patients. Relative frequencies were compared using the chi-squared test.
The study was approved by the Ethical Committee of the Fundação de Medicina Tropical do Amazonas (FMT-AM), the reference institution for tropical diseases of Amazonas State.
The objectives of the study were explained to each participant in simple and comprehensible Portuguese. The right to withdraw at any time was described in plain words. Patients had time to meditate about their decision and were given the possibility to discuss any doubts with the researchers. Each participant, or in the case of minors, their legal guardian, signed the written informed consent form. In case of illiteracy consent was given via fingerprint. The informed consent form was written and read out loud, and after each paragraph, the participant was asked whether she/he understood its meaning. Patients with other skin diseases than CLM were referred to the nearest primary health care centre or to the outpatient department of the FMT-AM, where treatment was provided free of charge.
Ninety-one patients were included in the study, 63 of them were male and 28 female. The median age was 10 years (IQR 7-12, range 5–44 years). The demographic and clinical characteristics of the patients are summarized in Table 2. Forty-four point eight percent of the patients had more than two lesions. The maximal number of lesions was 51. 88% of the patients had noted the appearance of the oldest track during the last four weeks. Figure 1 shows a typical example of an inflamed and superinfected track at a visible body part.
Nearly all study participants showed a reduction of life quality (mDLQI≥2 points) at the time of diagnosis (Table 3). The majority of the patients (51.6%) showed a moderate life quality impairment.
At baseline, the median mDLQI score was 5 (IQR 3-8). 6 (IQR 3-9) for adults and 5 (IQR 3-8) for children (p = 0.7; Table 4). Pruritus, sleep disturbance, feeling of shame and the need to dress differently were the most frequent restrictions. Significant differences in perceived restrictions between adult and child patients existed for problems faced at work/school and impairment in social relationships (p = 0.040 and p = 0.026, respectively). There was no difference in mDLQI scores between boys and girls (5 [IQR 3-8] versus 6 [IQR 3-7]; p = 0.86) and men and women (6 [IQR 3.-9] versus 4 [IQR 2-9]; p = 0.63).
The degree of skin disease-associated life quality impairment correlated strongly with the severity of the infection (rho = 0.76; p<0.001) (Figure 2) and the number of affected body areas (rho = 0.30; p = 0.004) (Figure 3). A significant correlation existed between the presence of lesions in clearly visible body areas and the mDQLI score (p = 0.002).Skin disease-associated life quality impairment did not depend on the number of CLM episodes experienced previously (p = 0.88), the duration of the infection (p = 0.52), or the presence or absence of bacterial superinfection (p = 0.80).
The follow–up examinations showed an improvement of skin disease-associated life quality two weeks after treatment with ivermectin (median mDLQI = 5 [IQR 3-8] versus 1 [IQR 0-3; p<0.001] Table 5). Four weeks after treatment, the median mDLQI score was zero and 82% of the patients reported a normalization of their skin disease-associated life quality. The normalization of skin disease-associated life quality was paralleled by a drastic reduction of the CLM severity score from a median of 4 points (IQR 3-6) to 1 point (IQR 1-1) two weeks after treatment with ivermectin and to 1 point (IQR 0-1) at the end of the study (both p<0.001). Figures 4 and 5 show the resolution of the inflammatory skin reactions around embedded hookworm larvae before and four weeks after treatment with ivermectin.
Diseases of the skin lead to various levels of suffering. First, they cause defined clinical pathology, such as visible inflammation, pruritus or pain. Second, skin diseases are frequently chronic in nature and patients have to take drugs, either topically or orally, for a protracted period of time. Third, if gross alterations of the skin are located on visible body parts, they may, at worst, lead to social withdrawal and/or to exclusion from society, as it is the case, for instance, with leprosy [21]. Additionally, patients may be confronted with ignorance or misconceptions regarding the aetiology of their skin disease, such as the fear that the condition is contagious or related to poor personal hygiene – assumptions which may lead to stigmatisation [22], [23]. Lymphatic filariasis with gross lymphoedema is a paradigmatic example of this category of skin diseases [24], [25], [26].
CLM is an extremely itchy skin condition characterized by signs of inflammationm such as erythema. Since lesions are frequently located at visible body parts they are difficult to hide from the public [13] and negative impact on emotional well-being of the patient is possible..In our study 94.5% of patients with CLM reported reduction of their skin disease-associated life quality with a median mDLQI score of 5 (Table 3). The degree of skin disease-associated life quality impairment was positively correlated with the intensity of the infection (Figure 2), the number of body areas affected (Figure 3), and the presence of lesions at clearly visible body parts.
In contrast to a study in patients with scabies [20] we did not find different degrees of impairment between women and men. This could be due to the fact that scabies lesions usually are less obvious to the patient and external observers/third parties than highly inflamed larval tracks. Besides, in scabies the lesions are frequently located at “hidden” topographic areas, such as the interdigital spaces. Finally, the preponderance of male participants in the study – a consequence of the higher prevalence of CLM in males in the area where the study was conducted – may have blurred the differences between the sexes.
The most common finding associated with an impairment of skin disease-associated life quality was pruritus (93.4% of the patients). Pruritus causes the patient to scratch repeatedly- a behavior which does not pass unnoted by other members of society [27]. In addition, since the intensity of itching increases at night, it causes alterations in the sleep pattern. The affective aspect of pruritus may induce a vicious cycle in which increasing mental harm and distress lead to increased itching which, in turn, augments scratching [27], [28].
Insomnia was reported as a cause of life quality impairment by 73.6% of the patients. A previous study has shown that CLM related insomnia manifests itself as a sleep maintenance disorder [13], probably due to an increased perception of pruritus during the night. In patients with pruritus-induced perturbation of sleep, quality and duration of sleep are reduced as a consequence of shorter non-REM sleeping periods [29]. This may cause daytime somnolence, irritability and psychological problems such as anxiety disorders [30], [31].
It seems paradoxical that insomnia has been cited as most important restriction by people living in an invasão. From an outside observer's point of view, getting rest and sleep in this setting seems to be very difficult anyway: poor housing and a high population density allow noise to enter the house almost unaltered and loud music is heard even late at night. However, our patients seem to have adapted to the extremely noisy environment of an invasão and considered sleep and recreation to be severely impaired by the CLM-related pruritus. In fact, after treatment, insomnia was reduced significantly already after two weeks (Table 3).
The feeling of shame was noted by 64.8% of the patients. In our study on knowledge, attitudes and disease perception, it was found that shame frequently resulted from the concept that the occurrence of CLM reflects poor personal hygiene and lack of care (H. Lesshafft, unpublished observation).
Another commonly noted restriction is related to the necessity of patients with CLM to dress differently. In the hot climate of northern Brazil a great part of the body remains uncovered. Hence, skin lesions are difficult to hide and the effort to cover them up with extra clothes or bandages may lead to a reduction of self-esteem and provoke shame and stigmatisation [23], [32], [33]. These somato-psychological interactions were confirmed by our finding that mDLQI scores were highest in patients in whom lesions were present at clearly visible parts of the body.
Problems faced at work or at school and impairment of personal relationships were also a frequently noted restriction of skin disease-associated life quality (Table 4). Several mechanisms may underlie these psycho-social consequences. First, and similarly to other skin diseases such as psoriasis, the erroneous assumption that CLM is contagious leads to alterations in personal relationships and eventually to social exclusion [23]. Second, as shown in a previous focus group discussion in the study area (unpublished data), mothers frequently ban affected children from playing outside, partly to prevent a new infection and partly to avoid teasing by other children, which may cause boredom and or lead to a feeling of social exclusion. Thirdly, bullying and interrupted personal relationships may provoke a feeling of disgust and shame about the skin condition and reinforce an active withdrawal from social networks due to the fear of stigmatisation [23], [22].
With regard to personal relationships, the significantly lower impairment of skin disease-associated life quality in children compared to adults might be explained by the fact that consciousness about their own appearance interferes less in children's relationships. The higher impairment perceived by adults at work is presumably related to a similar mechanism. At work, adults are confronted with the “outside world” in which CLM reflects a life in poverty. In contrast, children - going to school in the community - do not leave their social environment and consequently may perceive less life quality impairment.
Hitherto, only a few studies have attempted to determine skin disease-associated life quality impairment in tropical parasitic skin diseases.
While in patients with active cutaneous leishmaniasis or onchocerciasis, the average impairment was found to be higher than in the CLM patients of our study, skin disease-associated life quality restrictions in lymphatic filariasis caused a similar or higher impairment depending on the severity of lymphoedema [24], [25], [26], [32], [33], [34]. In contrast, patients with scabies living in an invasão in Northeast Brazil percieved less impairment than our patients with CLM [20]. In scabies the duration of infection, but not the number of infested body areas, correlated with skin disease-associated life quality impairment. This is probably due to the rather slow development of the clinical pathology in scabies, where the degree of skin alteration increases gradually, whereas in CLM inflammatory skin reactions develop within a couple of days.
We think that our data clearly indicate a cause-effect relationship between cutaneous larva migrans and impaired quality of life. First, the severity of disease was significantly correlated to the degree of impaired quality of life (rho = 0.76; p<0.001) and number of body areas affected (rho = 0.30; p = 0.004), indicating positive “dose-response” relationships. Second, already two weeks after the regression of skin lesions due to treatment with ivermectin, the degree of life quality impairment decreased significantly. Taken together, these findings provide substantial evidence that the impairment of life quality is the consequence of the parasitic skin disease as it has been observed in patients suffering from other parasitic infections [24]–[26] [32]–[34]. These findings also suggest that a treatment that costs approximately 40–80 eurocents, not only abrogates clinical pathology, but also averts stressful psycho-social consequences and prevents the development of secondary morbidity when given promptly.
When interpreting our results one has to take into account that skin disease-associated life quality of people living in misery in an urban slum is very low per se [35]. Housing is poor, sanitary infrastructure is deficient, crowding is common and social problems such as unemployment, alcoholism, illiteracy, and violence prevail. Obviously, these characteristics should mitigate perceived restrictions on skin disease-associated life quality in our patients. In fact, the results of another study in the same setting showed that members of the community considered parasitic skin diseases negligible in comparison to the existential problems of daily life (H. Lesshafft, unpublished observation).
In conclusion, CLM impairs the physical and mental wellbeing as well as social interaction of patients in a setting where skin disease-associated life quality is generally low. A single dose of ivermectin caused a complete resolution of the lesions within one month and restored skin disease-associated life quality to the normal level.
|
10.1371/journal.pntd.0004963 | Live Attenuated Leishmania donovani Centrin Knock Out Parasites Generate Non-inferior Protective Immune Response in Aged Mice against Visceral Leishmaniasis | Visceral leishmaniasis (VL) caused by the protozoan parasite Leishmania donovani causes severe disease. Age appears to be critical in determining the clinical outcome of VL and at present there is no effective vaccine available against VL for any age group. Previously, we showed that genetically modified live attenuated L. donovani parasites (LdCen-/-) induced a strong protective innate and adaptive immune response in young mice. In this study we analyzed LdCen-/- parasite mediated modulation of innate and adaptive immune response in aged mice (18 months) and compared to young (2 months) mice.
Analysis of innate immune response in bone marrow derived dendritic cells (BMDCs) from both young and aged mice upon infection with LdCen-/- parasites, showed significant enhancement of innate effector responses, which consequently augmented CD4+ Th1 cell effector function compared to LdWT infected BMDCs in vitro. Similarly, parasitized splenic dendritic cells from LdCen-/- infected young and aged mice also revealed induction of proinflammatory cytokines (IL-12, IL-6, IFN-γ and TNF) and subsequent down regulation of anti-inflammatory cytokine (IL-10) genes compared to LdWT infected mice. We also evaluated in vivo protection of the LdCen-/- immunized young and aged mice against virulent L. donovani challenge. Immunization with LdCen-/- induced higher IgG2a antibodies, lymphoproliferative response, pro- and anti-inflammatory cytokine responses and stimulated splenocytes for heightened leishmanicidal activity associated with nitric oxide production in young and aged mice. Furthermore, upon virulent L. donovani challenge, LdCen-/- immunized mice from both age groups displayed multifunctional Th1-type CD4 and cytotoxic CD8 T cells correlating to a significantly reduced parasite burden in the spleen and liver compared to naïve mice. It is interesting to note that even though there was no difference in the LdCen-/- induced innate response in dendritic cells between aged and young mice; the adaptive response specifically in terms of T cell and B cell activation in aged animals was reduced compared to young mice which correlated with less protection in old mice compared to young mice.
Taken together, LdCen-/- immunization induced a significant but diminished host protective response in aged mice after challenge with virulent L. donovani parasites compared to young mice.
| Visceral leishmaniasis (VL) is caused by the protozoan parasite Leishmania donovani. There is no effective vaccine available against VL for any age group and importantly, there are no previous studies regarding immune responses against experimental Leishmania vaccines tested in aged animals. We have reported earlier that immunization with a live attenuated L. donovani parasites (LdCen-/-) induced protective immune response in young animals viz, mice, hamsters and dogs. In this study we analyzed LdCen-/- mediated modulation of innate and adaptive responses in aged mice and compared to young mice. We observed that LdCen-/- infected dendritic cells from young and aged mice resulted in enhanced innate effector functions compared to LdWT parasites both in vitro and in vivo. Further, upon virulent challenge, LdCen-/- immunized young and aged mice displayed protective Th1 immune response which correlated with a significantly reduced parasite burden in the visceral organs compared with naïve challenged mice. Although there was no difference in the LdCen-/- induced dendritic cell response between aged and young mice; adaptive response in aged was reduced, compared to young which correlated with less protection in aged compared to young mice. This study supports the potential use of LdCen-/- as vaccine candidate across all age groups against VL.
| Visceral leishmaniasis caused by the protozoan parasite, Leishmania donovani, is fatal when left untreated. Epidemiological studies show that more than 90% of VL infections are concentrated in five countries viz, India, Brazil, Bangladesh, Nepal, and Sudan [1]. According to the World Health Organization report, Leishmaniasis accounts for roughly 1 out of every 1000 deaths due to infectious disease [2] and the mortality rates are similar for both young and old adults [2]. However, other reports suggest that 85% of cases of VL occur in children [3] thereby presenting a higher prevalence of VL in this group [4, 5]. In addition, there are contradictory reports on the outcome of cutaneous leishmaniasis (CL) infection caused by Leishmania major (L. major) parasites in aged mice [6, 7]. This disparity in the outcome of Leishmaniasis indicates a need to better understand the modulation of immune response in aged host during Leishmania pathogenesis.
With increased age, the immune system declines slowly in its efficiency to fight off infectious agents which in turn results in severity of symptoms and prolonged duration of infection [8, 9]. In addition, reactivation of chronic infections occurs at a higher frequency in aged population [7]. The dysfunctions in the immune system in the aged population are mainly caused by alterations in the components of the innate and adaptive immune systems. However, in the context of the innate immune system, there are substantial evidences suggesting that innate cells, specifically APCs (macrophage, dendritic cells), maintain unaltered immune response with aging [10–13]. Nevertheless, with regard to the adaptive immune system, there is evidence for broad-ranging, age-associated deficiencies in the development and function of B and T cells [14]. Specifically, aging is associated with diminished and/or altered cytokine patterns, expression of delayed type hypersensitivity reactions to antigens encountered earlier in life, and reduction in clonal expansion of Ag-specific T and B cells [11, 15]. Importantly, the impaired ability to mount adaptive immune responses to new pathogens may result in a higher susceptibility to infectious diseases and can cause an insufficient vaccine response [16]. Indeed, reduced immune responses to vaccination have been observed for variety of vaccines including Streptococcus pneumoniae, influenza, hepatitis, and tetanus [17–20]. However, Tdap vaccine (‘tetanus toxoid, reduced diphtheria toxoid and acellular pertussis’) was found to be immunogenic in subjects ≥65 years, and resulted in a satisfactory but diminished protective antibody response in the elderly compared to young adults [21]. Additionally, it was demonstrated that live attenuated VZV vaccine (“zoster vaccine”) can elicit a significant increase in cell-mediated immunity to VZV in immunocompetent older adults which eventually prevents reactivation of herpes zoster partially or attenuates the severity of post-herpetic neuralgia [22–26]. The non-uniformity in host response against different vaccines suggests that the immunological mechanisms induced by vaccines in distinct age groups are not fully understood.
Over the past decades, chemotherapeutic treatment of VL with different type of drugs has shown only moderate success due to serious side effects and widespread emergence of drug resistant strains [27]. Given the insufficiencies associated with current treatments, vaccination could to be a practical alternative for an effective preventative control for the disease [28]. Previous attempts at vaccination based on killed Leishmania parasites or defined parasite antigens resulted in a limited protection [29, 30]. It is known that a cure from disease either due to a natural cutaneous infection, cutaneous leishmanization [31] or visceral leishmaniasis [32–34] protects the individual from reinfection. The notion that persistence of a few parasites in the body can sustain a life-long protection against leishmaniasis implies that adequately attenuated live Leishmania parasites can likewise yield protection [35–37]. Live-attenuated vaccines allow the host immune system to interact with a wide repertoire of antigens considered to be crucial in the development of protective immunity and more importantly cause no pathology [38]. However, a significant barrier to development and evaluation of vaccine candidates against leishmaniasis is to define the immunological correlates associated with protection both in young and old individuals. Therefore studies are needed to evaluate the efficacy of live attenuated vaccine candidates in all aged populations.
To address these questions, our laboratory has developed a L. donovani strain deleted for the centrin1 gene (LdCen-/-) that is essential for parasite virulence [39, 40]. LdCen-/- parasites have been shown to persist a maximum of 5–8 weeks in mice [40]. We have previously demonstrated safety, immunogenicity of LdCen-/- and protection against infection with virulent L. donovani as indicated by control of parasitemia and the induction of a protective adaptive immune response in various animal models (mice, hamster and dog) [40–43]. Of particular note, all these studies were performed in young animals. It is yet to be ascertained whether LdCen-/- immunization can similarly be protective in aged animals. In the present study, we report the immunogenicity profile of LdCen-/- parasites in aged mice (18 months old) and compared it with that in young mice (2 months old) in terms of both innate and adaptive immune response. We investigated the innate response in bone marrow derived dendritic cells (BMDCs) isolated from young and aged mice after in vitro infection with LdCen-/-. We found that LdCen-/- infection significantly enhanced production of pro-inflammatory response in both age groups compared to LdWT parasite infection. Further, LdCen-/- infected BMDCs derived from both young and old mice were able to promote the proliferation of OVA-specific CD4+T cells and induced strong Th1 type immune responses in vitro. Similarly, parasitized splenic dendritic cells isolated from LdCen-/- infected young and aged BALB/c mice after 4 days of infection displayed up-regulation of many proinflammatory cytokine genes along with generation of enhanced CD4+Th1 response compared to LdWT infected mice. In vivo studies further showed that immunization with LdCen-/- protected both young and aged mice from a challenge with virulent L. donovani parasites via generating enhanced Th1 predominant immune response along with generation of NO from restimulated splenocytes. Comparison of both innate and adaptive immune response showed that aged animals induced lower adaptive response with no change in the innate response. The reduced adaptive immune response correlated with satisfactory but reduced protection in aged mice compared to young mice upon virulent L. donovani challenge.
5 to 6-wk old Female BALB/c mice were obtained from the National Cancer Institute, National Institute of Health, Bethesda, MD, USA. Mice belonging to younger group were 8-wk old and mice belonging to aged group were 72-wk old. All mice were maintained in the FDA/CBER AAALAC-accredited facility under standard environmental conditions for this species. Among parasites, the wild type L. donovani (LdWT) (MHOM/SD/62/1S) maintained in golden Syrian hamsters and centrin1 gene–deleted (LdCen-/-) line of L. donovani (Ld1S2D) were used [39, 40]. The parasites were cultured according to the procedure previously described [39, 44]. Red fluorescent protein (RFP)-expressing LdWT parasites were developed using the pA2RFPhyg plasmid for integration of a RFP/ Hygromycin B resistance gene expression cassette into the parasite 18S rRNA gene locus as described previously [45]. mCherry expressing LdCen-/- parasites were generated using the pLEXSY-cherry-sat2 plasmid and following the company’s protocols (Jena Bioscience). The parasites were cultured according to the procedure previously described [44].
The animal protocol for this study has been approved by the Institutional Animal Care and Use Committee at the Center for Biologics Evaluation and Research, US FDA (ASP 1995#26). Further, the animal protocol is in full accordance with ‘The guide for the care and use of animals’ as descried in the US Public Health Service policy on Humane Care and Use of Laboratory Animals 2015 (http://grants.nih.gov/grants/olaw/references/phspolicylabanimals.pdf).
The young and aged mice were immunized via tail vein with 3X106 stationary phase LdCen-/- promastigotes; 5-wk post-immunized mice were then challenged via tail vein with 105 virulent L. donovani (LdWT) metacyclic parasites. In each study, at least 6 mice were used per group. Age matched naive mice used as controls were also similarly challenged with 105 virulent L. donovani metacyclic parasites. After 10-wk of post-challenge period, parasite load was recorded from spleens of challenged mice by culturing the separated host cell preparations by limiting dilutions as previously described [40]. After 4-wk post challenge period proinflammatory cytokines, nitric oxide, antibody titers and T cell effector responses were measured.
Splenocytes were plated in 24-well culture plates and stimulated with freeze-thaw L. donovani Ag (80μg/ml FTAg) in complete RPMI 1640 medium and cells were incubated at 37°C in 5% CO2, with 95% humidity. After 72h of culture, cell supernatants were collected and stored in −80°C until cytokines were analyzed using multiplex kits, MILLIPLEX Mouse Cytokine/Chemokine Magnetic Bead Panel (Millipore). The plate was read in a Luminex-100 (Luminex) system using Bio-Plex Manager software 5.0. The cytokine analysis procedure has been performed according to the manufacturer’s instructions, and the level of cytokine concentration was measured using a standard curve of each specific cytokine.
Splenocytes or macrophages obtained from peritoneal fluid were cultured in complete RPMI 1640 medium in the presence of FTAg (80μg/ml) for 24h at 37°C. NO (nitrite/nitrate) production was determined from the supernatants of the cultures by the Griess Reaction Kit (Sigma-Aldrich) [40].
The IgG specific Ab responses were measured by conventional ELISA method. Briefly, ELISA plates were coated overnight at room temperature with FTAg (80 μg/ml). A serial dilution of the sera was carried out to determine the titer, which is defined as the inverse of the highest serum dilution factor giving an absorbance of >0.2. Titers for the Abs were determined using the following HRP-conjugated secondary Abs: Rabbit anti-mouse IgG (H+L)–HRP, Rabbit anti-mouse IgG1-HRP, Rabbit anti-mouse IgG2a-HRP (Southern Biotech; all with 1:1000 dilutions). SureBlue (KPL) was used as a peroxidase substrate. After 15 min, the reaction was stopped by the addition of 100 μl 1M H2SO4, and the absorbance was read at 450 nm.
The proliferative capacity of T cells was assessed by a CFSE dilution assay in LdCen-/- immunized mice before and after challenge with virulent L. donovani parasites. Age-matched naive mice served as negative controls for Ag-specific proliferation. Splenocytes from different groups of mice were isolated, incubated in 5μM CFSE (Molecular Probes/Invitrogen) for 10 min in RPMI 1640 without fetal calf serum (FCS), followed by 5 min of quenching in ice-cold RPMI 1640 plus 10% FCS, and subsequently washed thoroughly before plating in 96-well tissue culture plates at 2×105 cells/well. Cells were cultured for 5 days at 37°C with 5% CO2 under stimulation with FTAg (80 μg/ml). Cells were harvested, washed, and blocked with anti-CD16/32 (5 μg/ml) for 20 min (4°C) and stained with anti-mouse CD3 allophycocyanin–eFluor 780, anti-mouse CD4 eFluor@450, anti-mouse CD8a eFluor@605NC (eBioscience, USA) (each with 1:200 dilution; 4°C) for 30 min. For analysis, single live cells (dead cells were excluded based on staining with the LIVE/DEAD Aqua dye) were gated for CFSE stained CD4+T cells and CD8+T cell and proliferation was calculated. Cells were acquired on an LSR II (BD Biosciences) equipped with 405-, 488-, 532-, and 638- nm laser lines using FACSDiva 6.1.2 software. Data were analyzed with FlowJo software version 9.1.5 (Tree Star, San Carlos, CA).
Splenocytes isolated from young and aged mice were cultured in 24-well plates in complete RPMI 1640 medium at 37°C and stimulated with FTAg (80 μg/ml). After 48h at 37°C, protein transport inhibitor (BD GolgiStop; BD Pharmingen) was added to the wells and the plate was incubated at 37°C for 6h. 6h after, cells were then blocked at 4°C with rat anti-mouse CD16/32 (5 μg/ml) from BD Pharmingen for 20 min. For surface staining, cells were then stained with anti-mouse CD3 APC-eFluor@780, anti-mouse CD4 eFluor@450, anti-mouse CD8a eFluor@605NC, anti-mouse CD44 FITC, and anti-mouse CCR7 PE-Cy5 Abs for 30 min (each with 1:200 dilution; 4°C). The cells were then stained with Live/Dead fixable aqua (Invitrogen/Molecular Probes) to stain dead cells. Cells were washed with wash buffer and fixed with the Cytofix/Cytoperm Kit (BD Biosciences) for 20 min (room temperature). Intracellular staining was done with anti-mouse IFN-γ PE-Cy7, anti-mouse TNF PerCp-eFluor@710, anti-mouse IL-2 BV711, anti-mouse IL-10 APC for 30 min (each with 1:300 dilution; 4°C). Cells were acquired on an LSRII (BD Biosciences) equipped with 405, 488, 532, and 638 laser lines using DIVA 6.1.2 software. Data were analyzed with FlowJo software version 9.1.5 (TreeStar). For analysis, first doublets were removed using width parameter; dead cells were excluded based on staining with the LIVE/DEAD Aqua dye. Lymphocytes were identified according to their light-scattering properties. CD4 and CD8 T cells were identified as CD3+ lymphocytes uniquely expressing either CD4 or CD8. Upon further gating, intracellular cytokines were measured in CD44HiCCR7Low cells. Fluorescence minus one controls were used for proper gating of positive events for designated cytokines.
Dendritic cells were cultured in vitro from bone marrow progenitors. Briefly, both young and aged BALB/c mice were sacrificed and their femurs and tibias were excised, cleaned of tissue, and flushed with RPMI medium. Bone marrow was isolated by depletion of erythrocytes with ACK lysis buffer, and cultured with complete RPMI medium supplemented with 10% (v/v) fetal bovine serum (FBS) and 1% penicillin (20 U/ml)/ streptomycin (20 μg/ml) and 20 ng/mL GM-CSF (Peprotech, London, United Kingdom) and IL-4 (Peprotech) for 7 days to obtain >75% purity of CD11c+ DCs. For Phagocytosis and parasite clearance assay DCs were infected with parasites (5:1, parasite to DCs ratio). After incubation for 6h at 37°C in 5%CO2, the cultures were washed with culture media to remove extracellular parasites, and the cultures were incubated in RPMI media for maximum of 72h. At 6, 24, 48 and 72h post-infection, cells were washed with PBS, fixed with methanol, stained with Giemsa, and intracellular parasite numbers were evaluated microscopically. For cytokine and NO measurements, DCs were infected with parasites and stimulated with LPS (1μg/ml) for 24h. Culture supernatants were collected at 24h post infection to evaluate cytokine (IL-12, TNF and IL-6) production with the use of sandwich ELISA kit (ebioscience) and nitric oxide production by the Griess assay. The assay was performed according to the manufacturer's instructions.
The young and aged mice were infected via tail vein with 3 X 106 stationary phase red fluorescent LdWT RFP, or LdCen-/-m-cherry promastigotes. In each study, at least 6 mice were used per group. Age-matched naive mice used as control. At day 4 post infection, mice were sacrificed and parasite load was recorded from spleens of the LdWT and LdCen-/- infected mice by culturing the separated host cell preparations by limiting dilutions as previously described [40].
In a separate experiment the splenic DCs were sort selected. Single-cell suspensions were prepared from spleens, and RBCs were lysed using ACK lysing buffer. Dendritic cells were enriched using pan dendritic cell isolation kit (Miltenyi Biotec) and subsequent sort selecting of Leishmania (RFP/m-Cherry) infected splenic DCs and uninfected splenic DCs from different age groups of infected mice were done by high speed FACS cell sorter system (BD FACS Aria-IITM). Infected splenic DCs were sorted by gating live single enriched DCs for Cd11c+Ly6G-Cd11b-RFP/m-Cherry+. Uninfected splenic DCs after enrichment [Cd11c+ Ly6G-Cd11b- RFP/ m-Cherry-] were also sort selected.
Total RNA was extracted from the parasitized/uninfected splenic DCs using RNAqueous Microkit (AM1931) and RNA was extracted from total mouse splenocytes using PureLink RNA Mini kit (Ambion). Total RNA (400 nanograms) was reverse transcribed into cDNA using random hexamers by a high-capacity cDNA reverse transcription kit (Applied Biosytems). Gene expressions were determined by TaqManGene Expression Master Mix and premade TaqMan Gene Expression assays (Applied Biosystems) using a CFX96 Touch real-time system (Bio-Rad, Hercules, CA). The data were analyzed with CFX Manager soft-ware. The TaqMan Gene Expression Assay ID (Applied Biosystems) of different primers used was as follows: TNF, Mm00443258_m1; IL-12, Mm00434174_m1; IFN-γ, Mm01168134_m1; IL-6, Mm00446190_m1; IL-4, Mm00445259_m1; IL-10, Mm00439614_m1; and GAPDH, Mm99999915_m1. Expression values were determined by the 2- DD Ct method where samples were normalized to GAPDH expression and determined relative to untreated sample.
Statistical analysis of differences between means of experimental groups was determined by unpaired two-tailed Student t test, using Graph Pad Prism 5.0 software. A p value <0.05 was considered significant, and a p value <0.005 was considered highly significant.
Our previous studies have shown that BMDCs from young mice manifest heightened effector function in response to LdCen-/- infection in vitro [46]. Hence, in this study we compared the ability of LdCen-/- parasites to modulate DC responses in BMDCs in vitro from both young and aged mice. LdWT and LdCen-/- parasites showed similar rates of infection along with identical parasite loads at 6h post-infection in both young and aged BMDCs (Fig 1A and 1B). These BMDCs cultures were then subsequently examined at 24, 48 and 72h post-infection, and the percentage of infected DCs and parasite loads were calculated. Interestingly, BMDCs from both young and aged mice infected with LdCen-/- displayed significantly lower percentage of infected cells (Fig 1A) and a lower number of parasites per infected cell (Fig 1B), compared to those infected with LdWT at all the time points post-infection. This indicates that LdCen-/- parasites infect both young and aged BMDCs similarly but are unable to persist for prolonged periods compared to LdWT parasites. DCs play a crucial role in coordinating immune response in leishmaniasis by providing the IL-12 necessary for the induction of protective Th1 immune response [47, 48]. LdCen-/- parasite infected young and aged BMDCs produced significantly more IL-12 than did those infected with LdWT parasites (Fig 1C). Next, we analyzed the level of proinflammatory cytokines (TNF, IL-6) and nitric oxide (NO), the key components in Leishmania control, in BMDCs cell supernatants 24h post infection. Interestingly, LdCen-/- infected BMDCs from both age groups secreted significantly more TNF, IL-6 and NO compared to LdWT infected cells (Fig 1D, 1E and 1F). Altogether, these observations suggest that BMDCs from both young and aged mice undergo a more pronounced effector response following infection with LdCen-/- compared to LdWT infected BMDCs. Importantly, BMDCs from aged mice showed similar innate response like that of BMDCs of young mice in response to LdCen-/- infection.
Next, we compared the antigen presentation capability of LdCen-/- infected BMDCs by measuring their ability to present LdCen-/- antigens and activate CFSE-labeled naïve CD4+ T cells as measured by IFN-γ and IL-10 production. BMDCs from young and aged mice were pulsed with OVA peptide followed by infection with LdWT or LdCen-/- and then co-cultured with OVA specific (DO11.10) CFSE labeled CD4+T cells. CD4+T cell proliferation was evaluated after 5 days by flow cytometry. LdCen-/- infection induced the proliferation of T cells in both young and aged BMDCs and was significantly higher compared to those co-cultured with LdWT infected BMDCs as demonstrated by the histogram overlay depicting the individual % proliferated CD4+T cells from different sets (Fig 1G) and the subsequent quantitative bar diagram (Fig 1H). Interestingly, LdCen-/- mediated induction of CD4+T cell proliferation was similar in young and aged BMDCs thereby highlighting the similar dendritic cell activation. Cytokine production resulting from CD4+T cell response was measured in culture supernatants after 5 days of co-culture. IFNγ levels from LdWT and LdCen-/- infected BMDCs–T cell co-cultures did not differ significantly between the two age groups and were comparatively higher in LdCen-/- infected BMDCs (Fig 1I). On the other hand, LdWT infected BMDCs–T cell co-cultures significantly induced IL-10 production compared to LdCen-/- infected BMDCs-T cell co-cultures in both age groups (Fig 1J). The IFN-γ: IL-10 ratio was significantly higher in LdCen-/- infected BMDCs-T cell cultures compared to LdWT infected BMDCs-T cells co-cultures (Fig 1K). Thus in vitro results suggest that LdCen-/- infected BMDCs from both age groups are capable of inducing the proliferation of Th1 type of CD4+T cells in vitro that may lead to enhanced adaptive immune response and also there is no diminution of LdCen-/- induced innate response between two age groups.
Further, we analyzed LdCen-/- mediated modulation of the dendritic cell (DCs) response in spleen using an in vivo mouse model. We infected young and aged mice intravenously with red fluorescent LdWT RFP, or LdCen-/- m-Cherry parasites. After 4 days we enriched DCs from spleen and sort selected uninfected DCs (UI DC) by gating live single cells for Cd11c+Ly6G-Cd11b-RFP/m-Cherry- whereas parasitized DCs (Inf DC) were sorted by gating live single cells for Cd11c+Ly6G-Cd11b-RFP/m-Cherry+ (Fig 1L). We observed, LdCen-/- infected young and aged mice had comparatively lower although not statistically significant level of splenic parasite burden compared to LdWT infected mice (Fig 1M) which is in agreement with our in vitro observation. We also assessed the expression of proinflammatory and anti-inflammatory cytokine gene expression in LdCen-/- infected DC population and compared them with LdWT parasitized DCs. RT-PCR analysis showed that 4 major proinflammatory cytokines such as IL-12 (Fig 1N), IL-6 (Fig 1O), IFN-γ (Fig 1P) and TNF (Fig 1Q) were significantly elevated in parasitized splenic DCs isolated from LdCen-/- infected young and aged mice compared to LdWT infected mice. In contrast, anti-inflammatory cytokine IL-10 (Fig 1R) expression was significantly reduced in parasitized DCs isolated from LdCen-/- infected young and aged mice compared to LdWT infected mice. We did not see any significant difference in the behavior between infected DCs (Inf DC) and bystander DCs (UI DC) from LdWT infected young and aged mice. However, we observed that bystander DCs from LdCen-/- infected mice resulted in comparatively lower production of proinflammatory cytokines and higher production of anti-inflammatory cytokines compared to infected DCs from both age groups (Fig 1N–1R). Moreover, un-infected splenic DCs (UI DC) from LdCen-/- infected mice showed slightly higher but not statistically significant level of proinflammatory cytokines (IL-12, IL-6, IFN-γ and TNF) along with down regulation of IL-10, compared to uninfected DCs from LdWT infected young and aged mice (Fig 1N–1R).
Additionally, we have investigated the antigen presentation capability of dendritic cells sort selected from young and aged mice which were infected with LdWT or LdCen-/- parasites for 4 days. We pulsed these DCs with OVA peptide followed by co-culture with OVA-specific CD4+T cells from both age groups for 5 days to analyze OVA-specific proliferation of T cells. Consistent with our in vitro study, the percentage of proliferating CD4+T cells after 5 days were comparatively higher upon co-culture with OVA-pulsed DCs from LdCen-/- infected young and aged mice compared to LdWT infected mice (S1B Fig). Cytokine production resulting from such proliferated CD4+T cell response was measured by flow cytometry. The results showed that DCs from LdCen-/- parasite infected young and aged mice stimulated CD4+ T cells to produce significantly higher levels of IFN-γ/IL-10 ratio compared to CD4+T cells cocultured with DCs from LdWT infected mice (S1C Fig), thereby indicating a similar Th1 response predominance in both age groups of mice. Thus, from the in vivo gene expression profile and antigen presentation assay, it is also evident that LdCen-/- infection results in similar and significant enhancement of innate effector responses in young and aged mice.
The ability of LdCen-/- to protect against virulent L.donovani infection was determined in 5-wk immunized mice followed by 10-wk of challenge with virulent L. donovani parasites. The results showed that immunization with LdCen-/- significantly reduced spleen and liver parasite burden in both young and aged mice compared to naive-challenged mice (Fig 2A and 2B). The parasite burden in spleen and liver of naive challenged aged mice was similar like that of naïve challenged young mice (Fig 2A and 2B). However, the reduction in parasite burden was significantly less in the LdCen-/- immunized–challenged aged group (13 fold in spleen and 8 fold in liver) than the immunized–challenged young mice (49 fold in spleen and 19 fold in liver; data in Fig 2A and 2B). Overall, these data suggest that LdCen-/- immunization confers substantial protection in young and aged mice against virulent L. donovani challenge; however, the protection is reduced in aged mice.
Effective clearance of Leishmania parasites requires a proinflammatory response which is mainly mediated by IFN-γ [49–53]. In order to evaluate the protective immunity induced after LdCen-/- immunization in both young and aged mice, we analyzed Ag-specific cytokine secretion by splenocytes from naive, immunized (5-wk), non-immunized challenged (naïve challenged) and immunized challenged mice (5-wk immunization + 4-wk of challenge) (Fig 3). Comparative cytokine profiles demonstrated a significant induction of Leishmania Ag–specific IFN-γ, IL-12, and TNF secretion in the culture supernatants of 5-wk post-immunized young and aged mice compared to naïve mice at protein level (Fig 3A–3C). Additionally, the steady state cytokine gene expression pattern at mRNA level (S2A–S2C Fig) was similar with Ag stimulated cytokine secretion at protein level. Interestingly, following challenge with virulent L. donovani parasites, the immunized young mice showed significantly enhanced IFN-γ, IL-12 and TNF secretion at both protein (Fig 3A–3C) and mRNA level (S2A–S2C Fig), while in aged mice significantly enhanced secretion of IL-12, TNF but not IFN-γ was observed (Figs 3A–3C and S2A–S2C). Importantly, overall induction of all the proinflammatory cytokines (IFN-γ, IL-12 and TNF) was significantly lower in aged mice compared to young mice from all the groups at both protein (Fig 3A–3C) and mRNA level (S2A–S2C Fig). Additionally, in immunized young and aged mice, compared with naive mice, significant induction of anti-inflammatory cytokines like IL-10 and IL-4 was observed at protein level (Fig 3D and 3E) whereas at mRNA level comparatively higher although statistically non-significant levels of IL-10 and IL-4 were observed (S2D and S2E Fig). However, young and aged immunized-challenged mice showed slightly higher although not statistically significant levels of IL-10 and IL-4, compared to naïve challenged groups at both the protein (Fig 3D and 3E) and mRNA levels (S2D and S2E Fig). Importantly, the IFN-γ: IL-10 or IFN-γ: IL-4 ratios were higher in immunized as well as immunized challenged young and aged mice compared to naïve and non-immunized challenged groups respectively (Fig 3F and 3G). Interestingly IFN-γ: IL-10 or IFN-γ: IL-4 ratios in immunized challenged aged mice were significantly less compared to young mice. Taken together, these results suggest that LdCen-/- immunization initiates mixed pro- and anti-inflammatory immune response against virulent L. donovani challenge in both young and aged mice albeit lower proinflammatory response in aged mice.
Since the preferential production of proinflammatory cytokines specifically IFN-γ results in increased synthesis of host protective nitric oxide (NO) by activated macrophages [54, 55], we determined the level of NO produced by Leishmania Ag restimulated splenocytes derived from the LdCen-/- immunized young and aged mice before and after challenge. It is important to note that NO product observed in the splenocytes is mainly from the macrophages present in such cultures. A significant amount of Leishmania Ag specific nitrite production was observed in both young and aged post immunized mice compared with non-immunized mice splenocytes (Fig 4A). Moreover, a much greater amount of NO was observed in LdCen-/- immunized young and aged mice upon virulent L. donovani challenge than in naive challenged mice (Fig 4A). Of note, the induction of nitrite concentration was significantly lower in aged mice compared to young mice from all the groups. Additionally, since, higher level of anti-leishmania antibody (IgG2a) dictates predominant Th1response [56, 57]; we analyzed sera from immunized-challenged young and aged mice for determination of Leishmania–specific IgG, IgG1, and IgG2a titers. The Leishmania Ag-specific IgGTotal, IgG1, and IgG2a production was significantly higher in immune challenged young and aged mice compared to naïve challenged groups (Fig 4B). Of note, specific increase in Th1 mediated IgG2a titers in immunized and challenged mice sera from both groups were observed. Importantly, the levels of all three Ab populations were reduced in aged mice compared to young mice from all the groups.
Aging is known to be associated with a decline in T cell proliferation [11, 58]. Hence to investigate if LdCen-/- vaccination induces Leishmania Ag specific T cell proliferation in young and aged mice, we performed proliferation assays using splenocytes from naïve, 8-wk post immunized, naive challenged (4-wk of challenge) and immunized challenged (8-wk immunization + 4-wk of challenge) mice that were stimulated with Leishmania FTAg. Results showed that both CD4 and CD8 T cells isolated from both young and aged mice immunized with LdCen-/-, have significantly higher proliferative capacity in response to Leishmania FTAg compared to T cells isolated from naive mice (Fig 5B). Further, T cells isolated from mice immunized and challenged with virulent L. donovani parasites have significantly higher proliferative capacity than naive challenged group. The significant proliferative capacity of T cells even in 8-wk post-immunized young and aged mice suggests that some of them may have derived from a memory T cell response since by that time the majority of mice have cleared the LdCen-/- parasite [40]. Interestingly significantly lower proliferative capacity of only CD8 T cells was observed in the spleen of immunized aged mice compared to young mice before and after challenge.
Since efficacy of several experimental Leishmania vaccine candidates has been shown to correlate with the T cells secreting multiple pro-inflammatory cytokines [59], we investigated whether LdCen-/- immunization could induce splenic CD4+ and CD8+ T cells to secrete multiple cytokines such as IFN-γ, IL-2 and TNF in infected aged mice as was previously observed in young mice [40]. Specifically, we demonstrated cytokine production from the antigen experienced effector memory T cells (CD44Hi/CCR7Low) in 8-wk post-immunized young and aged animals. Spleen cells cultured in vitro in presence of Leishmania FTAg followed by multicolor staining were gated according to their surface expression of CD44 and CCR7. Seven distinct populations of cytokine-producing cells were defined from the CD44Hi/CCR7Low cells based on different combinations of IFN-γ, IL-2 or TNF (Fig 6A). The results showed that at 8-wk post-immunization in both young and old mice, significantly higher number of CD4 and CD8 T cells were single, (IFN-γ, IL-2 and TNF) double (IFN-γ+ TNF+, IFN-γ+IL-2+, TNF+IL-2+) and triple cytokine producing cells (IFN-γ+ TNF+ IL-2+) (Fig 6B and 6C) compared to naïve mice. However, the number of single and multiple cytokine producing cells were significantly lower in aged mice compared to young mice. Similar analysis conducted with spleen cells of immunized and challenged young and aged mice showed significantly higher percent of the single and multiple cytokine-producing CD4 and CD8 T cells compared to naive challenged mice (Fig 7A and 7B). Again we observed a reduced percent of single and multiple cytokine expressing CD4 and CD8 T cell in aged immunized challenged mice.
We also quantified Ag-experienced CD4 T cells that produce IL-10, a crucial anti-inflammatory cytokine in the pathogenesis of VL from the spleen of young and aged mice (Fig 7C). Ag-experienced CD4 T cells from LdCen-/- immunized and challenged mice produced significantly lower IL-10 in both age groups compared to naïve challenged mice (Fig 7C). Since the balance between pro and anti-inflammatory cytokines predicts the outcome of protective immune response, we measured IFN-γ/IL-10 ratio. In Ag-stimulated CD4 T cells, the IFN-γ/IL-10 ratio was significantly higher in the immunized mice both before challenge and after challenge compared with either naive or naive-challenged controls (Fig 7D). However in aged mice the IFN-γ/IL-10 ratio was significantly lower compared to young mice. In summary, LdCen-/- immunized young and aged mice induced a strong antigen experienced effector T cell memory response after 8-wk post-immunization, at a time point when the majority of mice have cleared the LdCen-/- parasite. However, the effector T cell mediated immune response is lower in aged mice.
Immune dysfunction is a hallmark of aging leading to an increased susceptibility to infectious diseases in older individuals [16, 60]. In an older adult, the benefits of vaccination to prevent infectious disease are limited, mainly because of the adaptive immune system’s inability to generate protective immunity [16]. This has been observed for a variety of vaccines including Streptococcus pneumoniae, influenza, hepatitis, and tetanus and is thought to be the result of a compromised immune response [17–20]. However, on the other hand Tdap vaccine (‘tetanus toxoid, reduced diphtheria toxoid and acellular pertussis’) gives a satisfactory but diminished protective response in the elderly (≥65 years of age) compared with that in young adults [21]. Additionally, live attenuated zoster vaccine renders effective but partial protection against herpes zoster and postherpetic neuralgia in immunocompetent persons 60 years of age and older via boosting cell-mediated immunity to varicella–zoster virus (VZV) [22]. These differences in the efficacy of vaccines in elderly population suggest that there is a need to validate the efficacy of each vaccine in aged population. As a first step, it is important to analyze the essential components of both innate and adaptive immunity in aged and compare with the young to find out the essential differences between the two, which can be exploited to enhance the immunity of vaccines for aged against various infectious diseases including Leishmaniasis.
Till date no licensed Leishmania vaccines exist for any age groups. Our laboratory has shown, using genetically modified LdCen-/- parasites, the induction of host protective innate [61], adaptive immune response and long-term protection against virulent L. donovani infection in younger animals (mice, hamsters and dogs) [40–43]. Since, there are no reports regarding the immuno-protective role of experimental Leishmania vaccines including live attenuated parasites in aged animals, in this report we have performed a comprehensive study to systematically analyze the preclinical efficacy of LdCen-/- parasites vis-à-vis through induction of innate and adaptive immune response in aged mice and to compare with young mice against virulent LdWT infection.
The dendritic cell (DC) mediated activation of innate immune response plays a critical role in initiating and shaping Th1-protective responses [62]. Activation of DCs converts them into fully functional APCs capable of priming T cell responses [63]. Substantial evidence suggests that DCs preserve immune responses with aging [10, 11, 64, 65]. Notably, DCs obtained from elderly persons is reported to be able to present antigen as well as DCs from young donors [65]. Moreover, peripheral blood dendritic cells re-induce proliferation or clonal expansion in in vitro aged T cell populations thereby postponing the clonal elimination of antigen-specific T cell populations [66]. These studies suggest that at least a subset of APC in the elderly retain optimal function [67]. Previously we had shown that LdCen-/- modulates immune responses in young mice by acting on dendritic cells [46]. We therefore assessed whether BMDCs phagocytic function, LdCen-/- mediated APC function and DC-T cell interaction is impacted in aged mice in vitro. We found that LdCen-/- parasites were phagocytized at similar rate to LdWT parasites and did not persist for longer time in both young and aged BMDCs. Further, we observed both young and aged mice derived BMDCs infected with LdCen-/- parasites produced significantly more IL-12, necessary for the induction of protective Th1 response [47], compared with those infected with LdWT parasite-infected BMDCs which may likely induce heightened Th1 response in both age groups of mice. In addition, enhanced secretion of TNF and NO, directly involved in the killing of intracellular parasites [68], in LdCen-/- infected BMDCs compared with LdWT infection clearly indicated that LdCen-/- parasites not only have impaired growth inside DCs, but they also induce a proinflammatory response in both age groups, suggesting that LdCen-/- parasites are safe and immunogenic in both age groups. Additionally, OVA-pulsed BMDCs from young and aged mice infected with LdCen-/- manifest similar level of enhanced antigen presenting function as indicated by OVA specific proliferation of T cells compared to those infected with LdWT. Further such an interaction between BMDCs and T cells in vitro also resulted in enhanced CD4+Th1 cell activation as evident by higher IFN-γ and reduced IL-10 release by the responding T cells compared to LdWT infected BMDCs. Taken together, these data indicate that LdCen-/- infection in DCs substantially boost the CD4+Th1 cell effector functioning in vitro in both age groups. Additionally, the in vitro observations were further confirmed by in vivo experiments. We observed, the parasitized splenic DCs from LdCen-/- infected young and aged mice at 4 days post-infection exhibited similar and significant up regulation of proinflammatory cytokine genes such as IL-12, IL-6, IFN-γ and TNF whereas down regulation of anti-inflammatory cytokine gene IL-10 compared to LdWT infected mice. Consistent with our in vitro observation, dendritic cells sort selected from LdCen-/- parasitized young and aged mice 4 days post infection, resulted in the generation of similar level of enhanced CD4+Th1 response in young and aged mice as indicated by the higher IFN-γ/IL10 ratio compared to LdWT infected mice. Thus, both in vitro and in vivo observation suggest that LdCen-/- infection induced innate effector response in young and aged mice and interestingly, there is no diminution of LdCen-/- induced innate response in aged mice. Further, these findings are in agreement with studies done with TLR agonists showing myeloid DCs manifest preserved TLR-mediated immune responses with aging [10]. Additionally, it was also shown that TLR ligand-treated DCs can enhance the otherwise defective response of aged naive CD4 T cells via the induction of inflammatory cytokine specifically IL-6 [69]. Likewise, we found LdCen-/- substantially induces the innate effector function of aged BMDCs via production of proinflammatory molecules such as TNF, IL-6 which could likely improve residual functions of ‘‘defective” aged naive CD4+T cells thereby generating good adaptive response against virulent L. donovani challenge.
With regards to adaptive immunity, there have been conflicting reports about the tendency of senescent mice to mount either a Th1 or a Th2-like cytokine response [70]. Some studies report a bias for a Th2 response in aging mice [71], while others demonstrate a predominance of murine T cells releasing IFN-γ [72, 73]. A recent study has also shown a, unique reversal to a Th1 response along with attenuation of Th2 cell response in senescent mice during L. major infection leading to the development of resistance against cutaneous leishmaniasis [6]. Effective clearance of Leishmania parasites requires Th1 cells to secrete a substantial amount of IFN-γ [53], whereas susceptibility to Leishmania parasites is mainly mediated by Th2 cytokines, such as IL-10. Consistent with that finding, in our current study we observed LdCen-/- immunization induced significantly enhanced secretion of proinflammatory cytokine like IFN-γ over IL-10 or IL-4 in the spleen cells of immunized as well as immunized-challenged young and aged mice compared to naïve controls thereby suggesting a protective role of LdCen-/- in mediating the shift from Th2 to Th1 response in L. donovani challenged mice, along with possible activation of IFN-γ producing cells. Thus similar like our in vitro observation, LdCen-/- immunization in aged mice also induces a predominant proinflammatory microenvironment in vivo in terms of higher production of IFN-γ, IL-12 and TNF at both protein and mRNA level that may improve the residual functions of aged naive CD4+T cells. These observations are consistent with a study where introducing a combination of the inflammatory cytokines markedly enhanced the effector responses of the aged CD4 T cells in vivo [74]. Therefore, LdCen-/- immunization induced heightened proinflammatory response in aged mice further corroborates with protection in this group.
Proinflammatory (Th1) cytokine induced NO production is the main leishmanicidal mechanism of murine macrophages [55] and in this context, we also observed augmented NO generation in LdCen-/- immunized young and aged mice before and after challenge compared to naïve control. Additionally, we also detected a robust Th1-specific serum Ab (IgG2a) response in the immune-challenged young and aged mice, further highlighting LdCen-/- induced a generalized Th1 type response in both age groups and might also contribute to pathogen clearance [56, 57]. Strikingly, remarkable increase in the level of NO and IgG2a titers related to LdCen-/- immunized aged mice when compared to naïve group would indicate heightened proinflammatory response and further reinforce the protection induced by immunization with LdCen-/-. Of note, when compared to young mice, the induction of nitrite concentration and IgG2a titer level was significantly lower in aged mice from all the groups which is in agreement with other studies [15, 75] and corroborate with the diminished protection in aged mice. Indeed, it has been shown that the attenuated response to immunization in elderly individuals vaccinated against tetanus, is associated with decreased numbers of specific Ab- secreting B cells and usually decreased potency of those B cells [20]. Nevertheless, it has been reported that age-related reductions in humoral responses are mainly due to defects in the cognate helper function of naive CD4+ T cells from aged individuals [15]. This was further substantiated with hepatitis B vaccination, where it was shown that the defect in antibody production for the majority of non-responding old people results from a T-cell dysfunction [19].
T cell responses are also critical in determining the outcome of infections with Leishmania [49]. Notably, T cell proliferation in response to Leishmania antigens is an important biomarker of immunogenicity of a vaccine in mice [76]. Consistent with these reports, our results showed that vaccination with LdCen-/- parasites induces a higher T cell proliferation (both CD4 and CD8) upon stimulation with Leishmania antigen in both young and aged mice which corroborates with protection in both groups. The increase in Leishmania specific T cell proliferation in the immunocompromised aged mice after immunization with live attenuated parasite, is in agreement with earlier studies in human that found live attenuated varicella-zoster virus (VZV) vaccine enhanced frequency of VZV-specific proliferating T cells in PBMC of elderly vaccines [24]. Of particular note, we observed significant decrease in the proliferative capacity of CD8 T cell but not CD4 T cells in immunized and immunized challenged aged mice compared to young mice. Notably, since CD8 T cells contribute to the reduction in parasite burden during Leishmania donovani infection through cytotoxic activity [77] suggesting that the lower proliferation of CD8 T cells might be the underlining cause of diminished protection observed in aged mice. Indeed, several studies have shown age-related decline in overall T cell functions and its impact on vaccine efficacy and immunity [58, 78]. For example, immunization of young adults with live vaccine against influenza virus provides 65–80% protection compared to 30–50% protection in the elderly. The poor responsiveness to influenza vaccination has been shown to be specifically associated with the presence of high proportions of a population of inactive CD8+T lymphocytes that lack expression of the co-stimulatory molecule CD28 [58, 78, 79]. Additionally, in order to compare further the correlates of immune protection for LdCen-/- parasites between young and aged mice, we analyzed Ag-experienced effector memory T cells. We observed, the level of Ag-specific T cell recall response was significant, as indicated by T cell proliferation even after 8-wk post-immunization in the absence of parasite persistence, suggesting the generation of Ag-specific memory cells in both age groups and further corroborates with protection in both groups against virulent challenge. Nevertheless, compared to young mice the effector memory T cell mediated immune response is lower in aged mice. Overall, our results are in accordance with other studies showing that decline in T cell proliferation is associated with cellular senescence [11, 58] but that response is significant enough to provide protection against Leishmania infection in aged albeit lower than in young mice.
In order to determine the effector function of proliferated T cells, we further assessed the production of intracytoplasmic cytokines after immunization in young and aged mice. Apart from CD4+T cells, CD8+T cells play a critical role in the control of L. donovani infection by contributing to the formation of granulomas in the liver of L. donovani infected mice [77, 80]. Therefore, we analyzed both CD4 and CD8 T cells that are CD44Hi/CCR7Low populations and represent Ag-experienced effector memory response to determine correlates of immune protection for LdCen-/- parasites in young and aged mice. Our ex vivo analysis showed that LdCen-/- immunized young and aged mice had an increased percent of Leishmania specific CD4+ and/or CD8+ T cells expressing Th1 cytokines (IFN-γ, TNF, and IL-2) either singly or in multiple combinations. It is worth mentioning here that the development of cell mediated immune responses capable of controlling L. donovani infection and resolving disease are critically dependent upon IFN-γ and TNF [49]. Additionally, multifunctional cytokine-producing cells are found to be associated with vaccine induced protection as reported by several other studies including ours [40, 59, 81]. Of note, the percent of Th1 cytokine secreting CD4+ and/or CD8+ T cell population increases after immunizations in both young and old mice, although the percent increase in aged mice is significantly lower than in young thereby further underscoring the reduced protection in the aged mice. Indeed, in the context of T cells, it is well documented that age-related changes in T cells, specifically CD4+T cells adversely impact the outcome of a humoral response contributing to reduced vaccine efficacy [11, 82]. Naive CD4+T cells from aged animals produce about half the IL-2 as young cells and resulting in reduced differentiation of effector populations that produce reduced amounts of effector cytokines [58, 74]. Additionally, age-related decline in CD4+T cell helper activity results in the generation of defective CD8+T cell responses in terms of memory response [11]. Interestingly, we found a significantly increased IFNγ/IL-10 ratio in CD4T cells in immunized-challenged mice compared with non-immunized challenged mice, suggesting that LdCen-/- altered the Th1/Th2 balance towards a protective Th1 type response which correlated with protection in both age groups of mice. However, comparatively lower IFNγ/IL-10 ratio in immunized challenged aged mice compared to young mice further corroborate with lower protection in the aged group. It is worth mentioning here that apart from the CD4+ Th1 and Th2 cells, age-associated CD4+ Treg cell accumulation is likely to play a major role in the increased severity during L. major infection in old mice [7]. Additionally, it has also been hypothesized further that manipulation of Treg activity may enhance immune responses in the aged population and may therefore be envisioned to improve the vaccine efficiency in aged population. Future studies are needed to elucidate the specific role of Treg cell population in LdCen-/- vaccine induced immunity in aged mice.
Overall the outcome of the above described mechanism of immune responses induced by live attenuated LdCen-/- parasites resulted in significantly reduced parasite burden in the visceral organs of aged mice compared to naive challenged controls. Thus, LdCen-/- parasites are capable of inducing protective immune response in senescent mice in spite of the various deficiencies which have been reported to take place in the general immune response during aging [15]. Finally, to our knowledge this is the first report on evaluation of a genetically modified live attenuated Leishmania vaccine in aged mice. It is worth mentioning here, the main reason for using intravenous immunization in the current study with aged animals was to be consistent with our previous studies where all the end points of vaccine induced immunity were derived by intravenous route of immunization in young mice [40]. Nevertheless, in more recent studies using dogs and hamster models, we have shown LdCen-/- parasites injected either by subcutaneous (dogs) [42] or intradermal (hamsters) [41] routes also induced a strong protective immunity against L. donovani infection. We are currently performing a comparative analysis of optimal immune responses of LdCen-/- parasites inoculated through different routes of immunization in a mouse model.
In summary, our results thus indicate that LdCen-/- immunization induces substantial protection in aged host against virulent L. donovani challenge via induction of heightened innate effector function and subsequent predominant Th1 response. However, due to immunosenescence the adaptive immune responses are lower in aged mice compared to young and attempts could be made in future in improving this vaccine efficacy in aged host via administration of either higher vaccine dose, changing the vaccination route (intradermal) or addition of adjuvants such as sand fly salivary gland proteins. In that regard, a recent study from our group has shown induction of a long-lasting protective immune response in young hamsters immunized intradermally with salivary protein LJM19 with LdCen-/- against challenge from virulent parasites [41]. Whether such strategy will work in aged animals remains to be seen and is part of future studies. Taken together, our studies suggest that live attenuated LdCen-/- vaccine candidate has the potential to be used across all age groups against VL.
|
10.1371/journal.pgen.1006155 | Complex Ancestries of Lager-Brewing Hybrids Were Shaped by Standing Variation in the Wild Yeast Saccharomyces eubayanus | Lager-style beers constitute the vast majority of the beer market, and yet, the genetic origin of the yeast strains that brew them has been shrouded in mystery and controversy. Unlike ale-style beers, which are generally brewed with Saccharomyces cerevisiae, lagers are brewed at colder temperatures with allopolyploid hybrids of Saccharomyces eubayanus x S. cerevisiae. Since the discovery of S. eubayanus in 2011, additional strains have been isolated from South America, North America, Australasia, and Asia, but only interspecies hybrids have been isolated in Europe. Here, using genome sequence data, we examine the relationships of these wild S. eubayanus strains to each other and to domesticated lager strains. Our results support the existence of a relatively low-diversity (π = 0.00197) lineage of S. eubayanus whose distribution stretches across the Holarctic ecozone and includes wild isolates from Tibet, new wild isolates from North America, and the S. eubayanus parents of lager yeasts. This Holarctic lineage is closely related to a population with higher diversity (π = 0.00275) that has been found primarily in South America but includes some widely distributed isolates. A second diverse South American population (π = 0.00354) and two early-diverging Asian subspecies are more distantly related. We further show that no single wild strain from the Holarctic lineage is the sole closest relative of lager yeasts. Instead, different parts of the genome portray different phylogenetic signals and ancestry, likely due to outcrossing and incomplete lineage sorting. Indeed, standing genetic variation within this wild Holarctic lineage of S. eubayanus is responsible for genetic variation still segregating among modern lager-brewing hybrids. We conclude that the relationships among wild strains of S. eubayanus and their domesticated hybrids reflect complex biogeographical and genetic processes.
| Yeasts are key industrial microbes, most notably Saccharomyces cerevisiae, which is used to make a variety of products, including bread, wine, and ale-style beers. However, lager-style beers are brewed with interspecies hybrids of S. cerevisiae x Saccharomyces eubayanus. After its discovery in South America in 2011, rare strains of S. eubayanus have also been isolated outside of South America. Here we compare the genome sequences of several new and recent isolates of S. eubayanus from South America, North America, Australasia, and Asia to unravel the relationships of these wild isolates and their domesticated European hybrids. Two South American populations have the highest genetic diversity. One of these populations is closely related to a relatively low-diversity lineage that is spread across the Northern Hemisphere and includes the S. eubayanus parents of lager yeasts. Interestingly, we find that none of the wild isolates of S. eubayanus is the sole closest relative of lager-brewing hybrids. Instead, we show that standing variation among wild S. eubayanus strains contributed to the genetic makeup of lager yeasts. Our findings highlight the complex ancestries of lager yeasts and the importance of broader sampling of wild yeasts to illuminate our understanding of the sources of genetic variation among industrial hybrids.
| Humans changed from living in hunter-gatherer societies to agricultural societies in part through the domestication of animals and plants [1,2]. At the same time, humans began unwittingly domesticating microorganisms for the production of fermented beverages and foods, but the underlying source populations and genetic processes for microbial domestication are not well understood [3]. Beer is the most common fermented beverage in the world and can be classified as ale or lager, depending on the fermentation conditions and yeasts used. Ale-style beers are mainly produced by strains of S. cerevisiae [4]. In contrast, 94% of the beer market is dominated by lager-style beers, which are fermented at colder temperatures by allopolyploid hybrids of S. cerevisiae x S. eubayanus (syn. S. pastorianus syn. S. carlsbergensis) [5].
Two hybrid lineages of lager-brewing yeasts have been described based on genome content and phenotypic traits [6–9], leading to extensive debate about their origins. The two simplest models proposed to explain the origins of the Saaz and Frohberg lineages are through a single shared hybridization event [9–11] or through two or more independent hybridization events [6,12–15]. More complex models involving backcrossing have also been discussed by several authors [9–11,14,15]. All known modern lager strains are aneuploid. Genetic contributions from S. eubayanus have been argued to confer enhanced cold-tolerance, while genetic contributions from S. cerevisiae may confer other adaptions to the brewing environment, such as maltotriose fermentation [16–19].
Although the S. cerevisiae parent of lager yeasts seems to be closely related to modern ale strains [6,13,15], identifying close relatives of the S. eubayanus parent has proven more challenging. Since the discovery of the species in 2011 in Patagonia, South America [5], rare strains of S. eubayanus have been isolated in North America [20], Asia [21], and New Zealand [22]. Other than interspecies hybrids [5,23], no European isolates of S. eubayanus have been reported. Genome sequence comparisons have shown the Patagonian type strain to be 99.56% identical to the S. eubayanus subgenome of a lager-brewing hybrid [5], while a Tibetan isolate was shown to be 99.82% identical [21].
Previous population and phylogenetic studies of S. eubayanus suggest that it may contain up to five known phylogenetically distinct clades. Two distinct and highly diverse populations have been described in South America (Patagonia A and Patagonia B) where they have been commonly associated with Nothofagus [20], as well as Araucaria araucana [24]. Recently, an isolate from New Zealand was shown to belong to the Patagonia B clade by multi-locus phylogenetic analysis [22]. Previously isolated North American strains were shown to be the result of recent admixture between the two Patagonian populations [20]. Three lineages have been isolated in Asia, mostly in association with Quercus, including the Tibetan lineage and two early-diverging lineages that could be regarded as distinct subspecies (Sichuan and West China) [21]. Analyses of population differentiation and genetic diversity have not been performed on the latter three lineages, and all five lineages have not been thoroughly analyzed together in the same phylogenetic study.
To improve our understanding of the genetic diversity and phylogeography of S. eubayanus and its domesticated European hybrids, we have integrated existing multi-locus datasets and added several new isolates from North America (North Carolina, Washington, and New Brunswick). To extend these analyses, we have also performed whole genome sequencing (WGS) on available isolates. These results support the existence of a relatively low-diversity Holarctic lineage, which includes wild isolates from Tibet and North Carolina, as well as the hypothetical ancestor of the European interspecies hybrids. Depending on the region of the genome examined, this Holarctic lineage is embedded within or sister to one of the Patagonian populations of S. eubayanus. Genomic analyses further show that none of the known wild S. eubayanus strains is the sole closest relative to the lager-brewing hybrids. Instead, we infer that lager yeasts drew from alleles that were segregating among a Holarctic lineage of S. eubayanus.
Our ongoing high-sugar enrichment surveys of yeast from soil, leaves, bark, mushrooms, and other natural substrates in North America isolated seven new strains of S. eubayanus: one from Washington State, USA; two from North Carolina, USA; and four from New Brunswick, Canada (Fig 1A, S1 Table). The new S. eubayanus strains were isolated from novel tree hosts, including the bark of Cedrus sp., the bark and soil of Pinus taeda, and the bark of Quercus rubra. North American isolates of S. eubayanus remained quite rare overall (<1% of yeast isolates), except at specific sampling sites, and were only slightly biased toward the tree order Fagales (S1 Text, S1 Fig).
To determine how the new North American strains are related to South American [5,20], Asian [21], and New Zealand strains [22], we performed multi-locus phylogenetic analyses. Specifically, we partially sequenced nine nuclear coding sequences and three nuclear intergenic regions, consisting of a total of ~9.8 kbp, as well as one mitochondrial gene (500 bp). Existing multi-locus data was utilized at this stage, rather than WGS data, because the Chinese strains are not available for study.
North American strains displayed three different types of ancestry: 1) the strain from Washington was embedded within the Patagonia B clade and was more closely related to the strain from New Zealand than any other Patagonia B strain, 2) the strains from New Brunswick were identical at these loci to three previously characterized admixed strains from Wisconsin, USA [20], and 3) the strains from North Carolina were closely related to the strains from Tibet and lager beer (Fig 1B, S1 Text). This latter "Holarctic" subgroup of strains (Tibet, North Carolina, and Lager) was well supported phylogenetically and was more closely related to the Patagonia B clade than to any other population (Fig 1B). Phylogenetic supernetwork analysis and examination of the individual gene trees revealed a complex history for the strains in the Patagonian populations and their close Holarctic relatives, but it failed to unambiguously identify the closest relative of lager yeasts (S2 and S3 Figs, S1 Text).
To determine the consensus relationships among the wild populations of S. eubayanus and the domesticated lager-brewing hybrids, we compared the complete genome sequences of 33 strains, including representatives of both known lager yeast lineages (Saaz and Frohberg) and S. uvarum as the outgroup. In contrast to previously reported topologies citing a personal communication [25] and weak support in a multi-locus dataset [22], WGS data strongly agreed with our multi-locus phylogenetic tree and placed the Patagonia A population as an outgroup to a clade containing the Patagonia B population plus the strains from the Holarctic lineage (Fig 1C). Even with WGS data, it remained unclear whether the Holarctic subgroup was embedded within the Patagonia B population or was sister to it. In contrast, the New Zealand strain was closely related to the Washington strain, both falling within Patagonia B. These analyses further showed that, on average, the S. eubayanus subgenomes of both the Saaz and Frohberg lager yeast lineages were more closely related to the representative strain from Tibet than to known strains from North Carolina or Patagonia. Nonetheless, analysis of the full single nucleotide polymorphism (SNP) dataset revealed extensively conflicting phylogenetic signals, which are displayed by the presence of nodes subtended by multiple edges in a phylogenetic network (Fig 1D).
Concatenated phylogenies display the consensus topology supported by a dataset, which can obscure phylogenetic incongruence due to recombination, incomplete lineage sorting, and other biological processes. When genome-scale datasets are used, maximum support values can be obtained, even when different loci strongly support conflicting topologies [26,27]. To explore how recombination within and between populations has influenced the ancestry of S. eubayanus strains, we developed a simple and easily visualized test statistic and assessed its performance on one of the seven nearly identical admixed strains from North America (Fig 2D). First, across the genome, we plotted the average pairwise nucleotide sequence divergence (and standard deviation) of this strain compared to the Patagonia B and Patagonia A strains, clearly demonstrating regions more closely related to one population or the other (Fig 2A). This approach also revealed genomic regions of high genetic diversity within populations (Fig 2A) (e.g. the broader standard deviations of the left arm of chromosome IV among Patagonia A, and of the left arm of chromosome VII among Patagonia B strains). Next, for each window, we calculated the log2 of the pairwise divergence ratio using the strain with the minimum pairwise divergence value from each population. This ratio produced sharp transitions between positive and negative values, which corresponded to likely recombination breakpoints (Fig 2B). Our quantitative log2 ratio approach was generally concordant with a well-established program (PCAdmix) that uses a principal component analysis (PCA)-based method with hidden Markov model smoothing to assign ancestry to chromosomal regions according to the population contributing to it (Fig 2C). All seven admixed strains shared the same population ancestry in each chromosomal region, suggesting a recent radiation of this admixed lineage across the Great Lakes-Saint Lawrence Seaway.
Similar plots were constructed to determine whether the sequenced Tibetan strain was the closest relative of lager yeasts at all loci or whether there was indeed evidence for a more complex ancestry (Fig 3). Although most of the genomes of both the Saaz and Frohberg lager representatives were more closely related to the Tibetan genome than to the North Carolina genomes (i.e. log2 divergence ratio values < 0), 19 regions were more closely related to the North Carolina genomes in both the Saaz and Frohberg strains (i.e. log2 divergence ratio > 0.118 or 0.096 for Saaz and Frohberg, respectively, unbiased P < 0.019, permutation test) (Figs 3B, 3D and 4A). Each of these regions was supported by PCAdmix (Fig 4B), and PCAdmix detected several additional regions where the lager strains seemed to be more closely related to the North Carolina strains than to the Tibetan strain. The log2 ratio statistic and PCAdmix define windows differently, either based on physical genomic distance or the number of SNPs, respectively. Therefore, as expected, the methods did not always partition genomes in exactly the same places.
Strong support for this alternative topology was confirmed by conventional phylogenetic analyses (Fig 4C and 4D, S4 Fig). In a handful of cases, a Patagonia B representative was actually more closely related to the parent of one or both of the lager lineages than the Tibetan strain was (S5 and S6 Figs). These regions could be due to incomplete lineage sorting, introgression, or different rates of evolution among wild S. eubayanus strains, but overall, they show that lager yeasts and wild strains of S. eubayanus have complex ancestries. In particular, none of the known wild isolates of S. eubayanus is the sole closest relative to lager-brewing strains. Instead, as in the case for most natural, sexually reproducing species, the data suggest an important role for outcrossing and incomplete lineage sorting in maintaining genetic variation and creating recombinant individuals.
Surprisingly, comparison of the log2 divergence ratio values of the Saaz and Frohberg representatives against the North Carolina strains and the reference of Tibet (Fig 4, S5A and S6A Figs) highlighted at least five genomic regions where the ancestries of the Saaz and Frohberg representatives differed dramatically (Fig 4A). Several additional loci also had non-overlapping log2 ratios between Saaz and Frohberg, which provides further evidence of the complex ancestries of these lineages (Fig 4A). We closely inspected seven regions where the log2 divergence ratio, PCAdmix, or both methods suggested that the lager lineages had different alleles (Fig 4). The discordant ancestries of three of these regions were strongly supported by conventional phylogenetic analyses (Fig 4E–4G). In each case, the North Carolina strains were more closely related to one lager strain, while the Tibetan strain was more closely related to the other.
To ensure that the phylogenetic signals in these three regions were not artifacts, we closely inspected them using several orthogonal methods, including de novo assembly, PCR, local investigation of conflicting phylogenetic signals, examination of heterozygosity, and examination of copy-number variants. For example, the strongest phylogenetic signal for the region on chromosome X came from a 3-kbp region that placed the Frohberg and Tibetan strains sister to each other on a long branch (S7 Fig). Although this region contains a solo LTR in most strains, de novo assembly confirmed that the solo LTR was absent in the Tibetan and Frohberg strains and was not responsible for the phylogenetic signal. Additionally, although the Frohberg strain had multiple copies of the S. eubayanus subgenome in this region, there was no detectable heterozygosity. Heterozygosity was also too low in the other regions of phylogenetic interest to confound results (S8 Fig); indeed, overall these regions had less heterozygosity (1.08*10−4 and 8.49*10−5 heterozygous sites/bp for Saaz and Frohberg, respectively) than the genome as a whole (2.08*10−4 and 4.86*10−4 heterozygous sites/bp for Saaz and Frohberg, respectively) (S9 Fig). Differences between the regions of interest and the genome as a whole in copy-number variation (S8 and S9 Figs) and genetic diversity (S8 Fig, S2 Table) were also not the cause of the phylogenetic incongruence. Instead, we infer that the Saaz and Frohberg strains examined possess different alleles that were drawn from standing variation segregating among wild strains of S. eubayanus.
To delineate the number of populations of S. eubayanus and determine how well differentiated they are, we analyzed the multi-locus data from the complete strain set using STRUCTURE (S1 Text). Strains from West China were inferred to be an independent population and excluded from subsequent analyses. Analyses of WGS data using multiple methods suggested that Patagonia A and Patagonia B-Holarctic were independent populations and recovered the admixed strains (Fig 5). Although divisions beyond K = 2 were not significant with STRUCTURE (Fig 5A), principal component and coancestry analysis with fineSTRUCTURE provided some support for dividing Patagonia A into two subpopulations (PA-1 and PA-2, Fig 5B and 5C). Similarly, these analyses split Patagonia B-Holarctic into three subpopulations, one containing most of the non-admixed strains from Holarctic ecozone (Holarctic: North Carolina, Lager, Tibet), one containing only S. eubayanus strains from South America (PB-2), and a final subpopulation containing South-American and non-South American strains (PB-1).
These analyses also provided additional information about closest relatives of the admixed and lager strains. The fineSTRUCTURE coancestry heatmap suggested that PB-1 and PA-2 were the closest relatives of the admixed strains (Fig 5B). These results were also supported by analysis of D-statistics, where the most significant values were obtained when PB-1 and PA-2 were tested as donors to the admixed strains (S3 Table). Analysis with PCAdmix suggested that PB-1 contributed about 58% of the genome to the admixed strains, whereas PA-2 contributed 42%, results consistent with the phylogenetic analyses and an f4-ratio test (S3 Table, Fig 1D). Analysis with PCAdmix for the lager genomes further suggested that strains more closely related to the Tibetan strain contributed 66% of the S. eubayanus genetic material, whereas strains more closely related to those from North Carolina contributed 34% (S1 Text). Nonetheless, we caution that the few available data are best interpreted as pointing to the existence of standing variation across the Holarctic lineage, rather than direct ancestry or admixture involving these specific extant strains.
These results, together with the nucleotide diversity statistics (Fig 6A), the pairwise comparison of Fst, the distribution of SNPs (Fig 6B), and phylogenetic analysis (Fig 1B) support at least four distinct populations of S. eubayanus: Patagonia A, Patagonia B-Holarctic, Sichuan, and West China (Fig 6A). The nucleotide diversities of the West China population and the Holarctic lineage were lower than either population from Patagonia (Fig 6A, S4 Table). In contrast to the other populations or groups, including the Holarctic lineage as a whole, only the 10 strains from Tibet had significantly negative values for Tajima’s D, Fu and Li's D, and Fu's F (S4 Table). The Tibet group’s Fay and Wu’s H value was not significantly different from zero (H = 0.76 P > 0.05, calculated using Patagonia B strains as an outgroup), which is consistent with a neutral demographic explanation, such as a recent local population expansion across the vast region of Tibet surveyed.
The patterns of diversification and differentiation between S. eubayanus populations are remarkably reminiscent of those described recently for its sister species, S. uvarum (S10 Fig) [23]. Specifically, both species include early-diverging subspecies in East Asia or Australasia. Both species have two highly diverse, partially sympatric populations in Patagonia that are about 1% divergent in DNA sequence. In both cases, one of these populations is closely related to a relatively low-diversity lineage with a Holarctic distribution that gave rise to domesticated hybrid yeasts that ferment economically important products. In contrast to the process of introgression seen in domesticated strains of S. uvarum, lager yeasts were generated through allopolyploidization of S. eubayanus and S. cerevisiae. Genetic mechanisms of hybridization aside, the deep parallels between the diversifications of these two sister species in the wild suggest that similar biogeographical and ecological forces may explain their distributions. The presence of wild S. uvarum in Europe further suggests that Holarctic representatives of S. eubayanus are present, or may have been present in the past, somewhere in Europe.
Although non-hybrid isolates of European S. eubayanus remain elusive, we expect European strains of S. eubayanus would have relatively low genetic diversity, belong to the Holarctic lineage, and be genetically similar to isolates from Tibet and North Carolina, as well as to the parents of lager yeasts. Importantly, any European strains that might eventually be discovered will not be the closest relative to all lager yeasts at all loci because, as this study shows, standing genetic variation in S. eubayanus made it through the bottleneck of hybridization that generated modern lager yeasts. All of the currently proposed models of hybridization are compatible with this data, including multiple hybridization events [6,12–15], differential loss-of-heterozygosity among heterozygous ancestors [11], or more complicated backcrossing scenarios [9–11,14,28]. The complexity of lager yeast ancestry means that identifying the alleles relevant for specific traits may require a broad sampling of S. eubayanus genetic diversity from across the Holarctic ecozone.
In contrast to the frequent isolation of S. eubayanus from Nothofagus in Patagonia [5], the rare Northern Hemisphere strains of S. eubayanus described here and in other recent studies [20,21] were isolated in association with several different tree genera (S1 Fig). These findings suggest that our understanding of S. eubayanus ecology is still quite limited or may be an indication of its generalist character, as has recently been argued for S. cerevisiae [29]. Expanded sampling of substrates beyond the conventional hosts of Quercus and Nothofagus [30], even in South America [24], may be critical to gaining a fuller view of the ecological and genetic diversity of S. eubayanus.
Additional isolates will also be key for evaluating competing demographic models to explain the relationship between the Holarctic lineage and the Patagonia B population. One possibility is that a large ancestral population was split by vicariance, perhaps as the climate warmed following the last glacial period. Alternatively, long-range dispersal could have occurred between the Northern Hemisphere and South America, potentially in either or both directions. The relative diversities of the Holarctic and Patagonia B lineages and the confinement of a signature of recent demographic expansion to the Tibetan strains argue that dispersal from South America into the Holarctic may be more likely. Nonetheless, the distribution of clades defies a simple explanation and appears to require cladogenic events in multiple locations, both for S. eubayanus and its sister species S. uvarum.
Although humans undoubtedly played a role in selecting for the allopolyploid hybrids that became lager yeasts, human activity is not required to explain the spread of wild S. eubayanus across the Holarctic ecozone. Even conservative molecular clock estimations place all S. eubayanus cladogenic events, including the origin of the Holarctic lineage, well outside of the range of written human history (S11 Fig). Moreover, no known strain is a close enough relative to the ancestor of lager yeasts to be compatible with human-mediated transfer to Europe via the Silk Road [21] or any hypothesis involving colonial era transfer to Europe from South America [5] or North America.
How yeasts migrate is still controversial. Proposed natural mechanisms include long-range dispersal by birds [31,32], short-range dispersal by insects [33], or dispersal by wind [34]. The former may be particularly relevant because some bird migration flyways from Patagonia to Greenland or Alaska, overlap with European or Asian migration routes, respectively [35]. Clear cases for human-associated yeast dispersal have been made for industrial strains of S. cerevisiae, including the dispersal of Wine/European strains to wine-making regions all over the world [36–41], as well as some interspecies hybrids used in wine production [42]. Interestingly, Wine/European strains of S. cerevisiae have retained considerable genetic diversity, perhaps because large effective population sizes were maintained and because of the semipermeable nature of the vineyard environment [41]. European strains of S. paradoxus have also been inferred to have been dispersed to North America and New Zealand, possibly in association with Quercus [25,39,43]. A recent population genomic analysis of the former case revealed extremely low levels of diversity and a coalescence date consistent with colonial era dispersal [44].
The genomic diversity that we observed among the admixed strains of S. eubayanus from Wisconsin and New Brunswick is also consistent with a very recent dispersal to opposite ends of the Great Lakes-Saint Lawrence Seaway. The number of inferred breakpoints (40 total crossovers, Fig 2B) is similar to the number observed in one round of meiosis in S. cerevisiae [45], and each Patagonian population seems to have contributed approximately half of their genomes. Since all seven admixed strains share the same breakpoints and have nearly identical genome sequences (of 325 variable SNPs, only 37 differentiate Wisconsin from New Brunswick, Fig 2D), they are likely descended quite recently from a single individual that underwent haploselfing after an outcrossing event and one round of meiosis. Although we cannot be certain whether this dispersal across North America and the dispersal of S. paradoxus to North America were anthropic [44], they demonstrate that recent continent-scale dispersal is detectable in yeast using WGS data. In contrast, the mean genetic distance among S. eubayanus Holarctic genomes is well over 100 times higher (0.1989% for the Tibetan, North Carolina, and lager strains versus 0.0013% for the admixed strains of S. eubayanus and 0.0009% for the North American strains of S. paradoxus from Europe).
In conclusion, S. eubayanus biogeography and the origins of lager yeasts have proven more complex, but also much richer, than initially hypothesized. Here we have presented evidence that lager yeasts are derived from a relatively low-diversity lineage of S. eubayanus with a Holarctic distribution. These strains from the Holarctic lineage diversified from within one of two diverse populations found primarily in Patagonia. This pattern of diversification is similar to that of its sister species, S. uvarum. Although the S. eubayanus subgenomes of lager yeasts were drawn from the Holarctic lineage, none of the known S. eubayanus isolates is their sole nearest relative. Indeed, for the first time, we have shown that variation segregating among wild S. eubayanus persists among the allopolyploid lager-brewing yeasts. These findings strongly suggest that further sampling of the Northern Hemisphere for S. eubayanus will, not only enhance our understanding of the natural history and genetic diversity of this important species, but offer valuable insight into the sources of diversity among modern brewing strains.
New S. eubayanus strains were isolated from two locations in the USA, Washington State (yHKS509) and North Carolina (yHRVM107, yHRVM108), by following previously described high-sugar enrichment protocols at 10°C [46]. Four new S. eubayanus were isolated by enrichment from New Brunswick (yHDPN421-yHDPN424), Canada, as previously described [47], with the exception that the samples were incubated in liquid medium for seven months at 4°C, followed by a second culture step on solid medium for two weeks at 4°C. Strains were initially identified by PCR and Sanger-sequencing of the ITS region of the rDNA locus (see S1 Text). Complete results of these yeast biodiversity surveys will be reported elsewhere, and our recent publications represent less than half of the yeast strains isolated [46,47].
For the phylogenetic and nucleotide diversity analyses, we selected genes and intergenic sequences to integrate the maximum amount of sequencing data available from previous studies [20–22] (S1 Table). Additional genes from Patagonian and the newly isolated S. eubayanus strains were PCR-amplified and Sanger-sequenced (S4 Table). Reads from sequenced genes were assembled using the STADEN Package v1.7 [48]. The COX2 sequence of strain CDFM21L.1 was assembled in GENEIOUS v6.1.6 using the reads retrieved by BLASTing the S. eubayanus COX2 sequence against SRR1507225 from the SRA database of NCBI [21]. Individual genes of strain P1C1 were retrieved by BLASTing against its genome assembly (S1 Text). New sequences generated were deposited in GenBank under accession numbers KR871406-KR871626.
Phylogenetic gene trees and the supernetwork were reconstructed following our previous approach [20]. The supernetwork was reconstructed using the relative average for edge weights and using the filter option to discard the splits from PDR10 (a gene undergoing balancing selection or reciprocal introgression between some populations) (Dataset A) (S1 Text). An additional Neighbor-Net phylogenetic network was reconstructed for the SNP dataset using SplitsTree v4.12.8 [49].
Genomic libraries for available S. eubayanus strains (S1 Table), one representative strain from the Saaz lineage of lager yeast (CBS1503), and one representative strain from the Frohberg lineage of lager yeast (W34/70) were generated as described previously [50] and sequenced using Illumina paired-end sequencing (S5 Table). Details on the identification of high-quality single nucleotide polymorphisms (SNPs) can be found in S1 Text. Illumina reads were deposited in the SRA database of NCBI under accession number SRP064616.
After removing positions with gaps in any strain, whole genome nucleotide divergence graphs were constructed by calculating the pairwise number of segregating sites per nucleotide or divergence (d) in windows of 50,000 bp using the PopGenome package for R [51]. To compare how closely related various strains of interest (i.e. lager or admixed) were to a portion of the genome of two defined reference strains (e.g. North Carolina and Tibet), the value of the log2 of the ratio of the d values were calculated for each window (see S1 Text).
The whole genome phylogenetic tree was reconstructed from WGS data using RAxML v8.1 [52]. For phylonetwork and population analyses, SNPs were selected using strict coverage and quality filters (details in S1 Text). Based on the comparisons of the log2 divergence ratios or the PCAdmix results, genomic regions of interest were extracted for phylogenetic analyses (see S1 Text). Regions of interest were extracted from whole genome assemblies reconstructed using iWGSv1.01 [53].
A multi-locus concatenated alignment from Dataset A (~7.7 kbp) was generated using FASconCAT v1.0 [54]. Multi-locus concatenated alignment and WGS data were used for diversity statistics, polymorphism comparisons, and population analyses (see S1 Text). The concatenated alignment was also used to reconstruct a Maximum-Likelihood phylogenetic tree in RAxML v8.1 using the same parameters as for the individual gene trees.
A second recombinant-free concatenated alignment of the coding sequences from Dataset B (Dataset A where IntMD, MET2, and MLS1 sequences, which had low information content, were discarded) was generated using IMGC [55] and FASconCAT. The 380 fourfold degenerate sites in this alignment were used to estimate divergence times. Divergence time reconstruction was performed as we described previously [20].
The number of populations for the SNP dataset were inferred using STRUCTURE v2.3.4 [56]. fineSTRUCTURE v2 [57] was used to generate coancestry heatmaps and to perform PCA. Parental contributions to the genomes of Wisconsin, New Brunswick, Saaz, and Frohberg strains were estimated using a hidden Markov model of evolution implemented in PCAdmix v1.0 [58], and chromosomes were partitioned according to the output results. Analyses of f- and D-statistics were performed in ADMIXTOOLS v3.0 [59].
|
10.1371/journal.ppat.1003490 | CD40 Activation Rescues Antiviral CD8+ T Cells from PD-1-Mediated Exhaustion | The intrahepatic immune environment is normally biased towards tolerance. Nonetheless, effective antiviral immune responses can be induced against hepatotropic pathogens. To examine the immunological basis of this paradox we studied the ability of hepatocellularly expressed hepatitis B virus (HBV) to activate immunologically naïve HBV-specific CD8+ T cell receptor (TCR) transgenic T cells after adoptive transfer to HBV transgenic mice. Intrahepatic priming triggered vigorous in situ T cell proliferation but failed to induce interferon gamma production or cytolytic effector function. In contrast, the same T cells differentiated into cytolytic effector T cells in HBV transgenic mice if Programmed Death 1 (PD-1) expression was genetically ablated, suggesting that intrahepatic antigen presentation per se triggers negative regulatory signals that prevent the functional differentiation of naïve CD8+ T cells. Surprisingly, coadministration of an agonistic anti-CD40 antibody (αCD40) inhibited PD-1 induction and restored T cell effector function, thereby inhibiting viral gene expression and causing a necroinflammatory liver disease. Importantly, the depletion of myeloid dendritic cells (mDCs) strongly diminished the αCD40 mediated functional differentiation of HBV-specific CD8+ T cells, suggesting that activation of mDCs was responsible for the functional differentiation of HBV-specific CD8+ T cells in αCD40 treated animals. These results demonstrate that antigen-specific, PD-1-mediated CD8+ T cell exhaustion can be rescued by CD40-mediated mDC-activation.
| Hepatitis B virus (HBV) infection is responsible for more than 500,000 deaths annually as a result of the immune-mediated chronic liver damage it induces. The HBV specific CD8+ T cell response contributes to the pathogenesis of liver disease and viral clearance, and the failure to induce and/or sustain a vigorous CD8+ T cell response results in viral persistence and causes chronic necroinflammatory liver disease. To understand how the HBV-specific CD8+ T cell response is generated in response to intrahepatically expressed HBV, we generated T cell receptor transgenic mice whose CD8+ T cells are specific for HBV core or HBV envelope antigens. We find that these T cells are primed in the liver when they are adoptively transferred into HBV transgenic mouse recipients whose livers produce infectious virus particles, and that they proliferate vigorously in situ but do not differentiate into functional effector T cells after antigen recognition. Functional differentiation is suppressed by dominant negative regulatory signals, including PD-1, unless they are suppressed by anti-CD40 activation of myeloid dendritic cells.
| Rapid clonal expansion of CD8+ T cells in response to antigenic challenge is a hallmark of adaptive immunity and a crucial element of host defense. Activation and differentiation of T cells are largely determined by their initial encounter with antigen-presenting cells (APCs), and the resultant responses range from full activation and memory T cell differentiation to clonal exhaustion or deletion, depending on the nature and abundance of inductive signals that T cells decode from APCs during priming [1], [2]. These events generally occur in secondary lymphoid organs because naïve T cells are usually not primed in nonlymphoid tissues [2]. The liver is, however, an exception to this rule, due to the unique architecture of the hepatic sinusoid which is characterized by a discontinuous endothelium, the absence of a basement membrane, and a very slow flow rate [3]–[5], allowing circulating T cells to make prolonged direct contact with resident liver cells including hepatocytes [6]. Furthermore, the liver is replete with diverse and unique antigen presenting cell populations, including liver sinusoidal endothelial cells (LSECs) [7], [8], hepatic stellate cells (HSCs) [9], Kupffer cells [10], [11], conventional and plasmacytoid dendritic cells [12]–[14], all of which are capable of priming and/or tolerizing naïve T cells, at least in vitro. Thus, because of its unique immunological environment, antigens expressed and/or processed in the liver appear to be more accessible to T cells than those in other nonlymphoid organs [4], [15].
The hepatitis B virus (HBV) is a noncytopathic, enveloped, double-stranded DNA virus that causes acute and chronic hepatitis and hepatocellular carcinoma [16], [17]. Similar to other noncytopathic viruses, the clearance of HBV requires functional virus-specific CD8+ T cell responses [18]. Using the HBV transgenic mouse [19] as a model to study the impact of intrahepatic antigen recognition by HBV-specific CD8+ T cells, we have shown that adoptively transferred HBV-specific memory CD8+ T cells rapidly secrete IFNγ upon antigen recognition in the liver, thereby inhibiting HBV replication [20]. Subsequently, PD-1 is upregulated in the intrahepatic CD8+ T cells and they stop producing IFNγ, start expressing granzyme B (GrB) and undergo massive expansion [21] thereby mediating a necroinflammatory liver disease and terminating viral gene expression whereupon the intrahepatic CD8+ T cell population contracts, liver disease abates and IFNγ production returns [21].
While the foregoing studies illustrate the profound impact of intrahepatic antigen recognition on the distribution, expansion and effector functions of memory CD8+ T cells, they do not address the response of immunologically naïve CD8+ T cells to antigen recognition in the liver. Indeed, the literature reveals significant differences between naïve and memory CD8+ T cells in terms of the peptide:MHC complex concentration and costimulation required for activation and the development of their proliferative and cytokine secretion potentials, cytolytic activity and their migratory range [2], [22]. While T cell priming to viruses that do not infect conventional pAPCs is believed to occur in lymphoid organs via cross-priming [1], [2], [23], [24], the consequences of naïve T cell priming by hepatocellularly expressed viral antigen are less well understood.
In the current study, we used transgenic mice whose CD8+ T cells express T cell receptors (TCRs) specific for the HBV nucleocapsid (COR) and envelope (ENV) proteins to study the early intrahepatic immunological events that are likely to occur during HBV infection. By analyzing the response of naïve COR- and ENV-specific TCR transgenic CD8+ T cells to hepatocellularly presented HBV antigens in vivo after adoptive transfer into HBV transgenic mice whose hepatocytes produce all the HBV gene products and secrete infectious HBV virions [19], and in vitro after cocultivation with primary HBV transgenic mouse hepatocytes, we show that HBV-specific naïve CD8+ T cells are primed in the liver by HBV+ hepatocytes and proliferate vigorously in situ, but do not differentiate into functional effector T cells unless PD-1 signaling is genetically ablated. Importantly, when the same T cells are transferred into HBV transgenic mice whose myeloid dendritic cells (mDCs) were simultaneously activated by agonistic antibodies against CD40 (αCD40), PD-1 induction is suppressed and the T cells differentiate normally, inhibit HBV antigen expression, and cause liver disease. Collectively, these results indicate that CD40-mediated activation of mDCs can rescue the effector functions of PD-1-inhibited naïve CD8+ T cells, apparently by suppressing the negative regulatory signals that are triggered by antigen recognition in the liver. These results imply that the balance achieved between these two opposing forces may regulate the pathogenesis and outcome of HBV and other hepatotropic virus infections.
A Kb-restricted CD8+ CTL clone (BC10) that recognizes an epitope located between residues 93–100 in the HBV core protein (MGLKFRQL) (COR93) was generated from a Balb/c (H-2d) by C57BL/6 (H-2b) F1 hybrid (CB6F1) mouse that was immunized by standard DNA-prime/vaccinia boost immunization as previously described [21], [25]. Importantly, when in vitro core peptide-activated BC10 T cells (1×107/mouse) were adoptively transferred into HBV transgenic mice (lineage 1.3.32) that express all of the HBV antigens and replicate HBV in the liver and kidney [19], they inhibited HBV replication, and caused liver disease on day 1 after adoptive transfer (data not shown) as previously described after adoptive transfer of polyclonal COR93-specific effector memory CD8+ cells [21]. TCRα (Vα13.1JαNEW06) and β (Vβ8.1Jβ1.2) cDNA clones derived from BC10 were inserted into TCR expression cassettes [26], and injected into fertilized CByB6F2 eggs to generate BC10 TCR transgenic mice. Two founders, BC10.1 and BC10.3 carrying both TCRα and β transgenes were derived, and lineage BC10.3 was chosen for further backcrossing based on its superior allelic exclusion rate (data not shown). The BC10.3 TCR transgenic (TCRtg) mice were backcrossed more than 10 times onto C57BL/6 (B6) background, and then mated once with CD45.1 mice (H-2b) so that the TCR transgenic T cells could be easily followed by anti-CD45.1 antibody staining. As shown in Figures 1A and 1B, >98% of the splenic CD8+ T cells (33.5% of total spleen cells) in these mice were COR93-specific and CD45.1 positive as determined by staining with COR93-multimers and CD45.1 staining. As expected, they were phenotypically characterized as CD44−, CD62Lhigh, CD25−, CD69−, (Figures 1C and 1D) and fewer than 2% of them produced IFNγ or expressed Granzyme B (GrB) after 5 hours peptide stimulation in vitro (Figure 1E), indicating that they were in fact naïve T cells.
We also generated a lineage of transgenic mice whose CD8+ T cells express TCRs specific for the well-described Ld-restricted ENV28 epitope [27], [28]. The TCRs of these mice consist of Vα4.1JαNEW and Vβ1.1Jβ2.5 chains cloned from CD8+ ENV28-specific CTL clone 6C2, whose functional properties have been extensively characterized [20], [27]–[29]. Lineage 6C2.36 was chosen for further characterization and backcrossed onto the Balb/c background for at least 6 generations and then mated once with CD45.1 mice (H-2b). As shown in Figure 1F, approximately 83% of splenic CD8+ T cells (20% of total spleen cells) in lineage 6C2.36 are ENV28-specific and all of them were CD45.1 positive (Figure 1G). Again, virtually all the ENV28-specific CD8+ T cells were CD44−, CD62Lhigh, CD25−, CD69−, (Figures 1H and 1I), and they did not express IFNγ or GrB after peptide stimulation (Figure 1J), indicating that they are naïve T cells.
To examine the response of HBV-specific naïve CD8+ T cells to hepatocellularly expressed HBV, we adoptively transferred 3–5×106 COR93-specific naïve CD8+ T cells from the spleen of BC10.3 TCR transgenic donor mice into HBV transgenic lineage 1.3.32 recipient mice [19], [20]. Groups of 3–4 mice were sacrificed at various time points after adoptive transfer, and their intrahepatic, lymph nodal, and splenic lymphocytes were analyzed for the total number of COR93-specific CD8+ T cells and the extent to which they coexpress Granzyme B (GrB) and IFNγ either directly ex vivo or after in vitro stimulation by cognate COR93 peptide. To determine the functional capabilities of the adoptively transferred COR93-specific CD8+ TCR transgenic T cells during a systemic infection in vivo, we also studied their response to cognate HBcAg antigen produced by MHC-matched nontransgenic mice that had been infected 2 hours before adoptive transfer with 2×107 pfu of a recombinant vaccinia virus that expresses the HBV nucleocapsid protein (cVac) [30].
As shown in Figure 2, COR93-specific CD8+ T cells were detectable in the liver of HBV transgenic mice as early as 1 hour after adoptive transfer, and they rapidly accumulated in the liver, constituting more than 17% of total intrahepatic lymphocytes on days 1.5 and 3 (Figure 2A white bars) and showing greater than a 20-fold increase in their absolute numbers between the 1 hour and 3 day time points (Figure 2B white bars). After rapid expansion, the number of intrahepatic COR93-specific CD8+ T cells remained relatively stable up to day 10, after which they decreased more than 10-fold by day 14 (Figure 2B) but still constituted a large fraction of total intrahepatic CD8+ T cells on day 28 (Figure 2A). In contrast, the COR93-specific CD8+ T cells in the lymph nodes and spleen expanded much less vigorously between the 1 hour and 3 day time points than their counterparts in the liver, although the absolute number of COR93-specific CD8+ T cells was greater in the spleen than the liver at 1 hour and 4 hour time points (Figure 2B). COR93 specific CD8+ T cells started disappearing from the lymph nodes and spleen on day 14, and became almost undetectable on day 28 (Figures 2A and 2B; gray and black bars). In contrast, in cVac infected nontransgenic C57BL/6 mice, the frequency of COR93-specific CD8+ T cells was similar in the liver, lymph nodes, and spleen (Figure 2F), and fewer COR93-specific CD8+ T cells were detectable in the liver than the spleen (Figure 2G) at all time points tested. These results suggest that the COR93-specific CD8+ T cells were primed and accumulated preferentially in the antigen expressing liver of HBV transgenic mice rather than peripherally as in the cVac infected nontransgenic mice.
Strikingly, despite vigorous expansion (Figures 2A and 2B), the COR93-specific CD8+ T cells in the liver, lymph nodes and spleen of the HBV transgenic mice did not produce IFNγ either directly ex vivo (data not shown) or after 5 hours peptide stimulation (Figure 2C) at any time point examined, and their ability to express GrB was severely compromised as well (Figure 2D). In contrast, the intrahepatic, lymph nodal and splenic COR93-specific CD8+ T cells in the cVac infected nontransgenic recipients were able to produce IFNγ in response to 5 hours COR93-peptide stimulation, and expressed GrB directly ex vivo (Figures 2H and 2I). These data suggest that adoptively transferred HBV-specific naïve T cells preferentially expand in the HBV transgenic liver, but the expanding T cells are functionally impaired. Interestingly, virtually all the intrahepatic COR93-specific CD8+ T cells in HBV transgenic mice strongly expressed the co-inhibitory molecule PD-1 on day 1.5 ex vivo and remained so until day 28 (Figure 2E), while PD-1 expression was virtually absent in their counterparts in cVac infected nontransgenic animals (Figure 2J), suggesting that PD-1 signaling may have contributed to the functional impairment of intrahepatic COR93-specific CD8+ T cell responses in HBV transgenic mice.
To determine if the dysfunctional T cell responses in the HBV transgenic liver reflect active suppression of functional differentiation by PD-1 signaling, the COR93-specific TCR transgene was crossed for two generations onto a MHC class I matched PD-1 deficient background (kindly provided by Dr. Arlene Sharpe, Harvard Medical School) [31], yielding PD-1 deficient COR93-specific TCR transgenic animals. Equal numbers of PD-1 deficient and wild type COR93-specific naïve CD8+ T cells were adoptively transferred into HBV-transgenic mice, and analyzed for expansion, IFNγ producing ability and Granzyme B (GrB) expression on day 7 after adoptive transfer. The results were correlated with the degree of liver damage and HBV gene expression monitored by serum alanine aminotransferase (ALT) activity and HBV gene Northern Blot (NB) analysis, respectively. As shown in Figure 3A, PD-1 deficient COR93-specific CD8+ T cells expanded much more vigorously in the liver than wild type COR93-specific CD8+ T cells, and a larger fraction of PD-1 deficient T cells expressed IFNγ and Granzyme B (Figure 3B and C) and they induced a more severe liver disease, monitored as serum alanine aminotransferase (ALT) activity (Figure 3D). Furthermore, HBV gene expression was significantly reduced in the recipients of PD-1 deficient COR93-specific CD8+ T cells but not in the wild type T cell recipients (Figure 3E), reflecting the superior cytolytic and interferon gamma-producing activity of the PD-1 deficient cells. Collectively, these results indicate that PD-1 signaling suppresses the expansion and functional differentiation of HBV-specific CD8+ T cells after antigen recognition in the liver.
Because the hepatocytes in HBV transgenic mice replicate HBV at high level and release viral particles and subviral antigens into the circulation [19], HBV derived antigen could be presented to naïve T cells either by the hepatocytes themselves or by professional antigen presenting cells (pAPCs) that acquire virus particles and/or subviral antigens in the liver or in peripheral lymphoid organs. Therefore, the expansion of dysfunctional HBV-specific CD8+ T cells in the liver could reflect T cell priming and expansion in the liver, or the intrahepatic accumulation of T cells that were previously primed in the lymph nodes. To distinguish between these alternatives, we monitored the expression of activation markers (CD69 and CD25) on HBV-specific CD8+ T cells in the liver, lymph nodes and spleen at very early time points (1 hour, 4 hours and day 1) after adoptive transfer into HBV transgenic mice, and the results were compared with the expression of these activation markers on the HBV-specific CD8+ T cells in the cVac infected nontransgenic animals.
As shown in Figure 4A (white bars), within 1 hour after adoptive transfer, approximately 85.0% of the intrahepatic COR93-specific CD8+ T cells in the HBV transgenic mice expressed the very early activation marker CD69, suggesting that nearly all the COR93-specific T cells that entered the liver rapidly recognized antigen. By 4 hours, virtually all the intrahepatic COR93-specific CD8+ T cells in the transgenic mice were CD69 positive, and a large fraction of them also began to express CD25 (Figure 4B, white bars), the IL-2α receptor that is required for high affinity binding of IL-2 [32], suggesting that they were fully activated and prepared to proliferate. In contrast, CD69 expression by COR93-specific CD8+ T cells in the lymph nodes (gray bar) and spleen (black bars) occurred later (Figure 4A) than their intrahepatic counterparts (Figure 4A), and fewer nodal and splenic COR93-specific CD8+ T cells expressed CD25 (Figure 4B, gray and black bars), suggesting that naïve HBV-specific CD8+ T cell activation primarily occurred in the HBV-expressing liver and that these intrahepatically primed T cells subsequently trafficked to the lymph nodes and spleen. In contrast, COR93-specific CD8+ T cells in cVac infected nontransgenic recipients rapidly upregulated CD69 in the spleen and the liver as early as 1 hour after adoptive transfer (Figure 4C). Interestingly, CD25 expression in cVac infected nontransgenic mice was mainly observed on the splenic COR93-specific CD8+ T cells (Figure 4D), suggesting that the activation of COR93-specific CD8+ T cells during systemic vaccinia infection is largely splenic. None of these changes occurred in uninfected control nontransgenic recipients (data not shown), indicating that they were antigen specific events. Collectively, these results suggest that hepatocellularly expressed HBV antigen primes naïve T cells in the liver.
Next, groups of 4 HBV-transgenic mice received intraperitoneal injections of either saline or anti-CD62L antibodies (αCD62L), that are known to block naïve T cell homing to the lymph nodes [33]–[35], followed by naïve COR93-specific CD8+ T cells 16 hours later. The mice were sacrificed 1 hour after adoptive transfer and COR93-specific CD8+ T cells were isolated from the liver, lymph nodes, and spleen and analyzed for CD69 expression. As shown in Figure 5A, αCD62L administration completely abrogated the homing of COR93-specific CD8+ T cells to lymph nodes, but had no impact on the intrahepatic accumulation of the T cells. Despite the absence of T cell homing to the lymph nodes, COR93-specific CD8+ T cells in the αCD62L treated HBV-transgenic mice were fully activated in the liver, similar to those in saline treated recipients (Figure 5B). These results confirm that intrahepatic T cell activation and expansion do not reflect redistribution of T cells that were activated in the lymph nodes.
Intrahepatic priming of HBV-specific CD8+ T cells could reflect recognition of either endogenously synthesized hepatocellular antigen or of antigen that is released by the hepatocytes and internalized, processed and presented by liver sinusoidal endothelial cells (LSEC), Kupffer cells, or dendritic cells that are capable of cross-presentation [8], [10], [12], [36]. In an attempt to identify the antigen presenting cell population responsible for priming COR93-specific CD8+ T cells in the liver of HBV transgenic mice, we adoptively transferred COR93-specific naïve T cells into MHC-matched HBV transgenic mice lineages 1.3.32 and MUP-core 50 (MC50) that produce a nonsecretable form of HBcAg, and compared T cell accumulation and activation 1 hour later. Lineage 1.3.32 replicates HBV and expresses HBcAg (which is nonsecretable) in their hepatocytes and it also secretes viral particles and HBeAg, a soluble viral protein that is highly cross-reactive with HBcAg [16], [19]. In contrast, lineage MC50 express only HBcAg whose expression is restricted to hepatocytes [37], reducing the likelihood of antigen presentation by professional antigen presenting cells that acquire secreted viral particles or subviral antigens. As shown in Figure 6, COR93-specific CD8+ T cells accumulated similarly in liver, lymph nodes and spleen in both HBV-transgenic mouse lineages (Figure 6A), and the fraction of CD69 positive COR93-specific CD8+ T cells in the liver and lymph nodes were comparable in these lineages (Figure 6B). Since HBV core expression in MC50 transgenic mice is restricted to hepatocytes, these results suggest that naïve COR93-specific CD8+ T cells were primed by recognition of endogenously synthesized hepatocellular HBcAg. To test this notion, COR93-specific naïve T cells were co-cultured overnight with hepatocytes, LSECs, Kupffer cells, and dendritic cells that were isolated from the liver of HBV transgenic mice lineage 1.3.32 and nontransgenic controls, and then examined for CD69 expression. As shown in Figure 7, approximately 25% of COR93-specific naïve T cells upregulated CD69 when they were cocultured with hepatocytes isolated from HBV transgenic mice (Figure 7A and 7E), whereas fewer than 5% (2.5±1.8%) did so when cocultured with transgenic LSEC, (Figure 7B and 7E) and virtually no T cells expressed CD69 when cocultured with transgenic DCs or KCs (Figure 7C, 7D and 7E) despite the ability of intrahepatic LSECs, DCs and KCs to activate the T cells if the APCs are pulsed with COR93-peptide prior to coculture (Figure 7F–J). As expected, neither hepatocytes nor LSECs isolated from nontransgenic mice stimulated COR93-specific CD8+ T cells to express CD69 unless they were first pulsed with COR93-peptide (Figures 7E and 7J, white bars) indicating the antigen specificity of the T cell response to the HBV transgenic hepatocytes. Collectively, these data suggest that intrahepatic priming of COR93-specific CD8+ T cells was primarily mediated by endogenously synthesized antigen produced by the HBV-transgenic hepatocytes.
To determine if intrahepatic priming of functionally defective CD8+ T cells is a general rule or restricted to COR93-specific TCR transgenic T cells, we adoptively transferred naive ENV28-specific T cells from CD45.1-6C2.36 TCRtg mice into MHC-matched HBV transgenic mice and nontransgenic littermates. Groups of 3 mice were sacrificed 4 hours, 3 days and 7 days after adoptive transfer to examine the ENV28-specific CD8+ T cell response in the liver (Figure 8; white bars), lymph nodes (Figure 8; gray bars) and spleen (Figure 8, black bars). ENV28-specific naïve CD8+ T cells were rapidly activated in the liver of the HBV transgenic mouse recipients (but not in nontransgenic recipients – not shown) as early as 4 hours after adoptive transfer (Figure 8A), suggesting that, like COR93-specific naïve CD8+ T cells, adoptively transferred ENV28-specific CD8+ naïve T cells are primed in the liver. The intrahepatically primed ENV28-specific CD8+ T cells expanded in the liver (Figure 8B), but did not express IFNγ or GrB (Figure 8C and 8D). These results recapitulate the immunological events observed after adoptive transfer of COR93-specific naïve T cells into HBV transgenic mice illustrated in Figures 2 and 4, indicating that intrahepatic T cell priming and the expansion of functionally defective T cells occur irrespective of the antigen specificity or MHC restriction of the T cells. Thus, T cell hyporesponsiveness represents a general outcome induced by intrahepatic T cell priming to endogenously synthesized hepatocellular antigen.
Ample evidence suggests that the induction of functional CD8+ T cell responses requires the activation of professional antigen presenting cells (pAPCs), which in turn provide secondary signals to naïve T cells [1], [2], [38]. Since the results shown in Figure 7 indicate that HBV-specific naïve T cells were primed by hepatocytes that are not known to express co-stimulatory molecules [39], it is possible that the dysfunctional HBV-specific T cell responses in the HBV transgenic liver reflected the absence of a second signal.
To determine if the differentiation defect of intrahepatically primed HBV-specific CD8+ T cells can be rescued by products of the immune response to an exogenous pathogen, COR93-specific naïve T cells were adoptively transferred into HBV transgenic mice that were either treated with saline (NaCl) or infected with 2×107 of cVac 2 hours before transfer, and the results were compared with their differentiation after transfer into nontransgenic recipients that had been infected with 2×107 of cVac 2 hours before transfer. Three and seven days later, mice were sacrificed, and intrahepatic COR93-specific CD8+ T cells were analyzed for expansion, IFNγ producing ability and Granzyme B (GrB) expression. The results were correlated with the degree of liver damage and HBV gene expression monitored by serum alanine aminotransferase (ALT) activity and Northern Blot (NB) analysis, respectively. To monitor the impact of cVac infection per se on liver disease and HBV gene expression, HBV transgenic mice were infected with 2×107 of cVac without receiving COR-93-specific naïve CD8+ T cells, and they were sacrificed 3 and 7 days later. As expected, COR93-specific naïve T cells expanded vigorously in the HBV transgenic mouse liver but did not express IFNγ or Granzyme B (Figures 9A–9C; white bars). In contrast, cVac infection of HBV transgenic mice triggered IFNγ (Figure 9B; black bar) and Granzyme B (Figure 9C; black bar) expression by a small but significant fraction of the transferred intrahepatic COR93-specific CD8+ T cells without significantly increasing their expansion in the liver (Figure 9A; black bars). Note, however, that the frequency of IFNγ+ and Granzyme B+ CD8+ T cells was lower in cVac infected HBV transgenic mice (Figures 9B and 9C, black bars) than in cVac infected nontransgenic recipients (Figures 9B and 9C, blue bars), suggesting that their effector functions were suppressed by continuous hepatocellular antigen recognition, similar to the response we have shown to occur when HBV-specific memory CD8+ T cells recognize antigen in the HBV transgenic mouse liver [21].
The COR93-specific CD8+ T cells induced only a modest elevation of serum ALT activity in saline injected HBV transgenic mice (Figure 9D), and they had little or no effect on HBV gene expression (Figure 9E) in the liver. In contrast, the CD8+ T cell-mediated liver disease was more severe (Figure 9D) and intrahepatic HBV gene expression was strongly suppressed in cVac infected HBV transgenic mice compared to saline treated HBV transgenic controls (Figure 9E). As expected, cVac infection per se did not induce liver disease (Figure 9D,; red bar), nor did it suppress HBV gene expression (Figure 9E), suggesting that the induction of liver disease and the suppression of HBV gene expression in cVac infected HBV transgenic mice after COR93-specific CD8+ T cell adoptive transfer were mediated by the T cells. These results suggest that functional differentiation of HBV-specific CD8+ T cells in the HBV transgenic mouse liver is sufficiently restored in the context of a systemic virus infection to both cause hepatitis and inhibit viral gene expression.
Activation of professional antigen presenting (pAPC) cells is believed to be essential for the induction of functional CD8+ T cell responses after virus infections, and several studies have demonstrated that ligation of CD40 induces pAPC activation, resulting in the induction of CD8+ T cell responses [40]–[42]. To examine whether CD40 activation could induce functional differentiation of HBV-specific CD8+ T cells in this model, we adoptively transferred naïve COR93-specfic CD8+ T cells into HBV transgenic mice and nontransgenic controls that had been injected intravenously either with 100 µg/mouse of an agonistic anti-CD40 antibody (αCD40) [43], [44] or with saline (NaCl) 16 hours before transfer. Seven days later, the mice were sacrificed, and intrahepatic COR93-specific CD8+ T cells were analyzed for expansion, IFNγ producing ability and Granzyme B (GrB) expression. The results were correlated with the degree of liver damage and HBV gene expression monitored by serum ALT activity and Northern Blot analysis, respectively. As shown in Figure 10A, by day 7, αCD40 treatment increased the intrahepatic expansion of COR93-specific CD8+ T cells in HBV transgenic mice by 5 fold compared to the saline treated transgenic animals. Furthermore, by day 7, approximately 40% of intrahepatic COR93-specific CD8+ T cells in the αCD40 treated animals produced IFNγ in response to 5 hours in vitro peptide stimulation (Figure 10B), and almost all the COR93-specific CD8+ T cells expressed GrB directly ex vivo (Figure 10C), contrasting strikingly to their counterparts in the saline treated animals. Interestingly the induction of T cell effector functions coincided with PD-1 downregulation in intrahepatic COR93-specific CD8+ T cells (Figure 10D), suggesting that activation of CD40 signaling counteracted the PD-1 mediated negative signaling. In contrast to these observations, neither the expansion nor the functional differentiation of the COR93-specific CD8+ T cells were enhanced in αCD40 treated nontransgenic recipients (not shown). Importantly, αCD40 treated HBV transgenic recipients displayed higher serum ALT activity (Figure 10E) and very strong suppression of intrahepatic HBV gene expression (Figure 10F), after adoptive transfer of naïve COR93-specific CD8+ T cells compared to saline treated animals. The induction of severe liver disease and the suppression of HBV gene expression by αCD40 treatment reflect the vigorous expansion and functional differentiation of adoptively transferred COR93-specific CD8+ T cells, since these changes were not observed in αCD40 treated transgenic recipients that did not receive naïve COR93-specific CD8+ T cells (Figure 10E; gray bars, and Figure 10F). These results suggest that CD40 activation during intrahepatic T cell priming converts T cell hyporesponsiveness into immunity.
We then attempted to determine the role of professional antigen presenting cells (pAPCs) in αCD40 induced functional differentiation of HBV-specific CD8+ T cells. To do so, HBV-transgenic mice were crossed with CD11c.DOG mice that express the human diphtheria toxin (DTX) receptor on CD11c+ cells and thus allow depletion of dendritic cells after DTX administration with no signs of toxicity [45]. Groups of three CD11c.DOG-HBV transgenic mice were treated with DTX or saline (NaCl) every other day in combination with single administration of clodronate liposome (CLL) that is known to induce apoptosis of macrophages and DCs in vivo and in vitro [46], [47], or control liposomes (NaCl-L), yielding 4 different groups of mice (i.e. NaCl+NaCl-L, DTX+NaCl-L, CLL+NaCl, and DTX+CLL.) On day 2 after CLL or NaCl-L treatment, we analyzed the numbers of myeloid dendritic cells (mDCs; F480+CD11c+), lymphoid dendritic cells (lymDCs: F480−CD11c+), Kupffer cells (F480+CD11c−) and B cells (B220+) in the liver to determine the efficacy of pAPCs depletion. As shown in Figures 11A and 11B, DTX and CLL independently depleted mDCs in the liver and their effects were additive. (Figure 11A), while intrahepatic lymDCs were depleted only by DTX treatment (Figure 11B). Surprisingly, the number of Kupffer cells paradoxically increased when mice were treated with DTX or CLL alone and with both together (Figure 11C). This might reflect that dendritic cell death stimulated proliferation and/or migration of Kupffer cells. None of these treatments significantly reduced the number of intrahepatic B cells (Figure 11D). To examine the impact of pAPC-depletion on αCD40 induced functional differentiation of HBV-specific CD8+ T cells, CD11c.DOG-HBV transgenic mice that were pre-treated with DTX, CLL or DTX plus CLL, were injected with αCD40, and 1 day later, adoptively transferred with COR93-specific naïve T cells. The mice were sacrificed on day 7 after adoptive transfer, and the intrahepatic COR93-specific CD8+ T cells were analyzed for expansion, IFNγ producing ability and Granzyme B (GrB) expression. The T cell responses were correlated with the degree of liver damage monitored by serum ALT activity. As shown in Figures 11E to 11G, expansion, IFNγ producing ability, and GrB expression of COR93-specific CD8+ T cells in αCD40 treated CD11c.DOG transgenic mice were directly correlated with the number of intrahepatic mDCs at the time of αCD40 administration, but not those of intrahepatic lymDCs, Kupffer cells, or B cells, suggesting that mDCs are required for αCD40 induced functional differentiation of HBV-specific CD8+ T cells. Taken together, these results suggest that activation of mDCs through the CD40 pathway can overcome PD-1-mediated suppression and induce functional CD8+ T cell responses in response to intrahepatically expressed HBV.
The current study examines the impact of hepatocellular antigen presentation on the expansion and functional differentiation of antigen-specific CD8+ T cells. Our results revealed that intrahepatic antigen presentation primes functionally defective antigen-specific CD8+ T cell responses, that the dysfunctional CD8+ T cell response reflects active suppression of expansion and functional differentiation by PD-1 signaling, and that such suppression can be overridden by activating myeloid dendritic cells (mDCs) through CD40 stimulation.
COR93-specific naïve CD8+ T cells were rapidly activated in the liver after adoptive transfer into HBV transgenic mice as indicated by more rapid expression of activation markers CD69 and CD25 by intrahepatic COR93-specific CD8+ T cells than their lymph nodal and splenic counterparts (Figures 4A and 4B). Hepatocellular activation of the COR93-specific CD8+ T cells was not due to hepatic migration of cells that had been activated in lymphoid organs, because the naïve T cells were equally activated when T cell homing to lymph nodes was prevented by anti-CD62L antibody (αCD62L) treatment (Figure 5) and when HBcAg secretion was precluded (Figure 6). Rather, it reflected T cell priming in the liver by recognition of endogenously synthesized hepatocellular antigen (Figure 7). As a consequence of hepatocellular antigen presentation, the COR93-specific CD8+ T cells upregulated CD69 and CD25 expression (Figure 4A and B) and expanded vigorously (Figures 2A and 2B) but they were functionally impaired as they did not express IFNγ or Granzyme B (GrB) either directly ex vivo or after 5 hours in vitro peptide stimulation (Figures 2C and 2D). Importantly, hepatocellular T cell priming and the expansion of functionally defective T cells also occurred when ENV28-specific naïve T cells were transferred into HBV-transgenic mice (Figure 8), indicating that hepatocellular T cell priming induces functionally defective T cells responses irrespective of antigen specificity and MHC restriction and illustrating the generality of these observations. The dysfunctional HBV-specific T cell responses likely reflected the suppression of functional differentiation by PD-1 signaling (Figure 3), but such suppression could be overcome by vaccinia virus infection (Figure 9) or simultaneous activation of mDCs via the CD40 signaling pathway (Figures 10 and 11).
While various hepatic cell populations have been shown to contribute to T cell priming in the liver [8], [10], [12], [36], [39], our data suggest that HBV antigen-positive hepatocytes are responsible for priming of HBV-specific CD8+ T cells in our system. The fenestrated liver sinusoidal endothelium permits circulating T cells to make direct contact with underlying hepatocytes [6]. Indeed, when hepatitis B envelope (ENV) specific effector CD8+ T cells were adoptively transferred into hepatitis B virus (HBV) transgenic mice that express the ENV protein in their hepatocytes, renal tubular epithelium, and choroid plexus cells [48], the ENV-specific effector CD8+ T cells were selectively sequestered and specifically activated in the liver where they caused a necroinflammatory disease [4], [29] but not in the other tissues. Importantly, whereas they failed to recognize antigen in the kidney or the CNS when injected intravenously, the CTLs were highly cytopathic for ENV-positive renal tubules and choroid plexus epithelial cells when they were injected directly into those tissues [4]. Furthermore, a series of studies by Bertolino and colleagues suggest that hepatocytes can prime alloantigen specific naïve T cells [15], [35], [39], suggesting that antigen expressed by hepatocytes is highly accessible to circulating native T cells. In those studies, however, the intrahepatic priming of alloantigen specific T cells by hepatocytes was shown to induce rapid T cell deletion [15], [35], [39], contrasting strikingly to the vigorous expansion of HBV-specific CD8+ T cell responses described in this study. The basis for the difference is unclear, but it could reflect the different TCR affinity of transgenic T cells and the level of cognate antigen expression in the liver. Nonetheless, in Bertolino's hands and ours, intrahepatic antigen recognition fails to trigger CD8+ T cell functional differentiation.
Strikingly, the intrahepatically activated HBV-specific CD8+ T cells in the HBV transgenic mice did not secrete IFNγ or display cytotoxic activity (Figures 2C and 2D). Consequently, they did not inhibit HBV replication (data not shown) or gene expression (Figure 3E) or cause a necroinflammatory liver disease (Figure 3D). The lack of effector functions was not due to intrinsic defects of HBV-specific TCR transgenic CD8+ T cells or the large number of adoptively transferred T cells, since the same transgenic CD8+ T cells differentiated into fully functional effector T cells in nontransgenic recipients that were infected with recombinant vaccinia viruses expressing the HBV core antigen (Figures 2F–2I). Instead, our data suggest that the dysfunctional intrahepatic CD8+ T cell response reflects the impact of antigen-induced, PD-1-mediated negative signaling. The intrahepatic HBV-specific CD8+ T cells upregulated PD-1 (Figure 2E) in HBV transgenic mice but not in nontransgenic mice infected with cVac (Figure 2J). Furthermore, PD-1 deficient COR93-specific naïve T cells expanded more vigorously than their PD-1 positive wild-type counterparts, and they differentiated into cytotoxic effector T cells in situ, caused severe liver damage, and inhibited HBV gene expression in the liver (Figure 3). These results suggest that HBV-specific CD8+ T cells can be primed in the liver by recognition of antigen expressed by hepatocytes but activation of their effector functions is suppressed by PD-1 signaling, consistent with the previous studies reported by us [21], [49] and others [50], [51]. Whether other negative signaling molecules such as CTLA-4 [52], Tim3 [53]–[55], 2B4 [56], [57], IL-10 [58], [59] and TGFβ [60], [61] that are known to suppress antiviral CD8+ T cell responses also contributed to the suppression of functional differentiation after intrahepatic priming remains to be determined. In chronic HBV patients, the inhibitory molecule 2B4 and Tim-3 are highly co-expressed with PD-1 on HBV-specific CD8+ T cells [62], [63]. Similarly, CTLA-4 is highly expressed on HBV-specific CD8+ T cells that express high levels of pro-apoptotic molecule Bim [64]. Furthermore, in vitro blockade of CTLA-4 or Tim-3 signaling appears to restore effector functions of HBV-specific CD8+ T cells after in vitro peptide stimulation, and this effect was even enhanced when combined with PD-1 blockade, suggesting that HBV-specific CD8+ T cell responses in chronically infected patients are suppressed by several non-redundant mechanisms [62], [64]. Importantly, our data suggest that such suppressive mechanism(s) can be overcome by myeloid dendritic cell (mDC) activation. HBV-specific CD8+ T cells differentiated into fully functional effector T cells in recipient HBV transgenic mice that were treated with agonistic anti-CD40 antibodies (αCD40), resulting in liver disease and the inhibition of HBV gene expression (Figure 10). Importantly, mDCs are required for αCD40 induced functional differentiation of HBV-specific CD8+ T cells (Figure 11). Collectively, these results suggest that activation of mDCs via CD40 signaling was essential to rescue HBV-specific CD8+ T cells from functional suppression through PD-1 and perhaps other regulatory molecules.
Several studies with αCD40 established a model postulating that the αCD40 activates pAPCs that then provide secondary signals to naïve CD8+ T cells upon antigen presentation [38], [40], [42]. According to this model, HBV-specific CD8+ T cells were programmed to differentiate into functional effector T cells during cross-priming by αCD40-activated pAPCs that acquired circulating HBV particles, subviral antigens or HBV expressing hepatocytes or hepatocyte fragments. However, our preliminary data suggest that hepatic DCs isolated from αCD40 treated HBV-transgenic mice cannot stimulate COR93-specific native T cells to express CD69 (data not shown), suggesting inefficient cross-priming by αCD40-activated pAPCs. Therefore, it is possible that αCD40-activated pAPCs released cytokines such as IL-12 and type I interferons that provide a third signal required for T cell functional differentiation [1], [65], [66]. In line with this notion, Maini and colleagues have recently showed that IL-12 potently augments the capacity of HBV-specific CD8+ T cells to produce IFNγ upon in vitro stimulation by cognate antigen in association with down-modulation of PD-1 [67]. Additional studies are required to test these hypotheses.
It remains to be determined if similar events occur during natural HBV infection. Ample evidence suggests that HBV-specific CD8+ T cell responses are functionally impaired during chronic HBV infections [51], [68], [69] and the functional impairment is associated with PD-1 expression [68], [70] by HBV-specific CD8+ T cells, similar to the transgenic T cells described in this study. Therefore, the expansion of functionally defective CD8+ T cells by hepatocellular priming may explain the weak CD8+ T cell responses observed during chronic HBV infections. In contrast, approximately 95% of adult onset acute HBV infection is characterized by a vigorous HBV-specific CD8+ T cell response. While our data indicate that activation of mDCs through the CD40 pathway can induce functional HBV-specific CD8+ T cell responses, it is currently unknown whether, and if so, how the mDCs are activated during natural HBV infection. While CD40 ligand (CD40L) is expressed on a variety of cells including platelets, mast cells, basophils, NK cells, and B cells, antigen-specific CD4+ T cells are primarily responsible for activating CD40 expressing cells, particularly professional antigen presenting cells (pAPCs), and CD4 T cell mediated CD40 activation appears essential for cross-priming functional CD8+ T cell responses [40]–[42]. Indeed, early priming of HBV-specific CD4+ T cells before or during viral spread in HBV-infected chimpanzees appears to be necessary to initiate a functionally efficient CD8+ T cell response, and the depletion of CD4+ T cells before HBV infection precluded functional T cell priming and caused persistent infection in experimentally infected chimpanzees [70]. Experiments are currently underway to determine whether fully functional CD8+ T cell responses can be induced in this transgenic mouse model by providing HBV-specific T cell help.
In summary, the data described herein demonstrate that endogenously synthesized hepatocellular antigen primes functionally defective HBV-specific CD8+ T cells via an instructional process involving PD-1 signaling that actively suppresses expansion and functional differentiation of hepatocellularly primed T cells. Importantly, such suppressive mechanisms can be overcome by activating mDCs through the CD40 pathway. Collectively, these results suggest that HBV specific CD8+ T cell responses are regulated by the balance between PD-1 mediated inhibitory signaling and stimulatory signals by activated DCs. More experiments are required to determine whether DC activation and/or PD-1 blockade may, individually or together, have therapeutic potential to terminate chronic viral infections of the liver and possibly other persistent viral infections as well.
All experiments involving mice were performed in the AAALAC accredited vivarium (Vertebrate Animal Assurance No. A3194-01) at The Scripps Research Institute. All animal studies follow the guidelines in the NIH Guide for the Care and Use of Laboratory Animals and are approved by The Scripps Research Institute Animal Care and Use Committee (Protocol # 08-0159).
HBV transgenic mouse lineage 1.3.32 (inbred C57BL/6, H-2b) and lineage MC50 have been previously described [19], [37]. Lineage 1.3.32 expresses all of the HBV antigens and replicates HBV in the liver and kidney at high levels without any evidence of cytopathology. Lineage MUP-core 50 (MC50) (inbred C57BL/6, H-2b) expresses the HBV core protein in hepatocyte under the transcriptional control of the mouse major urinary protein (MUP) promoter. In all experiments, the mice were matched for age (8 weeks), sex (male), and (for the 1.3.32 animals) serum HBeAg levels in their serum before experimental manipulations. PD-1 deficient mice and CD11c.DOG mice (both inbred C57BL/6, H-2b), kindly provided by Drs. Arlene Sharpe and Günter Hämmerling, respectively, have been previously described [31], [45]. All experiments were approved by The Scripps Research Institute Animal Care and Use Committee.
Synthetic peptides corresponding to the previously described [27], [28], [71] HBV envelope (ENV)-specific CTL epitope, ENV28 (Ld; IPQSLDSWWTSL) and HBV nucleocapsid protein (COR)-specific CTL epitope, COR93 (Kb; MGLKFRQL) were purchased from Mimotope (Victoria, Australia). Recombinant vaccinia viruses that express the nucleocapsid protein (core) (subtype ayw) of HBV (designated cVac) and the major envelope protein (S) (subtype adw2) (HBs4) were kindly provided by H.J. Schlicht [30] and B. Moss [72], respectively.
A CD8+ CTL clone termed BC10, that is Kb-restricted and specific for an epitope located between residues 93–100 in the HBV core protein (MGLKFRQL) (COR93), was generated in Balb/c (H-2d) by C57BL/6 (H-2b) F1 hybrids (CB6F1) that were immunized by standard DNA-prime/vaccinia boost immunization to induce an HBcAg specific CD8+ T cells response as previously described [21], [25]. Fourteen days after the booster immunization, mice were sacrificed and spleen cells were collected. 4×106 spleen cells were cocultured with 1×105 of irradiated RBL5 cell transfectants that express the HBV core protein (RBL5/c) in complete RPMI 1640 medium (GIBCO, Frederick, Md.) containing streptomycin (100 µg/ml), penicillin (100 U/ml), 2-mercaptoethanol (5×10−5 M), 10% fetal calf serum, and 2.5% EL-4 supernatant in 24-well plates (Costar, Cambridge, Mass.). The RBL5/c cell line was a gift from Dr. Jorg Reiman [71]. Seven days later, the spleen cells were restimulated with RBL5/c, and on day 14, they were cloned in 96-well round bottom plates (Costar) at 1 cell/well. After 2 to 3 weeks of repetitive stimulation, wells containing growing cells were expanded and four H-2b restricted CTL clones were established. A CTL clone, termed BC10, was chosen for further characterization based on its superior cytolytic activity against RBL5/c. The fine antigen specificity of BC10 was confirmed by intracellular IFNγ production in response to a dominant, Kb restricted COR93-CTL epitope [71].
TCR-α and -β cDNAs were synthesized from 20 ng of total messenger RNA extracted from COR93-specific CD8+ T cell clone BC10 and ENV28-specific CD8+ T cell clone 6C2 [27], [29], and amplified by PCR as described by Yoshida, et al [73]. The PCR products were cloned into the pGEM-T Easy Vector (Promega. Madison, WI) and then sequenced. The sequence analyses revealed that CTL clone BC10 expressed a TCR composed of Vα13.1JαNEW06 and Vβ8.1Jβ1.2, while the CTL clone 6C2 expressed a TCR composed of Vα4.1JαNEW and Vβ1.1Jβ2.5 chains. Flanking primers were designed to amplify the rearranged Vα13.1JαNEW06 and Vβ8.1Jβ1.2 genomic DNA based on genomic sequence from these TCR loci. PCR products were sequenced again and then inserted into the TCR expression cassettes pTα and pTβ [26], kindly provided by Dr. Diane Mathis. Prokaryotic DNA sequences were removed from both vectors and injected into fertilized CByB6F2 eggs as previously described [19]. Founders were screened by PCR and analyzed for the specific TCR expression on CD8+ T cells in the peripheral blood by FACS analysis. A founder, BC10.3, expressed TCR specific for COR93 epitope and was bred against C57BL/6 mice (H-2b) for 6 generation and then against CD45.1 mice (C57BL/6 background; H-2b) for at least 6 more generations. A founder, 6C2.36, expressed TCRs specific for ENV28 epitope and was bred against Balb/c mice (H-2d) for more than 6 generations before being mated with CD45.1 mice to produce H-2bxd F1 hybrids.
Spleen cells were isolated from TCR transgenic mice BC10.3 (CD45.1+;H-2b) or 6C2.36xCD45.1 F1 hybrids (CD45.1+; H-2bxd) as previously described. Spleen cells from BC10.3 were transferred into either lineage 1.3.32 mice that were homozygous for HBV, or into MC50 mice heterozygous for the HBV core antigen, or into nontransgenic C57BL/6 mice (H-2b). In selected experiments, the mice were either intravenously infected with 2×107 of recombinant vaccinia viruses expressing the HBV core antigen (cVac) or received saline as a control. Spleen cells from 6C2.36xCD45.1 F1 hybrids were transferred into HBV transgenic mice lineage 1.3.32× Balb/c F1 hybrids (H-2bxd) and syngeneic nontransgenic recipients. Groups of 3–4 mice were sacrificed at various time points after adoptive transfer and their livers, lymph nodes, and spleen were harvested for further analysis.
The FGK45 hybridoma producing rat IgG2a mAb against mouse CD40 (αCD40) was provided by Dr. A. Rolink (Basel Institute for Immunology, Basel, Switzerland) [43]. CD40 was purified from FGK45 culture supernatants as previously described [44]. Mice were intravenously injected with 100 µg of CD40 16 hours before adoptive transfer of HBV-specific naïve T cells. The mice were sacrificed at different time points after injection, and their livers were harvested for further analysis (see below). A monoclonal anti-CD62L antibody (clone Mel-14) was purchased from BD Bioscience, and mice were intraperitoneally administered with 100 µg of anti-CD62L mAb (αCD62L; clone Mel14) in 200 µl of PBS at 16, and 4 hours before adoptive transfer. If necessary, further doses of antibodies were administered on day 2 after adoptive transfer.
Diphtheria Toxin (DTX) was purchased from Sigma-Aldrich, dissolved in PBS, and intraperitoneally administered (200 ng/mouse) every other day. Clodronate Liposome and control Liposome were both purchased from Encapusula NanoSciences, and intravenously injected once (200 µl/mouse).
Spleen cells, lymph node cells, and intrahepatic lymphocytes (IHL) were prepared as previously described [21], [74]. Briefly, spleen cells and lymph node cells were isolated by pressing through a 70 µm cell strainer (Becton Dickinson) with the plunger of a 1-ml syringe and were washed three times with PBS and used for further analysis. For IHL isolation, livers were perfused with 10 ml of PBS via the portal vein to remove circulating lymphocytes and the liver cell suspension was pressed through a 70 µm cell strainer and digested with 10 ml of RPMI 1640 medium (Life Technologies), containing 0.02% (w/v) collagenase IV (Sigma) and 0.002% (w/v) DNase I (Sigma), for 40 minutes at 37°C. Cells were washed with RPMI 1640 and then overlaid on Percoll/Histopaque solution consisting of 12% Percoll (Pharmacia) and 88% Histopaque-1083 (Sigma-Aldrich). After centrifugation for 20 min at 1500× g, the IHLs were isolated at the interface. The lymphmononuclear cells were washed twice with RPMI 1640 medium and used for further analysis
The livers of HBV transgenic mice lineage 1.3.32 were perfused slowly via the inferior vena cava with 25 ml of warm Liver Perfusion Medium (Gibco-Life Technologies) at a rate of 5 ml/minute, and then digested with 50–75 ml of warm Liver Digest Medium (Gibco-Life Technologies) at a rate of 5 ml/min. Following complete digestion of the liver (10–15 min.), the gall bladder was removed and the liver carefully excised. Cells were collected from the liver by disrupting the liver capsule and swirling the tissue in a petri dish containing Liver Digest Medium. Liver nonparenchymal cells (LNPCs) containing LSECs were separated from hepatocytes by centrifuging the cell suspension at 100× g for 2 min at room temperature. For LSEC isolation, the supernatant containing LNPCs was washed twice with RPMI 1640 (Cellgro), and the cell pellet was resupended with BD IMag Buffer at a concentration of 1×107/ml. The cell suspension was incubated with biotinylated antibody specific for lymphatic vessel endotherial hyaluronan receptor 1 (LYVE-1) antibody (eBioscience) (20 µl for every 1×107 of LNPCs) for 15 min on ice, washed twice, and resuspended in IMag buffer at a concentration of 2×107/ml. The cell suspension was then incubated with BD IMag Streptavidin Particles Plus-DM (BD Bioscience) (50 µl for every 1×107 of LNPCs) for 30 min on ice, and washed twice, and resuspended in IMag buffer at a concentration of 2 to 8×107. LYVE-1+ cells were then positively selected using BD IMagnet (BD Bioscience) following the manufactures instruction, and then stained with PE-conjugated CD147, APC-conjugated CD31, and Alexa 700 conjugated CD45. The purity and viability of LSEC (CD147+CD31+CD45−) cells [75], [76] was routinely greater than 80% and 90%, respectively. For hepatocyte isolation, the pellet from the initial centrifugation step were washed in DMEM containing 10% FCS and centrifuged at 100× g for 2 min at room temperature. This washing step was repeated until the supernatant was no longer cloudy. Hepatocyte viability was routinely higher than 80%. For intrahepatic Dendritic cell (DC) and Kupffer cell (KC) isolation, intrahepatic lymphocytes (IHLs) were isolated as described in the previous section, and DCs and Kupffer cells were positively selected using biotinylated CD11c and CD11b and streptavidin Magnetic Particles (BD Biosciences), following the manufactures instruction. The purity of DCs and KCs were then analyzed by staining with PE-conjugated CD11b, APC-conjugated CD11c, PE-Cy7-conjugated F480, and FITC-conjugated Ly6G. The purities of DCs (CD11c+CD11b+Ly6G−) and KCs (CD11b+F480+CD11c−Ly6G−) were routinely greater than 50%, and their viabilities higher than 90%. All antibodies were purchased from BD Bioscience and eBioscience.
Lymphmononuclear cells isolated from the liver, spleen, peripheral blood and lymph nodes were incubated with a mixture containing the COR93-dimer or ENV28-dimer, APC-, or Pacific Blue-conjugated anti-mouse CD8+, PE-Cy7 conjugated anti-mouse CD69, APC-conjugated anti-mouse CD25 or CD62L, FITC-conjugated anti-mouse CD45.1, and PE-conjugated PD-1 or CTLA-4 (BD Bioscience) for 1 hour on ice. After washing, the cells were incubated for 30 minutes with APC-conjugated anti-mouse IgG at 4°C to detect dimer positive cells. Dimer without peptide was used as a control. Intracellular cytokine staining (ICS) was performed using PE- or APC-conjugated anti-mouse IFNγ, APC-conjugated anti-mouse Granzyme B (GrB) (Caltag) after incubating for 5 hours at 37°C in the presence of brefeldin A (BFA), as previously described [21], [25], [49]. All antibodies were purchased from BD Bioscience and eBioscience.
Total liver DNA and RNA were analyzed for HBV replicative intermediates by Southern blot, and for HBV RNA by Northern blot, exactly as previously described [19], [20]. The relative abundance of specific DNA and RNA molecules was determined by phosphor imaging analysis, using the Optiquant image analysis software (Packard).
The extent of hepatocellular injury was monitored by measuring sALT activity at multiple time points after treatment as previously described [20].
Student t test was performed using Microsoft Excel. Data are depicted as the mean ± SD, and P values<0.05 were considered significant: *P<0.05, **P<0.01, ***P<0.001.
|
10.1371/journal.pmed.1002811 | Retention and viral suppression in a cohort of HIV patients on antiretroviral therapy in Zambia: Regionally representative estimates using a multistage-sampling-based approach | Although the success of HIV treatment programs depends on retention and viral suppression, routine program monitoring of these outcomes may be incomplete. We used data from the national electronic medical record (EMR) system in Zambia to enumerate a large and regionally representative cohort of patients on treatment. We traced a random sample with unknown outcomes (lost to follow-up) to document true care status and HIV RNA levels.
On 31 July 2015, we selected facilities from 4 provinces in 12 joint strata defined by facility type and province with probability proportional to size. In each facility, we enumerated adults with at least 1 clinical encounter after treatment initiation in the previous 24 months. From this cohort, we identified lost-to-follow-up patients (defined as 90 or more days late for their last appointment), selected a random sample, and intensively reviewed their records and traced them via phone calls and in-person visits in the community. In 1 of 4 provinces, we also collected dried blood spots (DBSs) for plasma HIV RNA testing. We used inverse probability weights to incorporate sampling outcomes into Aalen–Johansen and Cox proportional hazards regression to estimate retention and viremia. We used a bias analysis approach to correct for the known inaccuracy of plasma HIV RNA levels obtained from DBSs. From a total of 64 facilities with 165,464 adults on ART, we selected 32 facilities with 104,966 patients, of whom 17,602 (17%) were lost to follow-up: Those lost to follow-up had median age 36 years, 60% were female (N = 11,241), they had median enrollment CD4 count of 220 cells/μl, and 38% had WHO stage 1 clinical disease (N = 10,690). We traced 2,892 (16%) and found updated outcomes for 2,163 (75%): 412 (19%) had died, 836 (39%) were alive and in care at their original clinic, 457 (21%) had transferred to a new clinic, 255 (12%) were alive and out of care, and 203 (9%) were alive but we were unable to determine care status. Estimates using data from the EMR only suggested that 42.7% (95% CI 38.0%–47.1%) of new ART starters and 72.3% (95% CI 71.8%–73.0%) of all ART users were retained at 2 years. After incorporating updated data through tracing, we found that 77.3% (95% CI 70.5%–84.0%) of new initiates and 91.2% (95% CI 90.5%–91.8%) of all ART users were retained (at original clinic or transferred), indicating that routine program data underestimated retention in care markedly. In Lusaka Province, HIV RNA levels greater than or equal to 1,000 copies/ml were present in 18.1% (95% CI 14.0%–22.3%) of patients in care, 71.3% (95% CI 58.2%–84.4%) of lost patients, and 24.7% (95% CI 21.0%–29.3%). The main study limitations were imperfect response rates and the use of self-reported care status.
In this region of Zambia, routine program data underestimated retention, and the point prevalence of unsuppressed HIV RNA was high when lost patients were accounted for. Viremia was prevalent among patients who unofficially transferred: Sustained engagement remains a challenge among HIV patients in Zambia, and targeted sampling is an effective strategy to identify such gaps in the care cascade and monitor programmatic progress.
| Retention and HIV RNA suppression in HIV treatment programs represent critical metrics of success, but regionally representative estimates in longitudinal cohorts remain uncommon.
Most treatment programs, whether at the national or sub-national level, lack data systems able to capture patient movement across facilities, which may lead to underestimates of retention.
HIV RNA suppression levels from routine program monitoring or large-scale cross-sectional studies may miss patients who are lost to follow-up, and therefore those who had been on treatment, thus underestimating the prevalence of viremia.
Intensive ascertainment of care status and HIV RNA levels in a numerically small but randomly selected sample of patients with unknown outcomes can improve our understanding of treatment success in real-world program settings.
We used a multistage sampling approach, in which we first selected facilities and then, within each facility, selected a random sample of patients who were lost to follow-up—defined as no contact with a health facility for 90 or more days after last their missed appointment—for intensive tracing. We also collected dried blood spots in Lusaka Province to determine viral load levels in both a sample of lost patients and in-care patients to estimate the prevalence of viremia in both populations.
We found that among 165,464 patients on treatment in 64 facilities, 28,111 (17%) were lost to follow-up. We traced 2,892 of the lost (16%): and found 412 (14%) had died and 1,751 (61%) remained alive. Of those alive, 1,293 (74%) continued to receive treatment, 255 (15%) had stopped, and care status could not be determined in 12%.
Among all ART patients, using data known to the program before tracing, retention was 67.7% at 2 years; after incorporating findings among the lost patients, retention was 91.2%.
Among 1,044 participants with a viral load determination (901 in care and 143 who were lost), viremia was present in 18.1% of those in care, 71.3% of those lost to follow-up (49.8% of those lost and in care elsewhere and 83.9% of those lost and not in care), and 24.8% overall.
We found that patient retention in public ART facilities in Zambia was higher than apparent in data collected during routine care and monitoring.
Estimates of viremia that do not account for elevated levels in patients who stop treatment (and are lost to follow-up from cohort studies) or are missing from cross-sectional studies may overestimate treatment success.
Viremia among patients lost from one facility who reported engagement in a new facility was markedly higher than among patients who remained engaged in their original facility: Even though durable discontinuation from care was relatively infrequent, strategies to consistently engage patients to enhance retention and viral suppression are urgently needed.
| Assessments of retention and HIV RNA suppression levels after HIV treatment initiation in routine program settings represent the backbone of data-driven public health efforts to bring the epidemic under control. As HIV treatment reaches more patients in the era of test-and-treat, the remaining gaps in the cascade are likely to shift toward retention and adherence as the key modifiable mediators of success [1–3], which warrant careful assessment. Identifying when and why patients miss clinical visits, fail to pick up medications, and become viremic can help programs focus attention on vulnerable periods [4]. In addition, identifying facilities where retention and viral suppression are lower than at other similar settings can also direct targeting of additional health systems investments. Indeed, at this phase of the HIV treatment response, relatively widespread geographical access to treatment and large numbers on treatment already mean that the next phase in improvement efforts should focus on retention and suppression.
The importance of accurate measures of retention and viral suppression in routine care delivery settings, however, brings critical challenges in monitoring into focus. First, many patients move for social or livelihood reasons. Most programs lack data systems that are integrated in a region to capture movement across facilities. Second, even networked systems will not link records when patients enroll in new facilities using different names or identifiers, which is common to avoid being considered uncommitted patients by healthcare workers. In either case, routinely available data may underestimate retention [5,6]. Similarly, routine clinic-based viral load monitoring will fail to account for patients who are not coming back to clinic (i.e., lost to follow-up). In an analysis from the International Epidemiology Databases to Evaluate AIDS (IeDEA), investigators found that 94% of retained patients were virally suppressed, but this figure dropped to 45% when all lost patients were assumed to be viremic [7]. The Zambia Population-based HIV Impact Assessment (ZAMPHIA) suggested viral suppression in nearly 90% of people self-reporting ART use; patients lost to follow-up from treatment programs (and who have not been on treatment for some time) may not be captured in the denominator, thus potentially overestimating suppression [8].
In this study, we examined retention and viral suppression in a large public health program across 4 provinces in Zambia, a country with an estimated 1,200,000 adults living with HIV [8–10]. Building on previous work, we used a sampling-based approach in which we first selected facilities from 4 provinces (with probability proportional to facility size) and then intensively tracked a random sample of individuals (inversely proportional to facility size) lost to follow-up in each of these selected sites. In addition, in 1 of the 4 provinces (Lusaka), we assessed data on plasma HIV RNA suppression levels among a sample of both in-care and lost-to-follow-up patients. This approach yielded both a representative estimate of overall retention and viral suppression in a large region of Zambia and site-level estimates of retention with enough precision to assess site-to-site variation [11].
The protocol and study were approved by the University of Zambia Biomedical Research Ethics Committee (004-06-14), and the institutional review board of the University of Alabama at Birmingham School of Medicine (F160122006). The full analysis protocol is available in S1 Appendix. The study adhered to good practice guidelines for reporting for cohort studies as presented in the STROBE statement (S2 Appendix).
Our sampling frame consisted of HIV-positive adults 18 years or older who sought HIV care and treatment services during a 24-month period (1 August 2013 to 31 July 2015) across 64 public health facilities supported with funding from the US President’s Emergency Plan for AIDS Relief/Centers for Disease Control and Prevention through the Centre for Infectious Disease Research in Zambia (CIDRZ) in 4 provinces (Western, Lusaka, Eastern, and Southern) in Zambia. We used a multistage-sampling-based approach to obtain corrected estimates of retention and viremia [12]. Briefly, we stratified 64 total facilities by province (Eastern, Western, Southern, and Lusaka) and facility type (hospital, urban health center, and rural health center) and selected facilities within these 12 joint strata with probability proportional to size. In each selected facility, we enumerated those lost to follow-up (defined as at least 90 days late for the last visit and not documented to have died or transferred out according to the electronic medical record [EMR] system), and selected a random sample with a sampling probability inversely proportional to facility size. In 14 Lusaka facilities selected for this study, we obtained dried blood spot (DBS) samples for determining plasma HIV RNA level (viral load) from both lost patients as well as a systematic sample of patients (defined as every 10th patient) retained at the facility (S1 Fig). This study predated routine plasma HIV RNA monitoring for treatment.
Data about patient appointments and visits and sociodemographic and clinical characteristics were obtained from the EMR system in Zambia (SmartCare) and used to enumerate the lost-to-follow-up patients. Lost patients were traced between October 2015 and June 2016 by chart review, phone calls, and in-person visits within the community. We recruited peer health workers with in-depth knowledge of patient flow within facilities and familiarity with the surrounding communities to carry out tracing. Patients were classified as died if review of EMR, paper records, or the tracing process found evidence that the patient was deceased. Patients were classified as alive if spoken to in person or an informant was contacted and reported knowledge of the patient but no knowledge of death. When information about a patient was collected from more than 1 informant and was discordant, we used information from closer relations (e.g., we prioritized information from a spouse over that from a neighbor). When lost patients were contacted in person, we asked, “Have you seen any doctor, nurse or other professional health worker (like, pharmacist) for treatment of HIV since your last visit which we have on file, which was on [X date] at the [original clinic]?” and recorded the date of that subsequent visit if the answer was yes (S3 Appendix). Current care status (i.e., retained in care) was established only if found through chart review or the patient was contacted in person. Identities were established by name, nicknames, age, occupation, height, sex, and location of residence. In Lusaka, we trained tracers to collect DBS, assess sample quality, and transport DBS cards to the CIDRZ central laboratory. We used the AmpliPrep/COBAS TaqMan HIV-1 Test, version 2.0, to quantify human immunodeficiency virus type 1 (HIV-1) RNA in DBSs. Viral suppression was defined as less than 1,000 copies/ml.
We determined “naïve” estimates of retention using only data available from the facility EMR for the entire cohort of ART users (which included all patients on ART in the 2-year period of observation, 1 August 2013 to 31 July 2015) as well as for new ART initiators (who started ART in this 2-year window). We carried out “revised” estimates that incorporated tracing outcomes through use of probability weights [13]. Weights were inverse to the probability of selection at both the patient and facility level, a process that seeks to yield regionally representative estimates [13,14] (S2 Fig). In the naïve analysis, we estimated the prevalence of 4 care states during the 2-year observation period using the Aalen–Johansen method [15]: (1) alive and in care at original clinic, (2) transferred to a new facility (which included only official transfers), (3) lost to follow-up, or (4) died. In revised estimates, after incorporating findings from tracing through probability weights, we estimated the prevalence of patients over time in the following 4 states: (1) alive and in care at the original clinic, (2) transferred to a new clinic (which included both official and unofficial transfers), (3) alive but out of care, or (4) died. We used Cox proportional hazards models to identify characteristics associated with being out of care or deceased in the revised estimates. We examined the proportional hazards assumption using Schoenfeld residuals [16] (S1 Table), and, in addition to the application of inverse probability weights to account for sampling, we used inverse probability weights to address missing predictor data for CD4 count, WHO stage, marital status, and level of education [17]. We used robust variance estimates to account for clustering by clinic.
In Lusaka Province, we estimated the prevalence of viremia (viral load ≥ 1,000 copies/ml) among lost patients alone, then among patients in care at their original clinic, and finally overall (combining both) by applying sampling weights. We managed bias incurred via the inaccuracy of DBS-based viral load results by using the documented sensitivity of 80.8% and specificity of 87.3% (for detecting a viral load of ≥1,000 copies/ml) as compared to plasma HIV RNA determination through an outcome misclassification correction approach [18,19] (S4 Appendix). We used inverse probability weights to account for sampling (S1 Fig) and missing data in all analyses [17]. Post hoc analyses not predefined in the protocol include analyses restricted to the contemporary cohort of those initiating ART on or after 1 August 2013, analyses using DBS viral load outcome misclassification correction methods, and an analysis of predictors of viremia.
As described in previous work, 165,464 patients on ART had at least 1 encounter in the 64 health facilities over the 24 months between 1 August 2013 and 31 July 2015 (Fig 1) [12], of whom 28,111 (17%) were considered lost to follow-up at the time of sampling. At the 32 selected sites, 104,966 patients made any visit during that time, and 17,602 (17%) were lost to follow-up. We selected a random sample of 2,892 lost patients (16% of 17,602 lost patients at 32 selected facilities and 10% of all 28,111 lost in all 64 facilities) for intensive tracing to ascertain current care status. Of the 2,892 lost and traced, updated information was found for 2,163 (75%), of whom 1,751 (81%) were alive. Among those found alive, 836 (48%) were still in care at the original health facility, 457 (26%) had transferred to another facility, and 255 (15%) were out of care; for 203 (12%) care status remained undetermined (Fig 1). Patient characteristics among those patients lost, traced, and for whom updated care status was ascertained were similar to those of the overall population of lost patients in the total ART cohort (Table 1). Compared to the total ART cohort, new ART initiates with updated care status (N = 483) had a shorter duration of ART (88 days; IQR 1–224), mostly enrolled in care between 2013 and 2015 (85%; the other 15% enrolled before 2013, but started ART in the 2-year observation period), and appeared younger in age (median 33 years; IQR 28–40) (S2 Table).
Among all patients at 2 years, using only EMR data, we found that 67.7% of patients were retained at the original clinic (95% CI 67.3%–68.3%), 26.5% were lost (95% CI 26.1%–26.8%), 4.6% had officially transferred to a new facility (95% CI 4.5%–4.7%), and 1.2% had died (95% CI 1.1%–1.2%) (Fig 2; S3 Table)—indicating that 72.3% (95% CI 71.8%–73.0%) were retained at 2 years (at original clinic or transferred). After incorporating updated tracing outcomes, the revised 2-year estimates suggested that 76.5% (95% CI 76.0%–76.9%) were retained at the original clinic, 14.7% (95% CI 14.5%–14.9%) had officially or unofficially transferred to a new site, 3.9% (95% CI 3.8%–4.1%) were alive and out of care, and 4.9% (95% CI 4.8%–5.0%) had died (Fig 2; S3 Table), resulting in an updated estimate of 91.2% (95% CI 90.5%–91.8%) retained (at original clinic or transferred) at 2 years.
Compared to the total ART cohort, new ART initiators were less likely to be retained. Two-year estimates for this group using only EMR data showed 35.9% were retained (95% CI 31.8%–39.9%), 55.1% were lost, 6.8% transferred (95% CI 6.2%–7.5%), and 2.2% died (95% CI 1.7%–2.8%) (Fig 3; S4 Table), with a total of 42.7% (95% CI 38.1%–47.0%) retained in care (at original clinic or transferred). Revised Aalen–Johansen estimates incorporating tracing outcomes through probability weights showed a cumulative proportion of 44.2% who were retained in care at original clinic (95% CI 40%–48%), 33.1% who had transferred to new clinics (95% CI 30.7%–35.6%), 9.6% who were out of care (95% CI 8.7%–10.5%), and 13.1% who had died (95% CI 12.2%–14.1%) (Fig 3; S4 Table), resulting in updated estimates of 77.3% (95% CI 70.5%–84.0%) retained (at the original clinic or transferred) at 2 years.
Revised rates of stopping care varied markedly across health facilities, ranging between 1.3 and 8.8 per 100 person-years (pyrs) in the total cohort (Fig 4a), and 1.8 and 26.3 per 100 pyrs among the new ART initiators (Fig 4b), and across the 4 provinces, ranging from 4.0 (95% CI 3.5–4.5) per 100 pyrs in Eastern Province to 5.5 (95% CI 4.9–6.2) per 100 pyrs in Lusaka Province in the total ART cohort, and from 9.1 (95% CI 7.5–11.0) per 100 pyrs in Eastern Province to 12.5 (95% CI 10.3–15.2) per 100 pyrs in Lusaka Province among new ART initiates (S3 Fig).
At the individual level, the characteristics most strongly associated with disengagement were male sex (hazard ratio [HR] 1.82; 95% CI 1.47–2.25; p < 0.001) and years on ART (HR 0.81; 95% CI 0.79–0.84; p < 0.001). In addition, a low CD4 count or being divorced as compared to married at enrollment showed an association with higher disengagement (Table 2). In the sample of new ART initiates, enrollment CD4 count, WHO stage, and divorce were associated with disengagement.
Among the 86,688 patients who initiated ART during the study period in Lusaka Province (where we sought to estimate the prevalence of viremia), 68,934 were retained in care and 17,754 were lost to follow-up. Of a random sample of 798 lost patients who were eligible for tracing, 400 (50.1%) could not be traced, and for 255 (32.0%), samples could not be obtained due to either refusals or logistical challenges, resulting in 143 (17.9%) DBS viral load samples. Characteristics were similar for eligible patients with and without viral load samples (S5 Table). In a systematic sample of retained patients, we obtained 901 DBS viral load samples. In combination, we analyzed 1,044 DBS viral load results (S1 Fig). After applying inverse probability weights for sampling and nonresponse, and bias correction for known misclassification of DBS-based HIV RNA levels (as compared to plasma HIV RNA levels), we found the prevalence of viremia among patients retained at their original health facility (using a threshold of 1,000 copies/ml) to be 18.1% (95% CI 14.0%–22.3%). Among the lost patients, which included both those reporting no care as well as those who unofficially transferred, 71.3% (95% CI 58.2%–84.4%) were viremic. Unofficial transfers and patients out of care had a prevalence of viremia of 49.8% (95% CI 28.1%–71.4%) and 83.9% (95% CI 67.2%–98.8%), respectively. Incorporating results among those lost and traced into the underlying cohort using probability weights yielded an overall prevalence of viremia of 24.7% (95% CI 21.0%–29.3%). In multivariable regression using pre-therapy patient characteristics (Table 3), male sex, younger age, and lower ART initiation CD4 count were associated with viremia. In a model with current care status, this factor was most strongly associated with viremia. Male sex and time on ART diminished in significance, but younger age remained strongly associated with viremia in the model with current care status.
We combined targeted supplemental data collection (for updated care status and viral loads) with large-scale data from a national EMR system to advance our understanding of the public health response to HIV in Zambia. First, we found that even though a large percentage of patients missed visits and became lost to follow-up, most patients returned to care and relatively few patients stopped care altogether. Second, these lapses in retention varied markedly from 1.8 to 26.3 per 100 pyrs among new ART initiates attending 32 health facilities across 4 provinces we studied, differences that were incompletely explained by measured patient and facility characteristics. Third, unsuppressed HIV RNA levels in a population of treated patients rose substantially when lost-to-follow-up patients were included in estimates. In this study, viremia rose by 7% on an absolute scale and nearly 40% on a ratio scale when lost patients were included in the estimates. Patients who were retained within the greater health system, but not at their original clinics, contributed substantially to the total viremia in the population: Patient-initiated transfers of care were not well coordinated and safe. These findings suggest that public health HIV treatment services in Zambia, while accomplishing an enormous task and saving thousands of lives, are of uneven success: Many patients do not achieve optimal sustained engagement, and they experience viremia and therefore attenuated clinical benefits of HIV treatment.
When compared to the large-scale cross-sectional ZAMPHIA—which documents a prevalence of viral suppression among current HIV ART users of 90% in Zambia [8]—our longitudinal data suggest several additional observations. First, we find the prevalence of viremia in a population treated within the last 2 years to be 25% when those lost to follow-up are incorporated into estimates. This is substantially higher than the 10% estimated in ZAMPHIA. Although this difference could be due in part to measurement error (i.e., limited sensitivity and specificity) of DBSs, we sought to manage these consequences through bias correction methods. Another important possibility, however, is that cross-sectional studies do not fully capture those patients who had been on treatment, but who stopped treatment prior to their participation in the survey. If these patients do not admit to previous treatment (due to social desirability bias) or if instruments only ask about current HIV treatment, the denominator could be artificially small (and viral suppression overestimated) compared to this analysis. In either case, we found our internal estimates of viral suppression among those in clinic care to be much higher when lost-to-follow-up patients were included.
The revised estimates of retention and viral suppression do not just change the numerical estimates, but further illustrate that retention is a multidimensional, complex outcome that likely requires adaptive, innovative, longitudinal public health practices to improve. On the one hand, the prevalence of true disengagement (i.e., being alive and out of care) is much lower in reality than as shown by estimates using EMR data alone. On the other hand, patients who drop out of care at one site and reenter at another are much more likely to be viremic (at approximately 50%). Collectively, these 2 observations direct our attention to an important reality in public health chronic disease management: Patients may move their residence, and their prioritization of treatment may wax and wane, and these changes represent periods of vulnerability [20–22]. Innovations to improve the effectiveness of ART programs, such as differentiated service delivery, as well as others, must adapt to patients’ lived realities in order to support durable, long-term engagement in care and viral suppression [23–25].
Ongoing efforts to enhance retention and viral suppression are underway, both in the environment of this study in Zambia and beyond, but will need continued monitoring, including among those inevitably lost to follow-up. Zambia is rapidly scaling up targeted quality improvement activities in response to these findings, including wider use of differentiated service delivery models. Although the improvement of estimates of retention through tracing a sample of lost patients has been demonstrated in sub-Saharan Africa [5,26,27], this analysis also highlights the marked heterogeneity across facility-level estimates. The heterogeneity implies that the intensification or prioritization of retention strategies should be targeted, to be most efficient, and one should endeavor to understand facility-level and facility–patient interaction dynamics when implementing support efforts. This targeting of health system improvements is aligned with current strategic thinking about targeting resources to those most in need [5,25,26]. In addition, interventions that facilitate reengagement in care for those who are found to be disengaged from care are needed. Data suggest that early tracing after a missed visit (within 1 week) can improve patient contact and return to care [27–29], an intervention that should be considered in this setting, but must occur alongside improvements in EMR systems and data management to minimize misclassification and wasted tracing efforts.
In our analysis, we observed that the strongest predictors of disengagement were male sex and advanced HIV disease [5,27]. The recent introduction of the concept of differentiated care for patients with advanced HIV disease could, over time, have an impact on this higher risk of disengagement among those who present late to HIV services [24]. Effective approaches to retaining men in care in sub-Saharan Africa are less clearly defined [30,31]; however, interventions such as home, mobile, or workplace ART distribution and financial incentives that target men should be conceptualized with consideration of the unique facility characteristics and community dynamics relevant to the settings where these services may be implemented. Conceptually, those lost to follow-up will have greater prevalence of viremia than those in care. If this fraction of lost patients is large, then their overall contribution to viremia in a population is important. These data offer proof of that concept. Among those in care, only 1 in 6 patients are viremic, whereas over half of those lost to follow-up are viremic. Of note, patients who silently transferred (i.e., had no official transfer documentation) between facilities also had a high risk of viremia, highlighting the contribution of treatment gaps during transfer to the overall infectiousness of this community. Efforts should be made to ensure rapid reengagement in care among those with missed visits by simplifying transfer systems and educating patients and staff to monitor and document reengagement processes at original or new facilities.
This study has a number of limitations. Our sample, even though randomly selected, was affected by imperfect response rates: We did not ascertain outcomes in all those who were traced. In addition, we ascertained true care status among lost patients by self-report, which could be influenced by social desirability bias. Self-reported retention status, however, was highly associated with viral load, lending credibility to this measurement. Furthermore, our competing risk estimates assumed that unofficial and official transfers remained engaged in care, an assumption that could lead to an overestimation of retention. Viral load measurements used DBSs, which have known limitations in sensitivity and specificity compared to the gold standard plasma-based assay; we however used established methods of bias analysis to correct for DBS inaccuracy.
This study demonstrates how a strategy of sampling and tracing of lost patients can be used to generate revised estimates of retention and viremia at a population level. For Zambia these estimates reflected better overall retention than routinely collected program data but also highlighted significant gaps in care, and marked variation of overall retention at the facility level among those retained, contributing to high overall viremia in the ART cohort. Substantial efforts need to be made to tailor services to the needs of patients in order to reduce lapses in care and maintain long-term viral suppression. Furthermore, understanding facility- and community-based barriers to retention in care and addressing these barriers remain critical to attaining the UNAIDS 90-90-90 targets.
|
10.1371/journal.ppat.1000766 | IFN-α-Induced Upregulation of CCR5 Leads to Expanded HIV Tropism In Vivo | Chronic immune activation and inflammation (e.g., as manifest by production of type I interferons) are major determinants of disease progression in primate lentivirus infections. To investigate the impact of such activation on intrathymic T-cell production, we studied infection of the human thymus implants of SCID-hu Thy/Liv mice with X4 and R5 HIV. X4 HIV was observed to infect CD3−CD4+CD8−CXCR4+CCR5− intrathymic T-cell progenitors (ITTP) and to abrogate thymopoiesis. R5 HIV, by contrast, first established a nonpathogenic infection of thymic macrophages and then, after many weeks, began to replicate in ITTP. We demonstrate here that the tropism of R5 HIV is expanded and pathogenicity enhanced by upregulation of CCR5 on these key T-cell progenitors. Such CCR5 induction was mediated by interferon-α (IFN-α) in both thymic organ cultures and in SCID-hu mice, and antibody neutralization of IFN-α in R5 HIV-infected SCID-hu mice inhibited both CCR5 upregulation and infection of the T-cell progenitors. These observations suggest a mechanism by which IFN-α production may paradoxically expand the tropism of R5 HIV and, in so doing, accelerate disease progression.
| Human immunodeficiency virus (HIV), a lentivirus, is the causative agent of AIDS. Chronic immune activation and inflammation are major determinants of disease progression in primate lentivirus infections and are associated with the production of type I interferon. To investigate the impact of type I interferon on HIV infection, we studied the human thymus implants of SCID-hu Thy/Liv mice infected with HIV that uses either CXCR4 (X4 HIV) or CCR5 (R5 HIV) as a coreceptor. X4 HIV was observed to infect T-cell progenitors in the thymus and to disrupt T-cell production by that organ. R5 HIV, by contrast, first established a nondisruptive infection of thymic macrophages and then began to infect intrathymic T-cell progenitors. We report here that the tropism of R5 HIV is expanded and T-cell disruption enhanced by increased expression of CCR5 on these key T-cell progenitors. Such CCR5 induction was mediated by interferon-α (IFN-α) in both thymic organ cultures and in SCID-hu mice. Moreover, antibody neutralization of IFN-α in R5 HIV-infected SCID-hu mice inhibited both CCR5 upregulation and infection of the T-cell progenitors. These observations suggest a mechanism by which IFN-α may paradoxically expand the tropism of R5 HIV and accelerate disease progression.
| HIV disease progression is marked by chronic immune activation associated with accelerated destruction of T cells in the periphery and diminished production of new T cells from progenitors in the thymus and elsewhere [1],[2]. Although the detection of X4 HIV as the predominant viral species in peripheral blood is clearly associated with a higher risk of disease progression, about half of patients progress to AIDS in the presence of R5 viruses alone [3],[4] or with only the transient appearance of X4 virus [5]. Since it is just a small fraction of CD4+ target cells that express the CCR5 coreceptor [6], the mechanisms underlying such intrinsic R5 virus pathogenicity remain unclear. Given the association between high levels of T-cell activation and more rapid disease progression in untreated HIV-infected individuals [7], however, it is possible that such activation might induce the upregulation of CCR5 and expand the tropism of R5 HIV to include essential T-cell progenitors that are normally spared.
To address the hypothesis that R5 HIV infection might lead to such an indirect expansion of tropism in vivo, we investigated the course of R5 HIV infection in the SCID-hu Thy/Liv mouse model of human T-cell production. This small animal model, in which severe combined immunodeficient (C.B-17 SCID) mice are implanted with human fetal thymus and liver under the kidney capsule, supports multilineage human hematopoiesis, including T lymphopoiesis, for periods up to one year [8] and represents a venue in which to study the effects of HIV on human thymopoiesis in vivo. After inoculation with X4 HIV, a key population of ITTPs (CD3−CD4+CD8−CXCR4+CCR5−) is rapidly infected and destroyed, impeding thymocyte maturation and depleting the implants of thymocytes within 4–5 weeks [9],[10]. In contrast, rapid destruction of the thymic organ is not observed after infection with the R5 isolate Ba-L, which follows a biphasic process involving nonpathogenic replication in medullary stromal macrophages followed by cytopathic replication in thymocytes after 6 weeks of infection [11]. CCR5 is expressed at much lower levels than CXCR4 (<5% versus 30–40% of thymocytes) at all stages of T-cell development in the thymus [6],[12],[13], and this may explain the decreased pathogenicity of R5 HIV in that organ.
We demonstrate here that R5 HIV causes eventual depletion of thymocytes that is associated with de novo IFN-α-mediated upregulation of CCR5 on ITTP, rendering these key progenitor cells permissive for R5 HIV infection and depletion. Moreover, we show that monoclonal antibody (MAb) neutralization of IFN-α in SCID-hu Thy/Liv mice inhibits CCR5 induction after HIV infection and prevents infection of ITTP with R5 HIV. The observation that IFN-α may be a driving force behind expanded HIV tropism in vivo offers a proximal mechanism for the relationship between immune activation and disease progression and suggests that immunomodulatory agents that suppress the production or the effects of IFN-α may serve to slow disease progression in the HIV-infected host.
The human thymus implants of SCID-hu Thy/Liv mice were inoculated with the X4 HIV clone NL4-3, the R5 HIV isolate Ba-L, or a chimeric R5 clone of NL4-3 containing the V1-V3 env regions of Ba-L (81A) and monitored for viral replication and thymocyte depletion at 21 and 42 days. As expected from our previous work in the Thy/Liv model [11],[14],[15],[16], viral replication resulted in time-dependent increases in implant HIV RNA, p24, Gag-p24+ thymocytes, and MHC class I expression on CD3intCD4+CD8+ (double-positive, DP) thymocytes (Figure 1A). Viral replication was accompanied by time-dependent decreases in implant cellularity, thymocyte viability, percentage of DP thymocytes, and CD4/CD8 ratio (Figure 1B) that were more rapid and of greater magnitude for X4 than for R5 HIV, a finding consistent with the far greater of expression of CXCR4 than CCR5 on human thymocyte subpopulations [6].
The slow but evident pathogenicity of R5 HIV may be dependent upon inductive events that take place after infection. For instance, progressive sequence variation in the env gene may enable a “switch” of envelope glycoproteins to a pathogenic X4 phenotype. Alternatively, R5 viral pathogenesis may proceed in a time-dependent manner through infection of ITTP, which constitute a minor but key thymocyte progenitor subpopulation, in a manner analogous to X4 thymic pathogenesis [10]. Since we have previously found that an R5-to-X4 phenotypic switch is not detectable during thymic infection with Ba-L [11], we more closely evaluated the possibility that ITTP, which are normally CCR5-negative [6],[12], might be infected at some time point after virus inoculation.
Intracellular Gag-p24 staining in concert with surface staining for CD3, CD4, and CD8 revealed that ITTP were infected by Ba-L and 81A at day 42 but not day 21 (Figure 1C). This finding was unexpected because ITTP do not normally express CCR5 [6] and were thus not considered targets of R5 HIV infection, in marked contrast to the susceptibility of ITTP to X4 HIV infection as a consequence of high-level expression of CXCR4 [6]. Reasoning that CCR5 expression might be indirectly induced by HIV infection, we evaluated CCR5 expression on thymocyte subpopulations after HIV inoculation and found, at day 42 but not day 21, statistically significant increases in the percentage of CCR5-expressing ITTP [to 6.4±1.5% (P = 0.017) for Ba-L and to 3.2±0.2% for 81A (P = 0.001) versus a mean of 1.3±0.2% for mock-infected implants] (Figure 1D and E). Less dramatic, but still statistically significant increases in CCR5-positive CD3+CD4−CD8+ (single-positive, SP8) thymocytes were also observed, as has been reported previously in NOD/SCID-hu BLT mice infected intravaginally with HIV and attributed to a heightened state of immune activation [17]. Significant increases in the percentage of CCR5+ ITTP were also observed in X4 NL4-3-infected implants (Figure 1D). Treatment of SCID-hu Thy/Liv mice with 3TC (lamivudine) inhibited the induction of CCR5 on thymic progenitors and prevented Ba-L-mediated thymocyte depletion (data not shown). These results indicate that induction of CCR5 in HIV-infected Thy/Liv implants occurs in a time-dependent manner that is dependent on active HIV (R5 or X4) replication.
Previous reports have demonstrated that CCR5 expression can be increased on several cell types after treatment with cytokines including IL-2 [18], IL-4 [19], IL-10 [20],[21], IL-15 [22], TGF-β [23], and IFN-γ [24], and with HIV Tat [25]. When human thymic organ cultures were incubated with these and other cytokines, significant induction of CCR5 expression on human thymocytes was only observed after treatment with IFN-α (Figure 2A). Analysis of CCR5 expression on thymocyte subpopulations demonstrated statistically significant upregulation on ITTP (Figure 2B), the same key subpopulation found to upregulate CCR5 in the HIV infected thymic implant and which we have previously reported to express the IFN-α/β receptor [26]. This receptor is expressed at high levels on ITTP and at progressively lower levels on more mature thymocytes (e.g., DP, SP4, and SP8 thymocytes) [26]. The ability of IFN-α to induce expression of CCR5 is consistent with the presence of STAT-binding sites at nt −55 and −116 in the CCR5 promoter; mutation of the proximal STAT site nearly abolishes promoter activity [27].
To determine whether IFN-α can induce expression of CCR5 on ITTP in vivo, we treated groups of SCID-hu Thy/Liv mice in three separate cohorts (A, B, and C) with IFN-α2b (Intron A or pegylated interferon alfa-2b) by once-daily intraperitoneal (i.p.) injection for 6 or 13 days. Significant increases in the percentage of CCR5+ ITTP were observed at both time points, normalizing to pretreatment levels after discontinuation of IFN-α (Figure 2B and C). Treatment with IFN-α for 13 days had no effect on the percentage and absolute number of ITTP or other more mature thymocyte subpopulations present in the implants (data not shown), so it is unlikely that the increase in CCR5+ cells is the result of either IFN-α-mediated apoptosis of CCR5-negative ITTP or an increased rate of CCR5+ ITTP cell division. Of note, the percentage of CCR5+ ITTP in IFN-α-treated mice (means in the three cohorts: 6–20%, range: 3–37%) (Figure 2B and C) tended to be higher than that found in Ba-L-infected mice at day 42 (mean: 6%, range: 0–12%) (Figure 1D and E), a difference that may be the result of virus-mediated depletion of infected CCR5+ progenitors.
To examine more closely the relationship between CCR5+ induction on ITTP, thymic organ infection, and thymocyte depletion, we studied two additional SCID-hu Thy/Liv cohorts (D and E) inoculated with Ba-L plus two cohorts (G and H) inoculated with the R5 isolate, CC1/85. Data for these cohorts were analyzed together with the data obtained from the Ba-L and 81A-infected mice shown in Figure 1 (cohort F). Implant viral loads measured 42–49 days after inoculation were within 1.0 log10 across all six infected groups (means of 4.4–5.4 log10 copies HIV RNA and 80–800 pg p24 per 106 cells) (Figure 3A). For the two additional SCID-hu cohorts inoculated with Ba-L (D and E), implants collected at much later time points (up to one year) after inoculation showed progressively more severe thymocyte depletion, while mock-infected implants remained intact with ∼80% DP thymocytes (Figure 3B). As we have reported previously [11], such depletion became noticeable 6 weeks after inoculation with Ba-L; we accordingly focused on implants collected from the five infected cohorts at this time (i.e., days 42–49) (Figure 3C). Although the decreases in implant cellularity, thymocyte viability, and percentage of DP thymocyte were often not statistically significant when individual experiments were compared (likely the result of the small number of mock-infected mice), animals in the R5 HIV-infected cohorts showed a trend towards decreases in each of these parameters. Concomitantly, there were increases in the percentage of CCR5+ ITTP compared to mock-infected mice, and the percentages of CCR5+ ITTP were comparable to the percentages of ITTP that were Gag-p24+ (Figure 3D).
To better document the relationship between CCR5 expression on ITTP and thymocyte depletion, data for each individual implant were plotted to show the correlation between the percentage of CCR5+ ITTP and viral load (Figure 4A) as well as the percentage of CCR5+ ITTP and thymocyte depletion (Figure 4B). Not only are the correlations highly significant for infected implants in statistical terms (e.g., P<0.0001 for CCR5+ ITTP versus both Gag-p24+ ITTP and thymocyte viability), but the proportion of CCR5+ ITTP corresponds closely to that of infected ITTP (Figure 4A). In contrast, there was no correlation between CCR5+ ITTP and markers of thymocyte depletion for mock-infected implants (Figure 4B). Accordingly, it is highly likely that induction of CCR5 expression on ITTP is a causal event precipitating thymocyte depletion after HIV infection.
To show definitively that upregulation of CCR5 on ITTP was mediated by IFN-α we treated three cohorts (I, J, and K) of SCID-hu Thy/Liv mice with a broadly neutralizing mouse MAb against multiple human IFN-α subtypes. Mice were treated by three times weekly i.p. injection, beginning 2 days before Ba-L, 81A, or NL4-3 inoculation and continuing until implant collection. For mice infected with Ba-L, neutralization of IFN-α was found to result in a lower percentage of CCR5+ ITTP (P<0.05 in cohort I and P<0.01 in cohort J), a lower percentage of Gag-p24+ total live thymocytes (P<0.05 in both cohort I and J), and a lower percentage of Gag-p24+ ITTP (P<0.01 in cohort I and J) (Figure 5). In cohort K, we directly compared the effects of IFN-α neutralization on HIV 81A (R5) and NL4-3 (X4) infection. This experiment was carried out with the expectation that infection of ITTP would be inhibited after 81A, but not after NL4-3, inoculation. We found that this was indeed the case: there was a 93% reduction (P = 0.005) in Gag-p24+ ITTP after IFN-α-treatment in 81A-infected mice yet an insignificant 25% reduction (P = 0.501) in Gag-p24+ ITTP in treated NL4-3-infected mice. This was accompanied by expected reductions in CCR5+ ITTP for both viruses (89% reduction for 81A; P = 0.004 and 67% reduction for NL4-3; P = 0.083). Given our previous data showing that infection of ITTP leads to interruption of thymopoiesis [10], these results indicate that IFN-α-induced upregulation of CCR5 on ITTP is likely to result in diminished production of T cells from the thymus.
R5 isolates of HIV have been associated with disease progression in HIV-infected individuals [28]. Likewise, as we have shown here, R5 HIV can be pathogenic in the SCID-hu Thy/Liv model of human thymopoiesis. Even though there is little CCR5 expression in the human thymus, R5 HIV was found to induce delayed but significant depletion of developing DP thymocytes and reduction in implant cellularity, and progression of R5 infection was found to correlate with the induction of CCR5 expression on early thymic progenitor cells. Such induction, in turn, is mediated by IFN-α both in vitro and in vivo. This finding is in contrast to a previous report showing that R5 HIV infection of thymic organ cultures induced CCR5 on CD4+ thymocytes through the production of IL-10 and TGF-β [29]. The ability of HIV to induce expression of its own coreceptor through the major antiviral cytokine, IFN-α likely evolved to dampen this antiviral defense mechanism, a counterbalancing act that has been likened to a détente through which virus and host achieve conditions for coexistence [30].
The above results indicate that expanded tropism of R5 HIV in the infected human thymus (to ITTP and DP thymocytes) is a secondary event that occurs after the induction of IFN-α production, most likely from plasmacytoid dendritic cells (pDC). These cells function as part of the innate immune response by secreting large quantities of IFN-α after contact or infection with a wide range of viruses, including HIV [31],[32],[33]. IL-3Rα+ pDC reside in the medulla of the human thymus [34], and we have previously shown that these cells produce IFN-α in response to HIV infection in both human thymic organ culture and in SCID-hu Thy/Liv mice [26]. Intrathymic pDC express both CXCR4 and CCR5 and are themselves targets for HIV replication [35], although it is not known if infection of these cells plays a role in IFN-α secretion. In sum, interactions between R5 HIV and pDC might lead indirectly to upregulation of CCR5 on cells that are normally not permissive for R5 infection. If so, these data point to a critical role for pDC-mediated IFN-α secretion in R5 HIV pathogenesis in the thymus of the SCID-hu mouse.
There is a low frequency of CCR5+ pDC in the CD3−CD4+CD8− thymocyte population (unpublished observations), but we believe our results are due to upregulation of CCR5 on the T-lineage component of this population for the following reasons: First, in vitro IFN-α treatment results in upregulation of CCR5 on ITTP and DP (Figure 2A), cell populations that are T-lineage and that express high levels of the IFN-α/β receptor [26]. In vivo, the same phenomenon occurs (Figure 2B). Second, The fraction of CCR5+ ITTP is similar to the fraction of p24+ ITTP (Figure 3D), and there is a significant relationship between the two when analyzed in a large number of animals (Figure 4A). Finally, neutralizing anti-IFN-α antibody blocks the upregulation of CCR5 on the ITTP population (Figure 5). All of these data (especially the data in Figure 5) are most consistent with IFN-α induction of CCR5 on the CD3−CD4+CD8−CCR5− ITTP and the CD3+CD4+CD8+CCR5− DP populations, both of which are permissive for infection and replication of HIV.
These data also illustrate the importance of cell-cell interactions that can occur in lymphoid tissue after HIV infection with profound influence on the course of disease progression and that are not easily replicated in dispersed cell cultures. In addition, available in vitro culture systems do not persist for the periods of time required to measure the impact of these interactions on HIV pathogenesis. The observations in this study thus underscore the need for a closer evaluation of the dynamics of HIV infection within lymphoid organs and provide experimental justification for such tissue analysis within HIV-infected human subjects.
The finding that IFN-α can enhance HIV infectivity is surprising, especially given the potent antiviral activity against HIV we and others have reported in IFN-α-treated thymic organ cultures [26],[36]. These counterposing effects of IFN-α may occur simultaneously in pDC-containing tissue, thereby contributing to the slow progression of thymocyte depletion usually seen after R5 infection. Persistently high levels of IFN-α and of IFN-inducible genes are associated with more rapid disease progression in SIV-infected macaques [37],[38],[39]. In contrast, nonpathogenic SIV infections are associated with transient IFN-α responses, possibly due to the inability of the virus to activate pDC [40]. There is likely a complicated set of kinetics at play during HIV infection of the Thy/Liv implant in vivo, including but not limited to: the rate of viral replication and spread; the rate of induction of IFN-α in pDC; the rate of upregulation of CCR5 on thymocytes that express the IFN-α/β receptor; the rate at which these cells are infected and destroyed by R5 HIV; the rate at which they are replenished from earlier, CCR5− progenitors; and, not least, the rate at which more mature DP thymocytes are depleted. We presume that the late events observed after HIV infection represent a sum total of these and other counterposing rates, resulting eventually in complete depletion of double-positive thymocytes (e.g., by day 300 in Figure 3B).
IFN-α has been shown to inhibit thymic T-cell differentiation in both the mouse [41] and human [42]. IFN-α-mediated inhibition of T-cell development may have also contributed to the depletion of thymocyte subsets observed in this study; however, we found that treatment of the mice with IFN-α for 13 days had no effect on the percentage and absolute number of ITTP or other more mature thymocyte subpopulations present in the Thy/Liv implants. It is possible that more prolonged exposure of the implants to IFN-α over months of HIV infection may have more deleterious cumulative effects on T-cell maturation than relatively short-term IFN-α treatment.
The role of chemokine coreceptor utilization in HIV disease progression has been studied extensively. The switch of viral phenotype from R5 to X4 has a profound and negative effect on absolute CD4 cell counts [43] and has been implicated as a determining factor in accelerated disease progression [5]. However, it appears that R5 HIV [28] and SIV [44] have the capacity to be pathogenic in their own right. HIV can also evolve in vivo with increased affinity for CCR5, thus acquiring the ability to infect cells expressing low levels of the coreceptor and potentially increasing pathogenicity [45],[46]. We present evidence here that CCR5 induction resulting from IFN-α secretion by pDC plays a significant role in the pathogenesis of R5 HIV in the human thymus implant of the SCID-hu Thy/Liv mouse. Given the close structural and functional similarities between this model and the intact human thymus [8] as well as prior evidence that HIV can infect the thymus in humans [47],[48],[49],[50],[51], it is likely that these observations are relevant not only to the HIV-infected child with abundant thymic tissue but also to the HIV-infected adult, in whom residual thymic function can continue to play a role in the de novo production of naïve T cells [1].
Since pDC are resident throughout the lymphoid system and migrate to inflamed lymph nodes [52], the expansion of R5 HIV tropism described here in the human thymus may also occur in other organs of the hematolymphoid system. Indeed, recent data indicate that IFN-α treatment causes significant increases in CCR5 mRNA expression in PBMC cultures from both HIV-infected and uninfected individuals [53],[54], and IFN-α treatment of patients with uveitis resulted in increases of CCR5 expression on peripheral blood CD4+ T cells [55]. Even if these events are restricted to thymic pDC, residual thymic function that persists in some adults with HIV disease [1] might thereby be abrogated. Alone or together, such interactions between HIV, pDC, and normally CCR5-negative target cells might underlie disease progression induced by R5 viruses in vivo.
The following reagents were obtained through the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH: pNL4-3 [56] from Dr. Malcolm Martin, HIV-1Ba-L [57] from Dr. Suzanne Gartner, Dr. Mikulas Popovic and Dr. Robert Gallo, and p81A-4 [58] (Cat#11440) from Dr. Bruce Chesebro. CC1/85 [59] was generously provided by Drs. Shawn Kuhmann and John Moore. Ba-L is a low-passage isolate that has been propagated exclusively in human monocyte/macrophages [57], and CC1/85 is a well-characterized patient isolate that has also been minimally lab adapted [59],[60]. Working stocks of NL4-3 and 81A were prepared by lipofectamine (Invitrogen) transfection of 293T cells and collection of supernatants on day 2. Ba-L stock was generated in monocyte-derived macrophages with the supernatant collected on day 8, and CC1/85 stock was generated in phytohemagglutinin (PHA)-activated peripheral blood mononuclear cells (PBMC) with the supernatants collected on day 4. Virus stocks were titrated by limiting dilution for 50% tissue culture infectious doses (TCID50) in PHA-activated PBMC with p24 detection by ELISA on day 7 as previously described [61].
Human fetal thymus and liver were obtained through services provided by a nonprofit organization (Advanced Bioscience Resources) in accordance with federal, state, and local regulations. Coimplantation of thymus and liver pieces under the kidney capsule to generate SCID-hu Thy/Liv mice and inoculation of the Thy/Liv implants with HIV was performed as described [14],[62]. Male C.B-17 SCID (model #CB17SC-M, homozygous, C.B-Igh-1b/IcrTac-Prkdcscid) mice were obtained at 6–8 weeks of age from Taconic, and cohorts of 50–60 SCID-hu Thy/Liv mice were implanted with tissues from a single donor. Implanted mice were maintained in a barrier facility under pathogen-free conditions and inoculated 18 weeks after implantation with 50 µl of stock virus (1,000 TCID50) or conditioned medium from PBMC cultures (mock infection) by direct injection into the implant. All procedures with mice were approved by the UCSF Institutional Animal Care and Use Committee. The Thy/Liv implants were collected from euthanized mice at the indicated time points, placed into sterile PBS-FBS, and dispersed through nylon mesh into a single cell suspension. Cells were counted and processed for p24 ELISA, branched DNA assay, and flow cytometry as previously described [14],[15].
Dispersed implant cells were stained with MAbs against CD3, CD4, CD8, MHC class I, CCR5, and intracellular Gag-p24. Pellets containing 106 cells were resuspended in 50 µl of a MAb mixture containing phycoerythrin cyanine dye CY7-conjugated anti-CD4 (BD Biosciences), phycoerythrin cyanine dye CY5.5-conjugated anti-CD8 (Caltag Laboratories), allophycocyanin cyanine dye CY7-conjugated anti-CD3 (eBiosciences), allophycocyanin-conjugated anti-CD195 (CCR5, clone 2D7) (BD Biosciences), and phycoerythrin-conjugated anti-W6/32 (DakoCytomation) in PBS containing 0.8 mg/ml human IgG (Biodesign International). Cells from one implant were also stained with conjugated, isotype-matched antibodies to control for nonspecific antibody binding. Cells were incubated for 30 min in the dark and washed two times with PBS/2% FBS. Cells were resuspended in 200 µl of a fixation/permeabilization mixture containing 1.25% human IgG (Biodesign International), 1.2% paraformaldehyde (Sigma), and 0.5% polyoxyethylenesorbitan (Tween 20, Sigma) in PBS/2% FBS. Cells were incubated for 60 min in the dark, washed two times with PBS/2% FBS, and then resuspended in 50 µl of PBS containing fluorescein isothiocyanate-conjugated anti-p24 (Beckman Coulter) and 0.8 mg/ml human IgG (Biodesign International). In addition, a “fluorescence minus one” (FMO) control was prepared in which the anti-p24-FITC was omitted from the antibody mixture to allow for discrimination of Gag-p24+ from Gag-p24− cells. Cells were incubated for 30 min in the dark, washed twice with PBS/2% FBS, resuspended in 200 µl of PBS/2% FBS in 1.5-ml tubes, and analyzed on an LSR II (BD Biosciences) with FlowJo software (Tree Star). Optimization of fluorescence compensation for correction of fluorescence spectral overlaps emitted from the fluorescent conjugated antibodies was achieved by staining cells with each antibody alone plus anti-mouse Ig kappa chain and negative control BD CompBeads (BD Biosciences), as directed by the manufacturer. After collecting 100,000 total cell events, percentages of marker-positive (CD4+, CD8+, and DP) thymocytes in the implant samples were determined by first gating on a live lymphoid cell population identified by forward- and side-scatter characteristics and then by CD3 expression. In addition, the fraction of cells positive for Gag-p24 and CCR5 was determined for all thymocyte subpopulations in each implant (Figure S1). W6/32-positive mean fluorescence intensity (MFI) of DP thymocytes was determined for each sample, and CD4/CD8 ratios were calculated by dividing the percentage of CD4+ cells by the percentage of CD8+ cells for each individual implant.
SCID-hu Thy/Liv mice from three cohorts (A, B, and C) were treated with 106 IU recombinant interferon alfa-2b (Schering), 10 µg pegylated interferon alfa-2b (Schering), or sterile water by once-daily i.p. injection for 6 or 13 days. Implants were collected and stained for flow cytometry either 1 or 7 days after the last IFN-α injection. SCID-hu Thy/Liv mice from three cohorts (I, J, and K) were treated with a mouse MAb with broadly neutralizing activity against multiple human IFN-αs (clone 9F3.18.5 [63], 500 µg every other day by i.p. injection) kindly provided by Drs. Andrew C. Chan and Kerstin Schmidt (Genentech) beginning 2 days before implant injection with Ba-L or 81A. The 9F3 MAb does not neutralize IFN-β [63].
Fetal thymus was dissected into small pieces and plated on sterile filters (Millipore) placed on gelatin sponges (Pharmacia and Upjohn) in 700 µl Yssel's medium containing 1% human serum (Gemini Bio-Products) in 24-well plates. Cultures were incubated in the presence of various cytokines or HIV Tat at concentrations shown previously to induce CCR5 upregulation, e.g., at 10 ng/ml for IL-10 [21], IL-15 [22], HIV Tat [25], and TGF-β [23]; 20 ng/ml for IL-4 [19] and IFN-γ [24]; and 20 IU/ml for IL-2 [18]. In the case of IFN-α, a dose of 1,000 IU/ml was selected on the basis of dose-ranging experiments, although CCR5 upregulation was observed at lower (300 IU/ml) IFN-α concentrations (data not shown). Cytokine-treated thymus cultures were dispersed after 3 days, and cells were stained with MAbs to CD3, CD4, CD8, and CCR5 for flow cytometry as described above.
Results are expressed as means±SEM. Nonparametric statistical analysis was performed by use of the Mann-Whitney U test (StatView 5.0, Abacus Concepts), and correlation P values were generated by the correlation Z test (StatView).
|
10.1371/journal.pntd.0003080 | Cytokine Responses to Schistosoma mansoni and Schistosoma haematobium in Relation to Infection in a Co-endemic Focus in Northern Senegal | In Africa, many areas are co-endemic for the two major Schistosoma species, S. mansoni and S. haematobium. Epidemiological studies have suggested that host immunological factors may play an important role in co-endemic areas. As yet, little is known about differences in host immune responses and possible immunological interactions between S. mansoni and S. haematobium in humans. The aim of this study was to analyze host cytokine responses to antigens from either species in a population from a co-endemic focus, and relate these to S. mansoni and S. haematobium infection.
Whole blood cytokine responses were investigated in a population in the north of Senegal (n = 200). Blood was stimulated for 72 h with schistosomal egg and adult worm antigens of either Schistosoma species. IL-10, IL-5, IFN-γ, TNF-α, and IL-2 production was determined in culture supernatants. A multivariate (i.e. multi-response) approach was used to allow a joint analysis of all cytokines in relation to Schistosoma infection.
Schistosoma haematobium egg and worm antigens induced higher cytokine production, suggesting that S. haematobium may be more immunogenic than S. mansoni. However, both infections were strongly associated with similar, modified Th2 cytokine profiles.
This study is the first to compare S. mansoni and S. haematobium cytokine responses in one population residing in a co-endemic area. These findings are in line with previous epidemiological studies that also suggested S. haematobium egg and worm stages to be more immunogenic than those of S. mansoni.
| In the developing world, over 207 million people are infected with blood-dwelling parasitic Schistosoma worms. Schistosoma haematobium and S. mansoni are the most widespread species. In Africa, they often occur together in the same area, with many people carrying both species. Yet, little is known about the differences in immune response that the human host develops against these two species. It is also unknown whether the presence of one species may affect the immune response to the other. We here investigated 200 people from an area in the north of Senegal where both species occur. They were examined for Schistosoma infections, as well as for immune responses to the two species. We observed that both infections were characterized by very similar cytokine responses. However, S. haematobium antigens induced higher levels of cytokines than S. mansoni. This suggests that S. haematobium may give rise to stronger immune responses, and may help to explain differences between the two most important Schistosoma species regarding the occurrence of infection and morbidity.
| Schistosomiasis is a parasitic disease of major public health importance. Schistosoma mansoni and S. haematobium are the main human species. Both species are endemic in Africa, where their distributions show a great overlap [1]. Schistosomes are known to down-regulate host immune responses and to induce so-called modified Th2 responses. The exact phenotype of the induced response depends on a complex immunological ‘dialogue’ that involves cytokines and immune cells of Th2, but also Th1, Th17 and regulatory components of the immune system [2].
So far, little is known about differences in host immune responses to schistosomes and possible immunological interactions between S. mansoni and S. haematobium in humans. Yet, epidemiological studies have suggested that host immunological factors may play an important role in co-endemic areas. Interspecies differences in immunogenicity for example, may explain why infection-age curves and morbidity patterns differ between S. mansoni and S. haematobium. Also, immunological interspecies differences and/or immunological interactions between S. mansoni and S. haematobium may explain differences in morbidity levels between single and mixed Schistosoma infections. Cheever et al. reported a more pronounced reduction of S. haematobium than S. mansoni worm loads with age [3]. Similarly, in a mixed focus in northern Senegal, we found the age-infection curve of S. haematobium to decline more steeply after adolescence than that of S. mansoni [4], indicating that protective immunity against S. haematobium may develop more rapidly. In addition, we found that mixed S. mansoni and S. haematobium infection as compared with single S. haematobium infection tended to decrease the risk of S. haematobium-specific urinary tract pathology [5]. This appeared mainly due to ectopically excreted, possible hybrid eggs [6]. Others also found S. mansoni to affect S. haematobium-specific morbidity and vice versa [7], [8], indicating that the two infections may have different effects on the egg-induced immune responses that provoke morbidity.
The present study set out to compare Schistosoma-specific cytokine responses induced by S. mansoni and S. haematobium antigens, and to relate these to Schistosoma infection in a S. mansoni and S. haematobium co-endemic area. Schistosoma infection status (single and mixed) and infection intensities as well as Schistosoma-specific cytokine responses were determined in residents from a co-endemic focus in northern Senegal. A multivariate (i.e. multi-response) approach was used to allow a joint analysis of multiple cytokine responses (interleukin (IL)-10, IL-5, interferon (IFN)-γ, tumor necrosis factor (TNF)-α, and IL-2) [9].
This study was part of a larger investigation on the epidemiology of schistosomiasis and innate immune responses (SCHISTOINIR) for which approval was obtained from the review board of the Institute of Tropical Medicine, the ethical committee of the Antwerp University Hospital and ‘Le Comité National d'Ethique de la Recherche en Santé’ in Dakar. Informed and written consent was obtained from all participants prior to inclusion into the study. For minors, informed and written consent was obtained from their legal guardians.
All community members were offered praziquantel (40 mg/kg) and mebendazole (500 mg) treatment after the study according to WHO guidelines [10].
This study was conducted in Ndieumeul and Diokhor Tack, two neighboring communities on the Nouk Pomo peninsula in Lake Guiers. Details on the study area have been described elsewhere [4], [5]. Between July 2009 and March 2010, parasitological data were collected from 857 individuals [4]. A random subsample of 200 subjects was followed up immunologically. These subjects were between 5 and 53 years of age. Individuals who had lived in an urban area in the 5 years preceding the study (n = 7), had taken praziquantel within the last year (n = 2), or had clinical signs of malaria (recruited upon recovery), and pregnant women (n = 18) were excluded from the immunological study.
Two feces and two urine samples were collected from each participant on consecutive days. Infection with Schistosoma spp. was determined quantitatively (by Kato-Katz and urine filtration), and infection with soil-transmitted helminths (STHs) Ascaris lumbricoides, Trichuris trichiura and hookworm, was assessed qualitatively (by Kato-Katz), as described elsewhere [4]. Aliquots of the first fecal samples were preserved in ethanol to confirm microscopy results by multiplex PCR (A. lumbricoides, hookworm and Strongyloides stercoralis) (n = 198) [11]. Infection with Plasmodium was determined by Giemsa-stained thick blood smears.
Five hours after venipuncture, heparinized blood was diluted 1∶4 in RPMI 1640 (Invitrogen) supplemented with 100 U/ml penicillin, 100 µg/ml streptomycin, 1 mM pyruvate and 2 mM glutamate (all from Sigma). This mixture (200 µl sample volume) was incubated in 96-well round bottom plates (Nunc) at 37°C under 5% CO2 atmosphere for 72 h, together with one of four schistosomal water-soluble antigen preparations at a final concentration of 10 µg protein/ml:
Medium (see above) without stimulus was used as a negative control. After harvesting, supernatants were stored at −80°C. Schistosoma eggs and adult worms were isolated from either S. mansoni- or S. haematobium-infected golden hamsters. SEAm, SEAh, AWAm and AWAh were prepared from this material using identical procedures. In brief, eggs or worms were freeze-dried and then homogenized in phosphate-buffered saline (PBS) with 10% n-octyl-β-D-glucopyranoside. Subsequently, this mixture was sonicated, frozen, thawed and washed with PBS. The resulting pellet was dialyzed and filter-sterilized. While AWAm and AWAh batches were lipopolysaccharide (LPS)-free, SEAm and SEAh antigens contained equivalent amounts of LPS (final concentrations of 1–5 ng/ml).
IL-10, IL-5, IFN-γ, TNF-α, and IL-2 in culture supernatants were analyzed simultaneously using custom Luminex cytokine kits (Invitrogen) according to the manufacturer's instructions. Samples with concentrations below the detection limit were assigned values corresponding to half of the lowest value detected. Lowest values detected were 0.063 pg/ml for IL-10, 0.044 pg/ml for IL-5, 0.090 pg/ml for IFN-γ, 0.051 pg/ml for TNF-α, and 0.063 pg/ml for IL-2.
Results were considered significant when the p-value was <0.05. The Pearson Chi-square test was used to determine the association between infection status on the one hand, and age and gender on the other. Nonparametric techniques were chosen because cytokine concentrations were not normally distributed. Univariate statistics were used to compare single antigen-induced responses within individuals (IBM SPSS 21.0). McNemar's tests were used to compare cytokine response frequencies between S. mansoni and S. haematobium antigen-induced responses within individuals (e.g. SEAm- versus SEAh-induced responses). Similarly, Wilcoxon Signed Rank tests were used to compare cytokine response levels between S. mansoni and S. haematobium antigen-induced responses within individuals. Multivariate (i.e. multi-response) statistics were used to collectively analyze multiple cytokine responses – i.e. cytokine profiles - in the study population, and to investigate interrelationships between these responses [9]. We chose the nonparametric technique nonmetric multidimensional scaling (nMDS; in R with the ‘Vegan’ package [12], [13]). This is a variant of the parametric principal component analysis (PCA), but with fewer assumptions about the nature of the data and the interrelationship of the variables [14]. This is important because cytokine response levels were not normally distributed, even after log-transformation. Also, levels of different cytokines typically correlate with one another. Upon computation of the cytokine profiles, associations between these cytokine profiles and Schistosoma infection were assessed. The approach is illustrated in Supporting Information S1. Before nMDS, cytokine concentrations in the negative control were subtracted from those in antigen-stimulated samples to obtain net cytokine responses. Negative values were set to zero. Net cytokine responses were normalized by log(base 10)-transformation after adding 1 pg/ml to allow for zeroes. Schistosoma infection intensities were normalized after adding half of the detection limit (i.e. 5 eggs per gram of feces and 0.5 eggs per 10 ml of urine for S. mansoni and S. haematobium, respectively). One nMDS was performed for each of the four Schistosoma-specific whole blood stimulations (either SEAm, SEAh, AWAm or AWAh) using the ‘metaMDS’ function [13]. Each nMDS was repeated several times to assess the robustness of the resulting pattern [14]. The Euclidean dissimilarity index was used [13], and cytokine profiles - i.e. the matrix of IL-10, IL-5, IFN-γ, TNF-α, and IL-2 - were plotted in three dimensions (3D) to adequately represent the variation in the data [14]. Afterwards, gradients of the separate cytokine responses, on which the nMDS was based, were fitted using the ‘envfit’ function [13]. The same function was used to fit infection intensities onto each 3D nMDS, and to statistically test associations of antigen-induced cytokine profiles with Schistosoma infection intensity or infection status, i.e. uninfected, single S. mansoni, single S. haematobium, versus mixed S. mansoni and S. haematobium infection. The ‘ordiellipse’ function was used to fit average group scores - with their 95% confidence intervals (CIs) - for different infection statuses [13]. In contrast to individual S. mansoni- and S. haematobium-induced cytokine responses which can be compared quantitatively within individuals as described above (univariate statistics), qualitative differences between S. mansoni- and S. haematobium-induced cytokine profiles could only be assessed visually by nMDS, not by formal statistical testing.
The study population consisted of 88 males and 112 females with a median age of 16 (range 5–53) years. Malaria and STHs T. trichiura and hookworm were absent in this population, and A. lumbricoides and S. stercoralis rare (n = 3 and 2, respectively, with 100% concordance between microscopy and PCR). In contrast, 137 (69%) subjects were infected with S. mansoni, and 116 (58%) with S. haematobium. Sixty percent (95/158) of all Schistosoma infections were mixed S. mansoni and S. haematobium infections (Table 1). The distributions of S. mansoni and S. haematobium infections in the study population according to age and gender are shown in Table 2. Both Schistosoma infections peaked in adolescents (10 to 19 year-olds), but gender differences were not statistically significant. Epidemiological patterns of infection have been described in more detail elsewhere [4].
Insight into the different antigen-induced cytokine responses relative to one another was obtained by nMDS. Figure 1 and 2 show the variation in multivariate cytokine responses in the study population, with dots representing individuals. Distances between dots approximate inter-individual dissimilarities in cytokine responses with stress values (i.e. discrepancies) of 0.051 for SEAm, 0.041 for SEAh, 0.058 for AWAm, and 0.061 for AWAh. Red arrows indicate increasing gradients of IL-10, IL-5, IFN-γ, TNF-α and IL-2 responses, respectively. The level of a cytokine response increases in the direction of the corresponding arrow (see also Supporting Information S1). The length of a cytokine arrow indicates the goodness of fit of that arrow (or cytokine gradient).
The nMDS outcomes for the first axis (nMDS1) show that for each of the four antigen stimulations, all cytokine responses point to the left. Individuals plotted on the left produced consistently higher levels of all cytokines measured than those on the right. In other words, nMDS1 indicates a gradient of high (left) to low (right) cytokine responses. In analogy, the second axis (nMDS2), indicates a gradient of Th1-like (IFN-γ and TNF-α, top) to Th2-like (IL-5, bottom) phenotypes for each of the antigen stimulations. In contrast to SEA-induced IL-5, AWA-induced IL-5 was not accompanied by production of IL-10. IL-2 levels increased with Th1 cytokines, except for SEAm. The third axis (nMDS3) indicates a gradient of TNF-α and IL-2 (left) to IFN-γ and IL-10 (right).
In contrast to antigen-induced cytokines, spontaneously induced levels of cytokines in the control (medium only), did not show significant gradients, except for IL-5 on the third nMDS axis (stress = 0.11, data not shown).
Figure 1 and 2 indicate that S. mansoni and S. haematobium antigens induced very similar cytokine profiles; cytokine profiles differed more between adult (AWA) and egg (SEA) life stages of the parasite than between the two Schistosoma species. Within individuals, S. haematobium-induced cytokine response levels were higher than those induced by S. mansoni (Table 3). This was statistically significant for all SEA- and AWA-induced cytokine responses that were measured, except for SEA-induced IFN-γ and IL-10.
Subsequently, we related the above-described variation in cytokine responses in the study population (i.e. plotted cytokine profiles) to infection intensity. Table 4 shows that all associations between Schistosoma antigen-induced cytokine profiles and infection intensity were statistically significant. In Figure 1, the direction of the black arrows represents the increasing gradients of S. mansoni and S. haematobium infection intensity, respectively (see also Supporting Information S1). On the first axis, which indicates cytokine response levels (see above), these arrows generally point into the opposite direction of cytokine responses. This indicates that people with elevated Schistosoma infection intensities are more likely to have lower cytokine responses, and vice versa. On the second axis, which indicates the Th1 versus Th2 response phenotype (see above), infection intensity generally increases with IL-5 and decreases with Th1 cytokines TNF-α, IFN-γ, and IL-2 (except for SEAm-induced IL-5 which decreases with increasing infection intensity). Briefly, as infection intensity increased, cytokine response levels decreased and the Th2 phenotype became more pronounced. The association between infection intensity and reduced cytokine responsiveness was more pronounced for SEA than for AWA stimulation. Schistosoma infection intensity increased with AWA-induced IL-5, but decreased with SEA-induced IL-5 levels, indicating that people with higher infection intensities produced more of a Th2-like response against AWA and more of a suppressive response (i.e. with low cytokine response levels) against SEA than people with lower infection intensities, and vice versa.
We did not observe differences in induced cytokine profiles between the two Schistosoma infections. Associations between cytokine profiles and infection intensity were comparable for S. mansoni and S. haematobium infections (Figure 1). Table 4 shows significant correlations between cytokine profiles and Schistosoma infection intensity for homologous combinations (i.e. infection intensity and antigen stimulation of the same species) as well as for heterologous combinations (i.e. infection intensity of one and antigen stimulation of the other species).
Schistosoma antigen-induced cytokine profiles were significantly associated with Schistosoma infection status, except upon stimulation with AWAm (Table 4). Figure 2 shows how antigen-induced cytokine profiles differed according to infection status (except for AWAm, which was not significantly associated with infection status), with 95% CI ellipsoids indicating the average nMDS scores per infection group: uninfected (‘N’), single S. mansoni (‘M’), single S. haematobium (‘H’), versus mixed (‘MH’) Schistosoma infection group. In analogy with Figure 1, uninfected individuals had higher cytokine responses than Schistosoma-infected subjects, and their cytokine profiles were skewed more towards the Th1 phenotype. On the whole, there was a gradient in cytokine profiles from uninfected individuals, to people with single and then mixed Schistosoma infections (Figure 2) and these profiles were in the same direction as the gradient of infection intensity (Figure 1). In other words, people with low cytokine responses of the Th2 phenotype tended to have both mixed and heavier infections, people with strong Th1 responses tended to be uninfected, and those with an intermediate cytokine profile tended to have both single and lighter Schistosoma infections.
For the SEAm-induced cytokine profile, there was a clear difference (i.e. separation between ellipsoids) between S. mansoni-infected individuals (with either single or mixed S. mansoni), and those without S. mansoni (no Schistosoma infection, or single S. haematobium infection; Figure 2A). There were no significant differences in this cytokine profile between single and mixed S. mansoni infections, or between uninfected individuals and those with single S. haematobium infections. This indicates that, in contrast to S. mansoni, S. haematobium infection status was not associated with SEAm-induced cytokine profiles. Schistosoma haematobium-induced cytokine profiles on the other hand, showed similar relationships with S. mansoni as well as with S. haematobium infection status. Cytokine profiles of people with single and mixed infections differed significantly from those of uninfected people, and cytokine profiles did not appear to differ between single S. mansoni and single S. haematobium infections.
The objective of this study was to compare cytokine responses induced by S. mansoni and S. haematobium antigens, and to relate these to Schistosoma infection in a S. mansoni and S. haematobium co-endemic area. We showed that Schistosoma infection intensity was significantly associated with Schistosoma antigen-induced cytokine profiles and that it may explain up to 18% of the variation in cytokine responses observed in this population. As Schistosoma infection intensity increased, cytokine responses decreased and the Th2 phenotype became more pronounced. This was exemplified by relatively higher IL-5 (and IL-10) and relatively lower IFN-γ, TNF-α and IL-2 levels. Lightly infected and uninfected subjects on the other hand, had elevated cytokine responses, with a Th1 phenotype. These patterns are consistent with the modified Th2 response characteristic for schistosomiasis [2]. nMDS also indicated that the association between infection and the Th2 phenotype was more pronounced for AWA, while that between infection and (reduced) cytokine responsiveness was more pronounced for SEA. These observations fit with a previous study by Joseph et al. describing similar immunological differences between Schistosoma adult worm and egg life stages in a population from a S. mansoni mono-endemic area, using more conventional analyses [15].
Secondly, we demonstrated that increased Schistosoma infection intensity and mixed (as compared to single) infections were associated with similar, modified Th2, cytokine profiles. This is probably due to the fact that subjects with mixed infections were more likely to have higher infection intensities than those with single infections [4]. Also, similar, modified Th2, cytokine profiles were observed for both S. mansoni and S. haematobium infection intensity, whether blood was stimulated with antigens from the homo- or heterologous species. This may be indicative of immunological cross-reactivity between species. For S. mansoni-induced cytokine profiles however, this was unlikely, because profiles did not differ between single and mixed S. mansoni infection groups. While S. haematobium-induced cytokine profiles did differ between single and mixed S. haematobium infection groups, we could not determine whether these differences were due to mixed infection per se, or to higher S. haematobium infection intensity in mixed as compared to single infections. Other potentially confounding factors such as age may have been involved as well [4], and future studies should be performed to assess their respective roles in determining cytokine responses. To obtain more evidence on the existence of cross-reactivity between the two major human Schistosoma species, it is important to compare immune responses between different co- and mono-endemic areas, using different immunological parameters (e.g. cytokine, humoral and cytological data). To our knowledge, only one human study reported on functional S. mansoni – S. haematobium cross-reactivity. This study from 1974 reported lethal in vitro activity of sera from subjects infected with one species against schistosomula of the same but not of the other species [16]. Indeed, S. mansoni and S. haematobium may share few if any epitopes that are involved in protective immunity because they belong to genetically distinct groups. Potential cross-reactivity or the lack thereof merits further investigation as this may have important implications for our understanding of the epidemiology of schistosomiasis as well as for the development of an effective schistosomiasis vaccine.
The present study demonstrated that nMDS can be used successfully to analyze host cytokine responses collectively. In this way, it was possible to analyze cytokine responses in relation to one another, and in relation to Schistosoma infection. nMDS is a nonparametric, multivariate and visual method. It is a robust and powerful tool because it avoids problems of multiple statistical tests and violations of data assumptions [14]. Moreover, nMDS makes it easier to interpret complex data than traditional one-by-one graphs, tables, and tests. Here, we used this approach to study multivariate cytokine responses, but it can be used equally well to increase our understanding of other complex, multidimensional data, such as cytological and/or serological data (Durnez et al, unpublished data), as well as infection data on multiple co-endemic parasite species.
Additional analyses showed that, within individuals, S. haematobium antigens induced higher cytokine responses in 72 h whole blood cultures than those of S. mansoni. A very similar pattern was observed in parallel investigations in Ghana, in a population which was - in contrast to the Senegalese study population - first exposed to S. haematobium and then to both S. mansoni and S. haematobium, and with lower prevalences of S. mansoni and higher prevalences of S. haematobium (unpublished data, A.S. Amoah et al, and ref [4]). This suggests that this finding does not depend on the level of transmission or on exposure history, and that the two Schistosoma species may differ in their immunogenicity. This hypothesis is in line with observations from Van Remoortere et al. who found S. mansoni to induce mainly IgM antibodies – which are thought to inhibit protective host immune responses [17] – while S. haematobium induced both IgM and IgG antibodies against shared carbohydrate epitopes [18]. It is therefore tempting to speculate that lower cytokine response levels may prevent Ig class switching from IgM to IgG for these epitopes in S. mansoni infection, while stronger cytokine responses may promote class switching in S. haematobium infection. Alternatively, differences in their biochemical composition may underlie interspecies differences in both immunogenicity and humoral immune responses. These two immunological interspecies differences may also have contributed to earlier epidemiological findings. Several studies observed a steeper decline of the age-infection curve of S. haematobium as compared to S. mansoni after adolescence, indicating that protective immunity against S. haematobium might develop more rapidly [3], [4]. Secondly, higher levels of S. haematobium- as compared to S. mansoni-specific morbidity have been observed in co-endemic populations [5], [7], [8], suggesting that the immune responses provoked by S. haematobium eggs might be more pathogenic. It should be noted however, that other factors may also explain these two epidemiological observations. For example, S. mansoni and S. haematobium eggs accumulate in different organs, i.e. the liver and the urinary tract, respectively, and these differences in anatomical context may also explain the differences in the extent of morbidity between the two species. More research is necessary to investigate the abovementioned immunological interspecies differences and their implications for epidemiological patterns of infection and morbidity in more detail.
In conclusion, this is the first study to comprehensively investigate S. mansoni- and S. haematobium-induced cytokine responses in a S. mansoni and S. haematobium co-endemic area, and to relate these cytokine responses to Schistosoma infection. The present study demonstrates that nMDS can be used successfully as a tool for the joint analysis of multiple cytokine responses in relation to Schistosoma infection. We showed strong associations between Schistosoma infection and Schistosoma-induced cytokine profiles, and provided a first insight into potential differences and interactions between human S. mansoni and S. haematobium infections. This knowledge will contribute to an improved understanding of the mechanisms underlying Schistosoma infection and morbidity in co-endemic populations.
|
10.1371/journal.pgen.1003841 | Whole Genome Sequencing Identifies a Deletion in Protein Phosphatase 2A That Affects Its Stability and Localization in Chlamydomonas reinhardtii | Whole genome sequencing is a powerful tool in the discovery of single nucleotide polymorphisms (SNPs) and small insertions/deletions (indels) among mutant strains, which simplifies forward genetics approaches. However, identification of the causative mutation among a large number of non-causative SNPs in a mutant strain remains a big challenge. In the unicellular biflagellate green alga Chlamydomonas reinhardtii, we generated a SNP/indel library that contains over 2 million polymorphisms from four wild-type strains, one highly polymorphic strain that is frequently used in meiotic mapping, ten mutant strains that have flagellar assembly or motility defects, and one mutant strain, imp3, which has a mating defect. A comparison of polymorphisms in the imp3 strain and the other 15 strains allowed us to identify a deletion of the last three amino acids, Y313F314L315, in a protein phosphatase 2A catalytic subunit (PP2A3) in the imp3 strain. Introduction of a wild-type HA-tagged PP2A3 rescues the mutant phenotype, but mutant HA-PP2A3 at Y313 or L315 fail to rescue. Our immunoprecipitation results indicate that the Y313, L315, or YFLΔ mutations do not affect the binding of PP2A3 to the scaffold subunit, PP2A-2r. In contrast, the Y313, L315, or YFLΔ mutations affect both the stability and the localization of PP2A3. The PP2A3 protein is less abundant in these mutants and fails to accumulate in the basal body area as observed in transformants with either wild-type HA-PP2A3 or a HA-PP2A3 with a V310T change. The accumulation of HA-PP2A3 in the basal body region disappears in mated dikaryons, which suggests that the localization of PP2A3 may be essential to the mating process. Overall, our results demonstrate that the terminal YFL tail of PP2A3 is important in the regulation on Chlamydomonas mating.
| Whole genome sequencing is a powerful tool to detect changes in genomic DNA. However, how to identify a causative mutation from over 20,000 changes remains a big challenge. For the unicellular green alga Chlamydomonas, we built a library that consists of over 2 million changes from 16 strains. A comparison of changes found in a mutant strain with a mating defect, imp3, to 15 other strains, leads to the identification of a three amino acid deletion in the catalytic subunit of a protein phosphatase 2A (PP2A3). The mating defect of imp3 is rescued by an HA-tagged PP2A3 gene. Introduction of the HA-tagged PP2A3 gene with various mutations in these three amino acids reveals that they play a key role in stabilizing and ensuring the proper localization of PP2A3. The ubiquitous enzyme PP2A is involved in diverse cellular processes. Our discovery that PP2A3 is involved in the Chlamydomonas mating signaling pathway, which also contains the polycystin2 homolog (PKD2), makes Chlamydomonas mating an excellent model to study ciliary/flagellar signaling. Since both PP2A and PKD2 play important roles in human health, further investigation of the relationship between these two proteins in Chlamydomonas will facilitate better understanding of their functions.
| Forward genetics allows the identification of mutants with phenotypes of interest and the mechanistic understanding of biological processes [1], [2]. While gene lesions generated by insertional mutagenesis can be identified by Southern blot analysis [3]–[6] or PCR-based approaches [7]–[9], identification of mutations induced by radiation or chemical mutagenesis rely on time-consuming meiotic mapping [10]–[14]. Recently, single nucleotide polymorphism (SNP) discovery by whole genome sequencing (WGS) provides a faster and more efficient method to identify causative mutations [15]–[18]. However, in model organisms such as Arabidopsis thaliana, the number of SNPs can vary from 2,000 to 900,000, depending on the strain background [19]. In Caenorhabditis elegans, the number of SNPs between two strains is ∼100,000 [16]. In the unicellular biflagellate green alga Chlamydomonas reinhardtii, ∼38,000 SNPs were identified in individual mutant strains [18]. Therefore, identification of the causative SNP from a large number of SNPs remains a challenge.
In Chlamydomonas, a model organism for the study of flagellar function, photosynthesis, biofuels, and sex determination, the causative genes in several hundred mutant strains generated by radiation or chemical mutagenesis remain unidentified [20]. UV mutagenesis mutant screens generated 12 impotent (imp) mutant strains that have either abolished or reduced mating efficiency [21], [22]. In Chlamydomonas, the sex of a cell is controlled by two alleles, plus or minus, at the mating-type (MT) locus [23]. The differentiation from exponentially growing vegetative cells to gametes is triggered by nitrogen starvation via an unknown mechanism. When gametes of the opposite mating-types are mixed together, they agglutinate via flagellar membrane-associated proteins, agglutinins, to trigger a mating signal transduction pathway. This signal cascade leads to cell fusion and the formation of zygotes. Among the previously characterized imp mutant strains, imp2, imp5, imp6, imp7, and imp9 are allelic and encode SAG1, which is the plus agglutinin [24]. The imp10 and imp12 mutant strains encode SAD1, the minus agglutinin [24]. The imp8 strain is defective in O-glycosylation and is allelic with the GAG1 locus [25]. The imp1 and imp11 mutant strains map within the MT locus [26]–[29] and carry mutations in FUS1 in plus and MID1 in minus cells, respectively. Only imp3 and imp4 remain unidentified among the original collection of impotent mutants. Unlike the other imp strains that abolish mating (<1%), the mating efficiency of imp3 and imp4 strains varies from 10% to 50%, in contrast to >80% in wild-type cells 1 hour after mixing of the gametes. Neither mutation is linked to the MT locus or to the other [21]. Saito et al. [30] suggested that activation of uncharacterized flagellar adenylate cyclases is blocked in imp3 cells, and IMP3 is required in the mating signaling pathway. However, the causative genes in imp3 and imp4 remain unidentified due to their partial, weak mating phenotype and the difficulty to map this phenotype.
Serine/threonine phosphorylation is generated by 300–400 kinases but is reversed by a relative small number of phosphatases. The serine/threonine phosphatase, protein phosphatase 2A (PP2A), plays an important role in signaling pathways. It is a ubiquitous enzyme that is involved in diverse cellular processes; they include cell cycle control, cell growth, microtubule stability, and signaling [31]. The PP2A heterotrimeric holoenzyme contains 3 subunits; they are a catalytic subunit (C subunit), a scaffold subunit (A subunit), and a regulatory subunit (B subunit). The catalytic subunit (PP2Ac) is highly conserved across species and it shares significant sequence similarity to the PP4 and PP6 protein phosphatases [32]. The catalytic activity of PP2Ac can be modulated by post-translational modifications that include phosphorylation/dephosphorylation in the conserved C-terminal motif T304PDY307FL309 on T304 and Y307 and methylation/demethylation of L309 [33]. The scaffold subunit of PP2A contains multiple HEAT repeats that confer conformational flexibility to both the catalytic subunit and the regulatory subunit [32]. The regulatory subunit of PP2A falls into four distinct families, which are known as B (PR55), B′ (B56 or PR61), B″ (PR72), and B′″ (PR93/PR110). It is believed that different families of the B subunit target PP2A to different cellular locations and bind to different substrates [32], [33]. In Chlamydomonas, sequence similarity reveals four catalytic subunits, two scaffold subunits, and five regulatory subunits [34].
In this study, we took advantage of whole genome sequencing of 16 different wild-type and mutant strains to generate a SNP/indel library. SNP/indel comparison, in conjunction with meiotic mapping, allowed the quick identification of the causative mutation in the imp3 mutant strain, which contains a C-terminal three amino acid deletion in the conserved TPDYFL motif of a PP2A catalytic subunit (PP2A3). The deletion of YFL affects not only the stability of PP2A3, but also the accumulation of PP2A3 around the basal body area.
A previous study using 101-bp paired-end Illumina sequencing of an IFT80 mutant strain, NG30/ift80, revealed that over 38,000 SNPs/indels are present compared to the Chlamydomonas reference genome [18]. It was a challenge to identify the causative mutation from such a large number of changes. We reasoned that if changes are found in other unlinked mutant strains or in wild-type strains, they are not causal and could be eliminated from further analysis. Therefore, a collection of changes from multiple strains would be necessary to remove as many non-causative changes as possible to reveal the causative mutation in a given mutant strain.
To build a library of changes, we first sequenced four wild-type strains (CC-124, CC-125, isoloP (CC-4402), and isoloM (CC-4403)). In Chlamydomonas, the major laboratory strains CC-124 and CC-125 were first isolated from a single diploid zygote, 137c, in 1945 [35] and these strains have been used as the parents in many mutant isolations. CC-125 is also the background strain of CC-503, the strain used for the Chlamydomonas reference genome assembly. CC-124 carries a minus mating-type locus (mt−) and CC-125 is mating-type plus (mt+). The other two wild-type strains, isoloP (mt+) and isoloM (mt−), were generated by crossing CC-124 by CC-125 to obtain meiotic progeny. Several progeny from this cross that gave the fastest and highest percentage of mating were backcrossed to CC-124. This procedure was repeated ten times in an attempt to obtain isogenic strains that were named isoloP and isoloM. Sequencing of these wild-type strains identifies 13,000 to over 100,000 changes relative to the reference genome. We also sequenced is a highly polymorphic strain S1C5 (CC-1952) that is frequently used in meiotic mapping with molecular markers. Over 2 million changes are identified from this strain. In addition, since we were interested in identifying the causative mutations from a variety of mutant strains, we also performed whole genome sequencing on ten mutant strains that were generated by either chemical or UV-mutagenesis. Five of them (fla18, fla24, fla9, uni1, ift80) have flagellar assembly defects, four (ida3, pf23, pf7, pf8) have motility defects, and one has a mating defect (imp3). One additional mutant strain, cnk10, which has a flagellar assembly defect, was generated by insertional mutagenesis of the CC-125 strain (Lin and Dutcher, unpublished). The number of changes in individual mutant strains varies from 22,000 to over 150,000 (Table 1). The sequencing coverage of individual strains ranges from 26× to 162× (Table 1). Overall, 2,557,197 changes are included in this SNP/indel library and it is available at http://stormo.wustl.edu/SNPlibrary/.
After collecting the SNPs/indels, we analyzed the distribution of the changes across the 17 chromosomes relative to the reference genome [36] (Figure 1). In the wild-type strain CC-125 (mt+), changes are spread evenly across all chromosomes from the reference genome (Figure 1A, Table 2). The wild-type strain CC-124 (mt−), which came from the same zygote as CC-125, contains 100,737 changes, which is about eight times the number of changes (13,218) found in CC-125. Around 90% of changes found in CC-124 are concentrated on five chromosomes: 3, 6, 12, 16, and 17 (Figure 1A, Table 2). A detailed analysis of numbers of SNPs/indels every 100 kb along the chromosomes in CC-124 reveals that the polymorphisms are not distributed evenly across these five chromosomes (Figure 1B). On chromosome 3, most SNPs/indels are between 8.5 Mb and 9.1 Mb. On chromosome 6, most SNPs/indels are within the first 1.9 Mb, which contains the MT locus. The mating-type locus is known to be polymorphic between the two sexes [37]. CC-124 carries the MT minus locus, which is not shared with the reference strain CC-503. On chromosome 12, most changes are observed between 9.0 Mb to 9.8 Mb. On chromosome 16, three distinct regions of SNPS/indels lie between 0.9–1.0 Mb, 1.5–2.0 Mb, and 6.4–7.8 Mb. A large number of changes within the 0.9–1.0 Mb were observed in a previous study of Chlamydomonas strains, ift80 and ac17 [18]. On chromosome 17, most changes are observed between 0.3 Mb to 1.5 Mb. The other two wild-type strains, isoloP (mt+) and isoloM (mt−), which are meiotic progeny of CC-124 and CC-125 after ten rounds of backcrosses to CC-124, were expected to show difference only on chromosome 6, which contains the MT locus. However, comparison of the sequence of these two strains indicates that they are not isogenic on chromosomes 3 and 17. The isoloM strain maintains large numbers of changes from the CC-124 parent on chromosome 3; isoloP contains large numbers of changes inherited from the CC-124 parent on chromosome 17 (Figure 1A, Table 2).
We performed the same analysis on the ten mutant strains, with the exclusion of fla18, because the sequenced strain came from a cross between the fla18 mutant strain and S1C5. Ninety-six percent of fla18 SNPs/indels are found in the S1C5 strain. Prior to Illumina sequencing, the ten strains were crossed to either CC-124 or CC-125 at least once. An accumulation of changes on chromosomes 3, 6, 12, 16, and 17 are observed in most strains (Figure 1C). To ask whether these changes are the same as found in CC-124, we subtracted SNPs/indels found in CC-124 from individual strains. The numbers of changes drop dramatically from 10,000∼35,000 to 1,000∼5,000 in almost all strains (Figure 1D). This suggests that changes in these strains are likely come from the CC-124 parent. In comparison, removal of CC-125 changes from these strains does not have an obvious effect on numbers of changes in these strains (Figure 1E). There is no correlation with the position of the causative mutant and an accumulation of SNPs along the chromosomes. One mutant strain, uni1, has a distinct distribution of changes along all chromosomes when compared to other strains (Figure 1C). Accumulation of changes is found on chromosomes 1, 2, 11, and 16. Removal of CC-124 or CC-125 changes from uni1 has no significant effect on the distribution (Figure 1D and 1E). A comparison between 30,958 changes on chromosome 16 found in CC-124 and 56,192 changes on chromosome 16 found in uni1 indicates that only 8,229 changes are common between these two strains. Thus, the SNPs/indels found in uni1 have significantly different distribution than all other strains we analyzed. This mutant strain was generated from either strain 89 or 90 (CC-1009 or CC-1010) [38]. Pröschold et al. [39] categorized the common used laboratory Chlamydomonas strains into three basic sublines based on several criteria, including their ability to utilize nitrate, the mating-type locus, the number of rDNA repeats, and the presence of cell wall digestion metalloproteases (autolysins). Strains 89 and 90 belong to Subline II and CC-124 and CC-125 belong to Subline III. They have difference in all criteria described above. Thus it is not surprising to observe the difference of SNP/indel accumulations between uni1 and 137c strains.
The 162× sequencing coverage of imp3 reveal 91,066 changes in this mutant strain, which came from a 137c background. Comparison and subtraction between imp3 and all other 15 strains finds 7,092 (7%) of these changes are unique to imp3 (Table 1). Out of 954 changes in predicted exons, 145 are predicted to be synonymous changes and were excluded from further analysis. Changes vary from 1 to 297 on individual chromosomes (Table 3). In order to identify the causative mutation, we meiotically mapped the imp3 mutant strain.
The imp3 phenotype confers reduced mating efficiency, a phenotype that is challenging to analyze quickly. When wild-type gametes mate, the zygotes develop thick cell walls and form a dark green, multi-layered sheet of cells called the pellicle (Figure 2A). The mating between imp3 mt+ and imp3 mt− do not form the thick pellicle sheet observed in wild-type cells or in mating between imp3 and wild-type cells (Figure 2A). Observation under the dissecting microscope reveals dark thick multi-layer pellicle between imp3 and wild-type gametes (Figure 2B). Mating between two imp3 strains produces a single cell-layer pellicle that is light in color (Figure 2C and 2D). Individual progeny from 30 complete tetrads from a cross between imp3 and wild-type cells were tested for this phenotype by mating with imp3 mt+ and imp3 mt− tester strains. All 30 tetrads showed 2∶2 segregation of the single layer pellicle phenotype. This phenotype made it easy to distinguish between imp3 and IMP3 cells; this assay facilitated the molecular mapping of the IMP3 locus.
To map the imp3 mutation, meiotic progeny from imp3 crossed by the highly polymorphic strain S1C5 were obtained. Progeny from twenty tetrads that showed the single layer pellicle phenotype were mapped with dCAPS markers (Table S1). The single layer pellicle phenotype showed very tight linkage (18 parental: 0 recombinant) to the SCA8-2 marker, which maps to ∼6.6 Mb on chromosome 2 in a previous version of the genome assembly (v4) and maps to ∼3.5 Mb on chromosome 9 in the latest version (v5.3) of the genome assembly. To fine map the imp3 mutation, additional progeny from over 100 tetrads were used for mapping. All three markers, 2-98, SCA8-2, and 55-193, which are about 17, 8, and 3 cM away from the imp3 mutant, map to chromosome 2 in v4 genome assembly (Table S1). Both 2-98 and 55-193 are at ∼6.6 Mb and ∼7.1 Mb on chromosome 2 in v5.3 genome assembly, respectively (Table S1). Thus, we believe that genomic DNA with an unknown length around the SCA8-2 marker is misassembled in v5.3 genome assembly to chromosome 9 and the imp3 mutation maps to chromosome 2. However, since we were unclear whether the imp3 mutation is misassembled in v5.3 genome assembly, we examined polymorphisms in predicted exons on both chromosomes 2 and 9. Two 3 nucleotide-insertion changes at ∼2.3 Mb on chromosome 2 in v5.3 genome assembly are found in imp3 (Table 3). Since both changes are over 4 Mb away from the 2-98 and 55-193 markers, they were eliminated from further study. Two changes are found on chromosome 9 in v5.3 genome assembly (Table 3) and they are both within the mapping region of ∼8 cM defined by the SCA8-2 marker. Both changes map to chromosome 2 in the v4 genome assembly. The first SNP change, which maps to position 4,049,338 on chromosome 9, has 5 Illumina sequencing reads and a Phred quality score of 16.9. It is a G to A change that causes a non-synonymous change from R (cGg) to Q (cAg) in a RegA/RlsA-like protein (g9750). The exact same change is observed in an aflagellate mutant cnc1 not included in the SNP/indel library (Dutcher and Nauman, unpublished) and thus is unlikely to be the causative change in the imp3 mutant strain. The second change, which maps to position 3,721,280 on chromosome 9, has 122 reads and a Phred quality score of 214. It is a deletion of 9 nucleotides immediately before the stop codon of a PP2A catalytic protein (PP2A3, g9684) and the deletion is predicted to remove the last 3 amino acids, YFL, which are conserved in almost all type 2A phosphatases (PP2A, PP4, and PP6; Figure 3B, S1, and S2). This change was confirmed by Sanger sequencing.
A previous study using sequence similarity indicated that the Chlamydomonas genome contains four potential PP2A catalytic subunits, PP2A-1c (g4366), PP2A3 (g9684), PP2A-c4 (Cre12.g494900), and PPA1 (Cre06.g308350) [34]. Due to the sequence similarities observed among PP2A, PP4, and PP6 in all organisms [32], we asked whether the four Chlamydomonas proteins are PP2A catalytic subunits using phylogenetic analysis. A phylogenetic tree was built based on 55 PP2A, PP4, and PP6 protein sequences from green algae (Chlamydomonas reinhardtii, Chlorella variabilis, Micromonas, Ostreococcus lucimarinus, Ostreococcus tauri, and Volvox carteri), yeast (Saccharomyces cerevisiae), land plants (Arabidopsis thaliana and Zea mays), invertebrates (Caenorhabditis elegans and Drosophila melanogaster), and mammals (Mus musculus and Homo sapiens) (Table S2, Figure 3A and S1). This phylogenetic tree shows that Chlamydomonas PP2A-1c and PP2A3 are PP2A-like proteins. PP2A-c4 is a PP4-like protein and PPA1 belongs to the PP6 family. Within the PP2A family, four subgroups are distinguished using a bootstrap analysis (Figure 3A). PP2A proteins from invertebrates, mammals, and a green alga Chlorella form subgroup 1 in the PP2A family. The subgroup 2 is composed of PP2A proteins from land plants and yeast. PP2A proteins from land plants and green algae are found in subgroup 3. The subgroup 4, which includes PP2A3, is a green algal-specific subgroup. Protein sequences of all 26 proteins from the PP2A family were aligned. The alignment reveals that while subgroups 1, 2, and 3 have the conserved T304PDYFL309 C-terminus, the proteins in the fourth subgroup do not have the conserved T304, instead, it is replaced with V, C, or M (Figure 3B and S2). This suggests that T304 is not conserved in the green algal-specific subgroup.
In order to demonstrate that the deletion of YFL at the C-terminus of PP2A3 is the causative mutation in imp3, we performed plasmid rescue. The PP2A3 gene contains only a single exon, which is predicted to encode a 315 amino acid protein with a predicted molecular weight of 35,676 daltons. An HA-tagged PP2A3 gene with the epitope tag HA immediately following the start codon to avoid compromising the C-terminus and under the regulation of the 637 bp endogenous PP2A3 promoter was transformed into the imp3 mutant strain, and whole cell extract from 6 putative transformants were screened by immunoblotting with an anti-HA antibody (Table S3). The HA-PP2A3 protein is predicted to have a molecular weight of 36,760 daltons. The anti-HA antibody recognized a ∼37 kD band in one transformant (imp3; HA-PP2A3, Figure 4A). Given the importance of Y313 and L315 (the equivalent of Y307 and L309 in mammalian cells) for the function of the PP2A catalytic subunit [33], we generated various N-terminal HA-tagged mutant forms (Y313Δ, L315A, L315Δ, and Y313F314L315Δ) under the same promoter and transformed them into the imp3 cells individually (Table S3). All mutant forms of the HA-PP2A3 protein are expressed, as detected by the anti-HA antibody but the amount of the protein is significantly less than the transformed wild-type HA-PP2A3 protein (Figure 4A). Additionally, we generated a V310T substitution to investigate whether this change has any effect on rescue of the imp3 mutant strain. The abundance of this protein is about 1.5 fold higher than the wild-type HA-PP2A3 transformant (Figure 4A). An immunoblot with a monoclonal antibody against α-tubulin was used to quantify protein loading (Figure 4A). Two smaller bands (∼27 kD and ∼23 kD) were also detected by the anti-HA antibody and they may correspond to proteolyzed/truncated PP2A3.
To ask whether the difference in protein abundance observed in the HA-PP2A3 transformants is due to the abundance of the transgenic HA-PP2A3 transcript or due to protein stability, we measured the transcript levels of PP2A3 by real-time PCR. In the wild-type strain CC-125 and imp3, real-time PCR detected only transcript levels of the endogenous PP2A3 transcript (Figure 4B, blue). In all HA-PP2A3 transformants, levels of two transcripts were detected. The first primer set detected the combined transcript levels of endogenous PP2A3 and transgenic HA-PP2A3 (Figure 4B, blue). Overall transcript levels of PP2A3 in all strains tested are comparable. The second primer set detected only the transcript level of transgenic HA-PP2A3 (Figure 4B, red). The transcript levels of transgenic HA-PP2A3 in all transformants are about one-quarter of the total PP2A3 transcript levels. There is no significant difference among different mutant transformants compared to the wild-type HA-PP2A3 transformant (Figure 4B, red). Therefore, we conclude the differences observed in the protein levels of HA-PP2A3 in different mutant transformants are not due to the abundance of the transgenic HA-PP2A3 transcripts, but rather due to the stability of the HA-PP2A3 proteins.
Transformation of wild-type HA-PP2A3 into imp3 cells successfully rescued the mating defect (Figure 4C, HA-PP2A3). Mating between imp3; HA-PP2A3 plus and minus gametes and mating between wild-type and imp3; HA-PP2A3 both produce thick pellicles (Figure 2A). The V310T change, which is found in 5 independent transformants, partially rescues the mating efficiency to about 60% (Figure 4C). None of the changes in the YFL motif rescues the mating phenotype (Figure 4B, Y313Δ, L315A, L315Δ, and YFLΔ), which indicates the importance of the last 3 amino acids in the function of PP2A3 during mating. Thus, we conclude that PP2A3 is encoded by the IMP3 gene and the deletion of nine nucleotides at the C-terminus of this gene causes the defective mating efficiency of imp3 cells.
We further asked whether inhibition of PP2A3 has any effect on mating efficiency. Okadaic acid (OA), a polyether fatty acid, was shown to inhibit the phosphatase activity of PP2A [40], enhance phosphorylation of Y307 [41], and inhibit methylation of L309 [42], [43] in vitro. OA inhibits PP2A at very low concentrations and the dissociation constant (Ki) between OA and PP2A is ∼0.032 nM [44]. From in vivo studies, however, the amount of OA required to inhibit PP2A varies from 10 nM in human lung cancer cells [45] to ∼1 µM in MCF7 breast cancer cells [46]. It is suggested that the entry rate of OA can be affected by pH, temperature, and exposure time to OA [47]. We tested the effect of OA on Chlamydomonas mating at concentrations of 1 nM, 10 nM, and 1 µM. The mating efficiency between wild-type CC-124 (mt−) and CC-125 (mt+) is around 75% (Figure 4D). The addition of DMSO and different concentration of OA, for one hour at room temperature, has no significant effect on the mating efficiency (Figure 4D, blue bars). Similarly, addition of DMSO or OA has no significant effect on the mating efficiency of imp3 mt+×imp3 mt− (Figure 4D, red bars) and imp3; HA-PP2A3 mt+×imp3; HA-PP2A3 mt− (Figure 4D, yellow bars). Pre-treatment of cells with autolysin, an enzyme that removes Chlamydomonas cell walls, before the addition of OA, leads to similar results (data not shown). In a study on phosphoproteome in Chlamydomonas, cells pre-incubated with 1.5 µM OA for 29 hours accumulate 38% more phosphorylated proteins [48]. Therefore, our OA results indicate that either one hour inoculation is not sufficient for OA to enter Chlamydomonas, or the effect of OA on Chlamydomonas is more complicated than simple inhibition of PP2A3.
To identify interacting proteins of PP2A3 and to investigate whether changes in the YFL motif lead to changes in protein-protein interactions, we performed immunoprecipitation with the anti-HA antibody. Two major bands of ∼65 kD and ∼37 kD are obtained by immunoprecipitation from whole cell extract from imp3 gametes transformed with wild-type HA-PP2A3 but not with untransformed imp3 gametes (Figure 5A). The ∼37 kD band is the HA-PP2A3, indicated by an immunoblot probed with an anti-HA antibody (Figure 5A). The ∼65 kD band was excised and subjected to mass spectrometry. The protein with the most number of peptides (94; 27 are unique) is PP2A-2r (Cre11.g477300) and it has a predicted size of 64,729 daltons. Changes in Y313Δ, L315Δ, V310T, or YFLΔ did not affect the pull-down of PP2A-2r by the HA antibody (Figure 5A and S3). Mass spectrometry of the ∼65 kD band pulled down by HA-PP2A3-V310T and by HA-PP2A3-YFLΔ resulted in 68 (27 unique) and 82 (26 unique) peptides of PP2A-2r, respectively. Thus, the interaction between the catalytic subunit PP2A3 and the scaffold subunit PP2A-2r is not affected by the changes at the C-terminus of PP2A3. PP2A-2r is one of the two PP2A scaffold proteins in the genome and was found in the flagellar proteome [34], [49]. Another scaffold subunit FAP14, which is also found in the flagellar proteome, has a predicted molecular weight of 100,787 daltons. Given that no significant band at ∼100 kD was identified in the immunoprecipitation (Figure 5A and S3), it is unlikely that PP2A3 interacts with FAP14.
Since PP2A-2r is present in the flagellar proteome, we asked whether PP2A3 localizes to the flagella. Flagella were isolated from imp3 and the HA-tagged transformants with the wild-type PP2A3, V310T, and YFLΔ genes. The HA-PP2A3 is detected in both wild-type and mutant transformants, but not in the untransformed imp3 flagella with the anti-HA antibody (Figure 5B). The YFLΔ transgene strain contains only ∼10% of HA-PP2A3 of those found in the wild-type HA-PP2A3 and the V310T strain; this is similar to observations in the whole cell extract immunoblots (Figure 4A). A smaller band (∼29 kD), which may represent proteolyzed/truncated HA-PP2A3, is again recognized by the anti-HA antibody (Figure 5B).
To compare the relative distribution of HA-PP2A3 in Chlamydomonas cell bodies and flagella, we perform immunoblots of HA-PP2A3 on the basis of cell equivalents (Figure 5C). Protein from equal numbers of whole cells and cell bodies, and from flagella isolated from about 40 times more cells, were used in the analysis. In both imp3; HA-PP2A3 and imp3; HA-PP2A3 YFLΔ strains, the HA-PP2A3 signal intensity is comparable in all three portions (Figure 5C). While the flagellar proteins represent less than 5% of protein found in whole cell extract [50], we do not find a significant enrichment of HA-PP2A3 in the flagella. In contrast, we observed a significant reduction of the HA-PP2A3 signal in imp3; HA-PP2A3 YFLΔ when compared to imp3; HA-PP2A3. Similar to previous observation (Figure 4A and Figure 5B), we noticed additional smaller bands in whole cells, cell bodies, and flagella. It is intriguing that the smaller bands observed in whole cells/cell bodies and in flagella are different in size and intensity. It is likely that these represent truncated PP2A3 proteins but the functions of these truncated proteins are unknown.
To ask where the HA-PP2A3 protein localize, we performed immunofluorescence with the HA antibody in six transformant strains (Figure 6A). In wild-type (CC-125) and untransformed imp3 gametes, there is some non-specific binding of the antibody in the cell body. In imp3; HA-PP2A3 gametes, robust signals are observed throughout the cells. In addition, in ∼80% of imp3; HA-PP2A3 gametes, accumulation of the signal is observed around the basal body area (Figure 6C, blue bars). The same signal intensity and localization is observed in imp3 cells transformed with HA-PP2A3 carrying a V310T mutation. In comparison, in imp3 cells transformed with the mutant forms of HA-PP2A3 (Y313Δ, L315A, L315Δ, and YFLΔ), the signal intensities of HA are significantly reduced, consistent with what we observed in the immunoblots (Figure 4A). Less than 10% of these cells showed basal body localization (Figure 6C). Therefore, we conclude that mutations of the terminal YFL affect the localization of PP2A3 to the basal body region.
Given the accumulation of the HA-PP2A3 proteins in wild-type HA-PP2A3 and V310T cells, we asked whether mating of these gametes with wild-type gametes would lead to change of localization of HA-PP2A3 (Figure 6B). When Chlamydomonas cells mate, the flagella adhere to each other, leading to cell fusion to form a single cell with four flagella and two nuclei, which is known as a dikaryon. We examined dikaryons one hour after mixing wild-type gametes (CC-125×CC-124); they show a low background of non-specific staining. In contrast, dikaryons formed between imp3; HA-PP2A3 (wild-type or V310T) and wild-type gametes show strong signals throughout the cells. However, less than 5% of these dikaryons show staining around the basal body area (Figure 6C, red bars). These results indicate that PP2A3 moves out of the basal body region in dikaryons.
Whole genome sequencing has become an important tool to allow quick identification of causative mutations in Chlamydomonas ([18] and Dutcher et al., submitted), Caenorhabditis elegans [16], Drosophila [51], and humans [52]. In Drosophila, direct comparison of sequences from parental and EMS mutagenized chromosomes leads to the identification of causative SNPs. This removes the need for sequence alignment to the reference genome sequences, which eliminates the natural variation of SNPs within different strains [51]. However, this approach is not feasible to identify mutants whose original strain backgrounds are unavailable. In C. elegans, a cross to a highly polymorphic strain and whole genome sequencing of a pool of 50 F2 progeny eliminates the need for meiotic mapping. The number of SNPs drops significantly within a ∼2 Mb region where the mutation resides. Thus, the number of SNPs of interest is reduced dramatically; it becomes easier to identify the causative mutation [16]. In the studies of human variants, databases such as dbSNP and The 1000 Genomes Project [53] are available to filter non-causative SNPs/indels. The filtering results in a reduction of ∼98% of SNPs/indels in a given individual and thus it becomes feasible to identify causative mutations for rare Mendelian diseases [52]. Similar to the human 1000 Genomes Project, a 1001 Genomes Project on Arabidopsis thaliana was initiated in 2008 [19]. Sequencing of 80 Arabidopsis strains identified ∼5.7 million SNPs/indels [54].
We previously used meiotic mapping to narrow the regions of interest to 269 kb in NG6/fla8-3 and 458 kb in NG30/ift80, respectively. Identification of one and six nonsynonymous changes in these regions eventually led to discovery of the causative mutations in these mutant strains [18]. In an approach similar to that used in C. elegans [16], we combined a pool of 14 progeny from a cross between a pf27 mutant strain and the highly polymorphic S1C5 strain for whole genome sequencing. This approach narrows the region of interest to ∼2 Mb on chromosome 12 ([55] and Alford et al., submitted), and eliminates the need for genome-wide mapping. However, the number of SNPs within 2 Mb remains large and it is hard to identify the causative mutation easily without further fine scale mapping. Therefore, generation of a SNP/indel library similar to the databases generated in other organisms is necessary to further eliminate common SNPs/indels found in Chlamydomonas strains.
While the two major laboratory strains (CC-124 and CC-125) were isolated from a single diploid zygote, 137c, they have >80,000 SNP/indel difference (Table 1). Both strains, and the reference strain used for genome assembly, belong to Subline III, as described by Pröschold et al. [39]. Therefore, strains that belong to Subline I, and strains from Subline II, such as uni1, are expected to contain more changes. The two isolo strains, originated from CC-124 and CC-125, after 10 rounds of backcrosses to CC-124, are expected to contain changes similar to each other and to CC-124. Instead, in addition to the expected difference observed on chromosome 6, the isolo strains show >30,000 polymorphisms between each other on chromosomes 3 and 17 (Figure 1A and Table 2). It is worth noted that the sequencing of isolo strains was performed on the Genome Analyzer IIx platform with 36 base-pair single-end reads and the sequencing of almost all other strains except uni1 was performed on the HiSeq platform with 101 base-pair paired-end reads. Thus, the quality of alignments and coverage on the genome are slightly lower in the isolo strain sequence. As a result, SNP/indel calling in the isolo strains may be less comprehensive than that in other strains (Figure 1A, Table 1, and Table 2). Nevertheless, the accumulation of changes on chromosomes 3, 6, 12, 16, and 17 found in CC-124 is observed in one or both isolo strains. The difference observed between isoloP and isolo M on chromosomes 3 and 17 suggests that in addition to the recombination-suppressed regions found on chromosome 6, there are additional regions on these two chromosomes that may show suppression of meiotic recombination. It is unclear why these regions are recombination-suppressed.
Given that many Chlamydomonas mutants were generated from the 137c background (Table S4), the SNP/indel library that we generated greatly facilitates quick identification of causative mutations in six mutant strains, fla18, fla24, fla9 (Dutcher et al., submitted), imp3 (this study), pf7, and pf8 (Dutcher et al., manuscript in preparation). The percentage of unique SNPs/indels after filtering is about 1∼2% in most strains and 7% in imp3 (Table 1). Even though we cannot eliminate the need for meiotic mapping in this study, we were able to combine this SNP/indel library and transcriptional profiles during flagellar regeneration [56] to identify mutations in flagellar assembly genes without the need of meiotic mapping (Dutcher et al., submitted). Thus, we anticipate that this SNP/indel library will facilitate rapid discovery of more causative mutations in Chlamydomonas mutant strains.
In humans, there are two isoforms (α and β) of PP2A catalytic subunits (PP2Ac). They share high sequence identity (97%), but the α isoform is ∼10 times more abundant than the β isoform at both transcript and protein levels [31]. In the invertebrates, Drosophila [57] and C. elegans [58], only one PP2A is present in each genome. In contrast, Arabidopsis has at least five PP2Ac proteins, and they can be divided into two subfamilies based on sequence similarity. The sequence identity between two subfamilies, I and II, is ∼80% [59]. The subfamily I proteins are involved in abscisic acid (ABA) and brassinosteroid signaling [60], [61]. The subfamily II proteins play an important role in auxin distribution and plant development [59]. In green algae, there are two PP2Ac proteins present in the genome. While one Chlamydomonas PP2A protein (PP2A-1c) belongs to the same group as the Arabidopsis PP2A subfamily I, the other Chlamydomonas PP2A protein, PP2A3/IMP3, is found in a green algal-specific cluster (Figure 3A). The sequence identity between PP2A-1c and PP2A3 is only 63%. While our results suggest that PP2A3 functions in flagella, it is likely that PP2A3 and its homologs in the green algae-specific cluster have additional functions since Ostreococcus lacks flagella.
Our study on PP2A3 provides several insights in the importance of the terminal YFL in PP2A. First, it is necessary for PP2A function in Chlamydomonas mating. Transgenes with mutations of individual amino acids or deletion of all three amino acids fail to rescue the mating defect (Figure 4C). Second, mutations or deletion of YFL do not affect the binding of PP2A3 to the scaffold subunit. In a study on the structure of the PP2A holoenzyme in vitro, the removal of the last 15 amino acids at the C-terminus of human PP2A does not affect the formation of the holoenzyme, which includes the catalytic subunit, the scaffold subunit, and the regulatory subunit PR61 γ1 isoform [62]. In cultured mammalian cell lines including COS7, NIH 3T3, and neuro-2a, changes of T304, Y307 and L309 alter the binding between PP2A and different regulatory subunits [63], [64]. Our study provides evidence that YFL is not required to form the core PP2A enzyme, which contains the catalytic C subunit and the scaffold A subunit (Figure 5A and S3), but it is unclear whether it can form the holoenzyme since the regulatory subunit that interacts with PP2A3 has not been identified. Third, loss of YFL affects the protein stability of PP2A3 in Chlamydomonas. It is not clear if this instability is observed in all kinds of cells. In mammalian cell lines, deletion or mutation of T304, Y307 and L309 do not affect the stability of HA-tagged PP2Ac as judged by immunoblots probed with an anti-HA antibody [63], [64]. However in yeast cells, contradictory results were obtained with tagged versus untagged proteins [65], [66]. It suggests that the HA tag might destabilize the PPH21p-L369Δ mutant [66]. In our study, the change or deletion of Y313, L315, and deletion of YFL, but not the V310T change, leads to reduction of PP2A3 protein levels. We think it is unlikely that the HA tag has an effect on the stability of PP2A3 mutants, but to completely rule out the possibility, a PP2A3-specific antibody will be necessary. Fourth, loss of YFL affects the localization of PP2A3 to the basal body region in Chlamydomonas. In C. elegans embryo, LET-92, the catalytic subunit of PP2A, localizes to centrosomes as well as in the cytoplasm. The centrosome localization of LET-92 depends on RSA-1 (Regulator of Spindle Assembly 1), a B″ type PP2A regulatory subunit that also localizes to centrosomes [67]. In monkey kidney CV-1 cells, immunofluorescence using a monoclonal antibody against PP2Ac revealed that PP2Ac localizes to both microtubule and centrosomes [68]. It was shown in HeLa cells both the B′ type PP2A regulatory subunit B56α and the PP2A scaffold subunit localize to the centrosomes and it was proposed that B56α serves as a chaperone that facilitates the translocation and function of the catalytic subunit [69]. In Chlamydomonas, the basal bodies serve as nucleation sites for assembly of flagella during interphase and function as centrioles during mitosis [70]. Our finding that PP2A3 localize to the basal body region is consistent with PP2Ac being found in centrosomes of other organisms. More importantly, our result suggests that mutations in the terminal YFL affect the basal body localization of PP2A3. It is likely that PP2A3 binds to a Chlamydomonas regulatory subunit similar to RSA-1 or B56α in the centrosomes and mutations in PP2A3 YFL attenuate the interaction between PP2A3 and the regulatory subunit.
In the Chlamydomonas mating signaling pathway triggered by agglutination of flagella from plus and minus cells, two proteins have identified. The TRPP2 protein Polycystin-2 (PKD2; Cre17.g715300) localizes to the flagellar membrane as well as to the cell body and is enriched fourfold in flagella during gametogenesis [71]. A cGMP-dependent protein kinase (PKG/CGK2; Cre02.g076900) is also present in both the cell body and flagella. Flagellar agglutination during mating leads to phosphorylation of a tyrosine residue in PKG that activates its protein kinase activity [50]. Knockdown of PKD2 and PKG via RNA interference leads to reduced mating efficiency, which resembles the imp3 phenotype [50], [71]. The signal transduction cascade is postulated to involve at least one additional uncharacterized tyrosine kinase that phosphorylates PKG [50].
Given the localization of PP2A3 to the basal body region as well as to the flagella, we propose that PP2A3 regulates the mating signaling pathway through one or more of the following mechanisms. Similar to other signaling pathways, phosphorylation and dephosphorylation of components in the mating signaling pathway need fine regulation. As a phosphatase, PP2A3 may act as a negative regulator of phosphorylation of PKG. The presence of PP2A3 keeps PKG unphosphorylated in gametes. In the imp3 gametes, since the amount of the PP2A3 protein is significantly reduced, dephosphorylation of PKG is likely to be compromised. Alternatively, PP2A3 could act as a positive regulator to facilitate the movement of signaling proteins as well as agglutinin proteins into the flagella to allow the proper functions of these proteins on the flagellar membrane during mating. It has been shown that the transport of PKD2 to the flagellar membrane requires intraflagellar transport (IFT) [71] but the entry of a truncated SAG1 protein to the flagellar membrane is IFT-independent [72]. It is possible that PP2A3 acts either on these proteins directly or on the transport machineries themselves. Comparison of PKD2 levels in wild-type and imp3 cells, as well as exploration of the relationship between PP2A3 and IFT will provide some answers to the function of PP2A3 in protein trafficking to the flagellar membrane.
The involvement of important signaling proteins such as PKD2 makes the mating process of Chlamydomonas an outstanding model to study ciliary/flagellar signaling. Our finding that PP2A3 involved in this process suggests that this ubiquitous enzyme may play an important role in this signaling. In addition, identification of the causative mutation in the imp4 mutant strain, which shares the same phenotype as imp3 but is unlinked to imp3 [21], may reveal additional players in this intriguing pathway.
Chlamydomonas reinhardtii strains, CC-124 (mt−), CC-125 (mt+), CC-1952 (S1C5), CC-3864 (fla18), CC-3866 (fla24), CC-1918 (fla9), CC-2668 (ida3), CC-465 (imp3), CC-3660 (pf23), CC-568 (pf7), CC-560 (pf8), CC-1926 (uni1), CC-916 (NG30, ift80), were obtained from the Chlamydomonas Resource Center. Both isoloM (CC-4403, mt−) and isoloP (CC-4402, mt+) were generated in this laboratory and deposited to the Chlamydomonas Resource Center. All mutant strains obtained were backcrossed at least once with either CC-124 or CC-125, and progeny with the mutant phenotype were used in further analysis. The cnk10 mutant strain was generated by insertional mutagenesis in the CC-125 background.
Chlamydomonas genomic DNA was prepared as previously described [18]. Approximately 108 cells were used in DNA preparation. About three µg of genomic DNA from each strain was submitted to Genome Technology Access Core (Department of Genetics, Washington University in St. Louis) for Illumina sequencing. The 36 bp single-ended (SE) sequencing of isoloP and isoloM and the 60 bp pair-ended (PE) sequencing of uni1 were performed with the Genome Analyzer IIx platform. The 101 bp PE sequencing of the other 13 strains was performed with the HiSeq platform. Almost all samples with the exception of ift80 and uni1 were individually tagged with a unique 7-nucleotide index and multiple samples were subjected to one sequencing flow cell lane (Table S4). The sequencing reads of ift80 were obtained from a previously published result [18]. The indexed sequencing reads were de-multiplexed before being subjected to read alignment.
For read alignment, the genome sequence of Chlamydomonas v5.3.1 (Creinhardtii_236.fa.gz) was downloaded from http://www.phytozome.net/chlamy.php [36]. An indexed database was built from 54 FASTA sequences (17 chromosomes+37 scaffolds) by Novoindex (novocraft.com). The SE sequencing reads were aligned to the database by Novoalign (version 2.08.02) with the following options: -o SAM -r random -l 25 -e 100 -a AGATCGGAAGAGCGGTTCAGCAGGAATGCCGAG -H -c 12. The PE sequencing reads were aligned to the database by Novoalign with the following options: -o SAM -r random -l 30 -e 100 -i 230 140 -a AGATCGGAAGAGCGGTTCAGCAGGAATGCCGAG AGATCGGAAGAGCGTCGTGTAGGGAAAGAGTGTA -H -c 12 -h 90 120. SAMtools (version 0.1.18) [73] was used to convert the resulting SAM (Sequence Alignment/Map) files to BAM (Binary Sequence Alignment/Map) format and sort the BAM files. Duplicated reads were removed by Picard (version 1.46; picard.sourceforge.net) and the resulting BAM files were sorted by SAMtools.
The same genomic DNA sequence (v5.3.1) was indexed by SAMtools and used as a reference file in SAMtools mpileup (options: -u -g), which outputs a BCF (Binary Call Format) file for BCFtools (options: -b -v -c -g; vcfutils.pl varFilter -D 999) to call for SNP/indel variants using Bayesian inference [73]. The numbers of SNPs/indels from individual strains were obtained from the resulting VCF (Variant Call Format) files.
VCF files of all 16 strains sequenced are available on the SNP/indel library webpage (http://stormo.wustl.edu/SNPlibrary/). The webpage is designed to allow direct comparison of SNPs/indels from one strain to one or more strains of interest. It also allows users to upload their own VCF files and compare to SNPs/indels from one or more of the 16 strains we provide. The output is a VCF file which maintains its original format minus the common SNPs/indels. To detect unique changes in the imp3 mutant strain, we compared the SNPs/indels to the SNPs/indels in all other 15 strains. The resulting new VCF file was then subjected to SnpEff (version 3.0j) [74] to identify SNPs/indels within the coding regions or exon/intron boundaries (options: -no-upstream -no-downstream -no-intergenic -no-intron -no-utr -hom -o txt). The SnpEff reference database for was built based on the genome assembly and gene annotation (Creinhardtii_236_gene.gff3.gz) of Chlamydomonas v5.3.1 [36].
In order to quickly identify the imp3 progeny from crosses between imp3 and wild-type cells as well as between imp3 and S1C5 cells, we developed a low-density mating assay. In this assay, approximately 1o6 cells were subjected to 0.1 ml of nitrogen-free medium (M-N/5) in a well of a 96-well plate for 4 hours at 21°C. Half of the volume of cells was mixed with imp3 mt+ testers and half were mixed with imp3 mt− testers. Starting at 30 minutes after mixing and at 30-minute increments for four hours, the cells were monitored under the dissecting microscope for agglutination and the formation of pellicle. The cells were incubated under constant light overnight at 21°C and then scored again after 18 hours. Mating between IMP3 cells and imp3 testers formed multi-layer, dark green pellicle around the four-hour time point, which was maintained until the 18-hour time point. Mating between imp3 and imp3 testers formed a thin, single-cell layer beginning around the two-hour time point but it never accumulated into the multi-layer pellicle within the next 16 hours. In the cross between imp3 and S1C5, over 100 independent imp3 progeny were picked from over 100 tetrads and used for meiotic mapping.
Crude DNA was obtained from about 106 cells from each progeny. Cells were resuspended in 10 µl 1× Vent Buffer (20 mM Tris-HCl, 10 mM (NH4)2SO4, 10 mM KCl, 2 mM MgSO4, 0.1% Triton X-100, pH 8.8) and 1 mg/ml proteinase K (Sigma-Aldrich). The cells were incubated at 58°C for 30 minutes for proteinase K activity and 95°C for 15 minutes to inactivate proteinase K. For PCR, 0.5 µl of crude DNA was used in 20 µl reactions. Each PCR sample contained 1× TAQ buffer (50 mM KCl, 10 mM Tris-HCL, 0.1% Triton X-100, pH 9.0), 1.5 mM MgCl2, 0.4 µM dCAPS primers, 0.4 mM dNTPs, 5% DMSO and 4% TAQ polymerase. The cycling protocol was 95°C for 2 minutes, and 30 cycles of 95°C for 20 seconds, annealing temperature for 20 seconds (Table S1), and 72°C for 1 minute, followed by a final extension time of 5 minutes at 72°C. If digestion of the PCR product was necessary to detect the polymorphism, 10 µL of the PCR product, 1.2 µL of an enzyme's corresponding buffer, 1.2 µL of 10× BSA if necessary, 0.2 µL of the enzyme, and the remaining volume with water to reach a final 15 µL. This was placed at 37°C from 5 minutes to 1 hour depending on the enzyme's efficiency as suggested by the manufacturer (New England Biolabs).
The protein sequence of PP2A3 was used in BLAST on NCBI and 55 proteins with the expected E-value less than or equal to 1E-100 were collected from green algae (Chlamydomonas reinhardtii, Chlorella variabilis, Micromonas, Ostreococcus lucimarinus, Ostreococcus tauri, and Volvox carteri), yeast (Saccharomyces cerevisiae), land plants (Arabidopsis thaliana and Zea mays), invertebrates (Caenorhabditis elegans and Drosophila melanogaster), and mammals (Mus musculus and Homo sapiens) (Table S2). Sequences alignment was performed with MUSCLE [75] and the output ClustalW (strict) file was subjected to Colorfy [76] for color-coded protein sequence alignment.
From the protein sequence alignment (Figure S1), the low similarity region (∼30 amino acids from the N-terminus) was trimmed from individual protein sequences (Table S2) and multi-sequence alignment was again performed by MUSCLE. The output Phylip interleaved file was subjected to SEQBOOT for 100 bootstraps in the PHYLIP package (3.69; http://evolution.genetics.washington.edu/phylip.html). The 100 sets of samples were subjected to PROTDIST, which uses protein distance matrix to calculate distance. An unrooted tree was generated by a neighbor-joining method implemented by NEIGHBOR for each set of samples. A consensus tree from the 100 unrooted trees was computed by CONSENSE. The output tree was visualized by DRAWGRAM and the bootstrap numbers for each branch were obtained from the outfile file generated by CONSENSE (Figure 3A).
Based on the phylogenetic tree, proteins from the PP2A subfamily were collected and subjected to another round of multi-sequence alignment by MUSCLE and color-coded by Colorfy (Figure S2). Sequence alignment of amino acids close to the C-terminus is shown in Figure 3B.
The full-length PP2A3 genomic DNA, which contains 637 bp upstream of the start codon, the 948 bp coding region, and 424 bp downstream of the stop codon was amplified by primers PP2A-inF2 and PP2A-inR (Table S5) with the Phusion DNA polymerase (New England Biolabs) with the following cycling condition: 98°C for 30 sec, and 30 cycles of 98°C for 10 seconds, 69°C for 10 seconds, and 72°C for 1 minute, followed by a final extension time of 2 minutes at 72°C. The resultant PCR product was gel-purified by UltraClean 15 DNA Purification Kit (MO BIO Laboratories) and cloned into a pBlueScript vector digested with EcoRV using the In-Fusion HD Cloning Kit (Clontech). The resultant PP2A3-pBS plasmid was subjected to Sanger sequencing (GENEWIZ) to ensure the sequence accuracy.
The HA tag was inserted right after the start ATG codon via nested PCR. DNA polymerase and cycling conditions were the same as above with the exception of extension time was reduced to 20 seconds. The primers (PP2A-HA-F, PP2A-3R, Table S5) in the first round of PCR with the PP2A3-pBS plasmid generated above as a template, introduced the HA tag and generated a 361 bp fragment, which was gel purified and used as a template in the second round of PCR. A second forward primer (PP2A-3F-HA), along with the same reverse primer (PP2A-3R), produced a 399 bp fragment. This fragment was then digested with AvrII and AflII, gel purified, and cloned into PP2A3-pBS digested with the same enzymes to generate the HA-PP2A3-pBS plasmid. The plasmid was subject to Sanger sequencing (GENEWIZ) to ensure the sequence accuracy.
The mutations to the C-terminus of the protein were introduced by knitting PCR. DNA polymerase and cycling conditions were the same as above with the exception of extension time was 30 seconds. Using the HA-PP2A3-pBS plasmid as a template, the PP2A3-inf-3F primer was paired with different reverse primers each bearing a mutation (Y313Δ-R; Y315A-R; L315Δ-R; V310T-R; YFLΔ-R, Table S5) to amplify ∼430 bp bands. The PP2A3-inf-3R primer was paired with different forward primers each bearing a mutation (Y313Δ-F; Y315A-F; L315Δ-F; V310T-F; YFLΔ-F, Table S5) to amplify ∼410 bp bands. All bands were gel purified and the corresponding bands (i.e., PP2A3-inf-3F-Y313Δ-R and Y313Δ-F-PP2A3-inf-3R) were used as templates in the second round of PCR, with PP2A3-inf-3F and PP2A3-inf-3R as primers. The resultant ∼830 bp products from the second round of PCR, were gel purified, digested with NcoI and PmlI and cloned into HA-PP2A3-pBS plasmid digested with the same enzymes. All mutant plasmids were subject to Sanger sequencing (GENEWIZ) to ensure they bear the correct mutations.
Chlamydomonas transformation was performed as previously described with modification [76]. One µg of plasmid DNA was used in each transformation into the imp3 cells, using 1 µg of the pSI103 plasmid, which confers resistance to paromomycin [77], for co-transformation. Paromomycin-resistant (paroR) cells were selected on TAP plates with 10 µg/ml paromomycin. In these TAP plates, the concentration of glacial acetic acid was 0.75 ml/L instead of 1 ml/L [35]. The reduction of acetic acid in the medium confers more robust paromomycin resistance. Crude DNA was prepared from paroR colonies and PCR with primers PP2A-HA-short and PP2A-4R (Table S5) was performed to identify HA-positive colonies. The HA-positive colonies were then subjected to immunoblots with anti-HA antibody to confirm the presence of HA-tag in the transformants. The number of paroR colonies, number of HA-positive colonies by PCR and by immunoblot are summarized in Table S3.
The primary antibodies used in this study were anti-HA High Affinity antibody (3F10, Roche, 1∶2000 dilution in immunoblots and 1∶200 dilution in immunofluorescence), anti-α-tubulin antibody (Sigma-Aldrich, T6199, 1∶2000 dilution in immnoblots), and anti-acetylated- α-tubulin antibody (Sigma-Aldrich, T7451, 1∶250 dilution in immunofluorescence). The secondary antibodies used for immunoblots were HRP-conjugated goat anti-mouse antibody (BioRad, 1∶5000), and HRP-conjugated goat anti-rat antibody (Sigma-Aldrich, 1∶5000). Alexa 488-conjugated goat anti-rat antibody (Invitrogen, 1∶500) and Alexa 594-conjugated goat anti-mouse antibody (Invitrogen, 1∶500) were used in immunofluorescence.
Cells used in immunoblots were gametes, which were cultured on R plates for 5 days at 25°C and resuspended in nitrogen-free medium (M-N/5) for 4 hours at room temperature prior to protein extraction. For whole cell extracts, ∼107 cells were resuspended in 30 µl of 20 mM HEPES (pH 7.0) solution with the addition of 1× ProteaseArrest (G-Biosciences) and 10% SDS. The samples were incubated at 37°C for 5 minutes and diluted with 90 µl of 20 mM HEPES (pH 7.0) and 1× ProteaseArrest [78]. For flagellar extracts, ∼5×109 Chlamydomonas cells were resuspended in 4°C HEPES/Sr/DTT buffer (pH 7.1) buffer. Flagella were then detached from cells by pH shock on ice, collected by centrifugation at 4°C [79], and resuspended in 40 µl of 20 mM HEPES (pH 7.0) and 1× ProteaseArrest.
Protein extract from all samples isolated were store at −80°C. The amount of protein in each sample was measured by Bradford assay (Bio-Rad protein assay) before loaded into 10% polyacrylamide gels. The proteins were transferred to PVDF membrane (Millipore) after SDS-PAGE, blocked with 5% nonfat dry milk in PBST (Phosphate Buffered Saline with 0.02% Tween-20) at room temperature for 1 hour, probed with primary antibody in 3% milk in PBST overnight at 4°C, washed with PBST 3 times, 5 minutes each, probed with secondary antibody in 3% nonfat dry milk in PBST 1 hour at room temperature, and washed with PBST 3 times, 5 minutes each. ECL Plus Western Blotting Detection Reagents (GE Healthcare Life Sciences) or SuperSignal West Femto Chemiluminescent Substrate (Thermo Scientific) was used to expose the signal, which was captured by a FluorChem H2 image (Alpha Innotech). Signal quantification analysis was performed by ImageJ (NIH).
For immunofluorescence, gametes were treated with autolysin for 30 minutes before fixed with 2% paraformaldehyde as previously described [80]. The autolysin treatment was omitted in immunofluorescence of the dikaryons. The images were captured with an UltraVIEW VoX laser scanning disk confocal microscope (PerkinElmer) and acquired by Volocity software (PerkinElmer).
Immunoprecipitation was performed as described by Olson et al. [81] with modifications. About 109 gametes were used and proteins were cross-linked with freshly prepared DSP (Dithiobis-succinimidyl propionate; Pierce) on ice for 30 minutes before the cross-link was stopped by the addition of 100 mM Tris-HCl (pH 7.5). Cells were lysed by sonication and lysates were centrifuged at 20,000 g for 30 minutes at 4°C to remove insoluble materials. Protein concentration from the supernatant was determined by Bradford assay. One microgram of anti-HA high affinity antibody (3G10) was inoculated with 20 µl of protein G magnetic beads (Invitrogen) for 30 minutes at room temperature and cross-linked with 5 mM BS3 (Bis[sulfosuccinimidyl] suberate; Pierce) in conjugation buffer (20 mM sodium phosphate, pH 7.0, 0.15 M NaCl) for 30 minutes at room temperature. The cross-link reaction was stopped by the addition of 50 mM Tris-HCl (pH 7.5) and incubated at room temperature for 15 minutes. The anti-HA-antibody-conjugated protein G magnetic beads were then incubated with cell lysates that contain ∼12 mg of proteins in each sample overnight at 4°C. The beads were washed and the bound proteins were eluted with 1× SDS PAGE loading buffer with the addition of 100 mM DTT at 90°C for 10 minutes. About 1/10 of the elution was used in SDS-PAGE and immunoblots with anti-HA antibody and the rest was separated by a 10% polyacrylamide gels and stained by either silver [82] or 0.1% Coomassie Brilliant blue R-250. The ∼65 kD band stained by Coomassie blue was cut out and sent to the Taplin Mass Spectrometry Facility (Harvard University) for mass spectrometry.
Chlamydomonas total RNA extraction and real-time RT-PCR was performed as previously described [76]. Primers used in these reactions were PP2A3-HA-short and PP2A3-4R for HA-PP2A3 transcripts and PP2A3-7F and PP2A3-3R for endogenous+HA-PP2A3 transcripts (Table S5).
|
10.1371/journal.pntd.0002001 | Field Cage Studies and Progressive Evaluation of Genetically-Engineered Mosquitoes | A genetically-engineered strain of the dengue mosquito vector Aedes aegypti, designated OX3604C, was evaluated in large outdoor cage trials for its potential to improve dengue prevention efforts by inducing population suppression. OX3604C is engineered with a repressible genetic construct that causes a female-specific flightless phenotype. Wild-type females that mate with homozygous OX3604C males will not produce reproductive female offspring. Weekly introductions of OX3604C males eliminated all three targeted Ae. aegypti populations after 10–20 weeks in a previous laboratory cage experiment. As part of the phased, progressive evaluation of this technology, we carried out an assessment in large outdoor field enclosures in dengue endemic southern Mexico.
OX3604C males were introduced weekly into field cages containing stable target populations, initially at 10∶1 ratios. Statistically significant target population decreases were detected in 4 of 5 treatment cages after 17 weeks, but none of the treatment populations were eliminated. Mating competitiveness experiments, carried out to explore the discrepancy between lab and field cage results revealed a maximum mating disadvantage of up 59.1% for OX3604C males, which accounted for a significant part of the 97% fitness cost predicted by a mathematical model to be necessary to produce the field cage results.
Our results indicate that OX3604C may not be effective in large-scale releases. A strain with the same transgene that is not encumbered by a large mating disadvantage, however, could have improved prospects for dengue prevention. Insights from large outdoor cage experiments may provide an important part of the progressive, stepwise evaluation of genetically-engineered mosquitoes.
| The absence of a commercially-available dengue vaccine or anti-viral drug makes control of Aedes aegypti, the principal dengue mosquito vector, the only available method to prevent this disease. Sustained, effective application of vector control, however, has been difficult and this led to the call for innovative strategies, including genetic approaches. Here, the authors investigated the ability of a genetically-engineered strain of Ae. aegypti to eliminate wild mosquito populations in large outdoor cages. Females of the engineered strain cannot fly and, therefore, cannot mate or take blood meals necessary to lay viable eggs. Wild females that mate with genetically-engineered males, therefore, will not produce reproductive female offspring. In this study, although population reductions were detected in 4 of 5 field cages, none of the wild mosquito populations were eliminated. A mating disadvantage for genetically-engineered males appeared to account for a significant part of their fitness disadvantage. Results suggest that this specific strain may not be effective in a large-scale release and that new strains with the same or similar transgene, but improved mating performance, may be more effective for preventing dengue. Results also indicate that large outdoor cage experiments may provide valuable insights into the progressive, stepwise assessment of genetically-engineered mosquitoes.
| The recent worldwide increase in dengue [1], [2] has made urgent the development and assessment of new tools for controlling the disease [3]. Because no vaccines or drugs are commercially available [4], [5], mosquito vector control by insecticides, insect growth regulators and larval development site elimination (source reduction) are the current means for dengue prevention [6]. Long-term control of Aedes aegypti, the most efficient dengue vector [7], is a challenging and expensive task that is difficult to achieve and maintain, especially in developing, resource-challenged environments [8]–[10]. Genetically-engineered (GE) Ae. aegypti strains that are unable to transmit dengue [11] or that bear sterility genes [12], [13] constitute new tools to control dengue and merit confined experimental evaluation while public and scientific discourse enables appropriate oversight of this new technology [14], [15]. Concern regarding the use of GE organisms, and the absence of guidelines to help researchers interact with local communities, motivated the elaboration of a framework for the development, evaluation, and application of genetic strategies for prevention of mosquito-borne disease [16]. These guidelines were followed carefully in the development and execution of the experiments described here.
The OX3604C strain of Ae. aegypti contains a tetracycline-regulated transgene that induces a female-specific flightless phenotype that cannot reproduce as a consequence of its inability to fly and mate [17]. Tetracycline is added to larval rearing water to allow normal female development during colony maintenance and amplification, but is not added during the generation used for mass-production of males. This enables genetic removal of females, because females carrying the transgene and reared in the absence of tetracycline cannot fly. Similarly, female offspring that result from matings between wild-type females and released OX3604C males are unable to fly or reproduce. The goal in releasing OX3604C males is to control dengue virus transmission by reducing or eliminating Ae. aegypti populations. The release of insects carrying a dominant female-lethal construct has four main advantages compared to a traditional sterile insect technique (SIT): (i) no need to sort males and females, (ii) no need for facilities to irradiate males, (iii) the transgene has an effect in subsequent generations because it is dominant and inherited by male offspring, and (iv) OX3604C contains a heritable, fluorescent marker (DsRed2) that distinguishes it from immature wild-type Ae. aegypti [17].
As an initial step in OX3604C evaluation, transgenic males were introduced weekly at an 8.5–10∶1 OX3604C∶target ratio into large laboratory cages with constant temperature, humidity, and photoperiod, that contained stable target populations of wild type Ae. aegypti [18]. Target populations were eliminated in all experimental cages in 10–20 weeks, supporting further analyses of this strain in contained or confined field trials to evaluate mating competitiveness and environmental and other effects on its performance [18].
Progressive evaluation of OX3604C from laboratory to field cages prior to open field release is valuable because it allows for systematic assessment of possible environmental effects on mosquito performance under conditions increasingly more natural. Comparison of transgenic mosquito performance in laboratory versus semi-field conditions is expected to provide valuable data for planning subsequent experimental assessments and refine strategies for disease prevention. Insectary studies in a laboratory, field cage experiments and deliberative community engagement activities are all part of the progressive transition of engineered insects from the laboratory to open field releases [19], [20]. This is particularly true when the transgene as well as all other genes in the transgenic strain can be introduced into natural target populations and transmitted to subsequent generations. Even though OX3604C is a self-limiting strategy that lacks a gene drive component, it can introduce through heterozygous males new alleles and genes into target populations.
We report the effect of OX3604C in reducing target Ae. aegypti populations in the first large outdoor field cage trial of a transgenic population suppression tool. While density reduction was significant in four of five target populations, we did not observe population elimination in any of the cages within the time expected. A series of subsequent experiments revealed a significant mating disadvantage for OX3604C males that was not observed in the laboratory study. We discuss the implications of our results for OX3604C and more broadly for future evaluations of genetically-engineered mosquitoes.
Our study was carried out on a plot of land (14°51′41″N, −92°21′15″W) referred to hereafter as the “field site.” The land was a 4.5 ha flat, rural area located 11.2 km southeast of the center of Tapachula, Mexico, in the village of El Zapote (Ejido Rio Florido). The study area is characterized by a tropical climate with a rainy season from May to October (average total rainfall of 2,100 mm) and a dry season from November to April (average total rainfall of 50 mm). Supportive laboratory and insectary facilities were located at Centro Regional de Investigación en Salud Pública (CRISP), Tapachula, 15 km from the field site.
The protocol used in this study was similar to that used during the previous OX3604C assessment in laboratory insectary conditions [18]. Materials and methods were adapted to different logistics and biosecurity conditions required for a field-cage experiment. The most important differences between laboratory and field-cage experiments are summarized in Table S1. Mexico has a mature regulatory system for the use of genetically modified organisms, having a law and implementing regulations in place since 2005. Field cage experiments must comply with basic biosafety procedures, oversight, and registration of all experimental procedures, installations and monitoring programs. Our protocol was approved by the Mexican institutions Instituto Nacional de Salud Pública (#581) and Secretería de Medioambiente y Recursos Naturales (S.G.P.A./DGIRA/DG/7074/09).
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Protocols were approved by the Institutional Animal Care and Use Committee of the University of California, protocol 15653 [UCD] and the Instituto Nacional de Salud Pública INSP Biosecurity permit #581 [CRISP].
Our semi-field system consisted of six caging units each measuring 6×6×2 m (LxWxH) (Figure S1). A solid plastic roof with sunscreen around the edges covered all six cages, and provided shade and protection from direct sunlight. One caging unit was used as a field laboratory for larval rearing and adult mosquito handling. Each of the other five caging units was divided in half by means of zippered mesh walls resulting in a total of 10 6×3×2 m cages (five pairs) (Figure S2). Each pair was provided with two vestibules that allowed access to both cages through three sleeves, two of which opened to shelves inside the cage and a third opening to the cage floor (Figure S3). Cages were made of white tricot mesh reinforced with white fabric and located on a platform elevated 1 m above the ground, with ∼3–5 m between the top of each cage and the roof covering the platform. Materials and general design used to build the cages is described in detail by Facchinelli et al. [21]. Temperature and relative humidity were measured by means of data-loggers (Hobo Pro v2 temp/RH, Onset Computer Corporation, Bourne, MA), located inside each cage and in the outdoor environment, 30 meters from the cages.
The protocol for field cage design, OX3604C colony maintenance, and field cage experiments was approved by the Instituto Nacional de Salud Pública (INSP) and Secretaría de Medio Ambiente y Recursos Naturales (SEMARNAT), Mexico, under the provisions of the law on genetically modified organisms (Ley de Bioseguridad de Organismos Genéticamente Modificados, (Marzo 2005)). The field cage experiment protocol included procedures for detection of potential escapes of OX3604C mosquitoes to the open environment. Ten ovitraps and 10 BG-Sentinel Mosquito Traps (Biogents AG, Regensburg, Germany) were distributed on the ground around the cage platform. Three BG Sentinel Traps were located inside the CRISP insectary where larvae were reared. Ovitraps were serviced weekly and collected eggs were hatched and larvae screened for the fluorescent marker. BG Sentinel Traps were checked daily and Ae. aegypti adults were processed by gene amplification (PCR) for transgene detection. No transgenic mosquitoes were collected outside the field cages at the field site. Two transgenic adults were collected by BG Sentinel Traps in the room dedicated to OX3604C colony maintenance in the CRISP insectary.
The OX3604C strain employed in this study was obtained by backcrossing the homozygous OX3604C strain [17] into the genetic background of the genetically-diverse laboratory strain #1 (GDLS1; [18]) derived from 10 geographically-distinct populations collected during 2006 in Chiapas, Mexico [22]. Approximately 96.9% of the genome not linked to the transgene is expected to consist of GDLS-derived sequences [18]. Before introducing OX3604C into treatment cages, it was determined that the strain was not homozygous for the transgene. Screening for the fluorescent marker showed that the wild-type allele had a frequency ranging between 4.7 and 8.7%. A mathematical model indicated that at the release ratios we used, this low level of wild-type alleles was not expected to significantly affect the outcome of the experiment (Figure S4 and S5). Wild-type females were removed prior to introducing OX3604C males into treatment cages.
The target and control Ae. aegypti populations in our study were the genetically-diverse laboratory strain #2 (GDLS2) derived from 2008 field collections in the same locations in Chiapas where GDLS1 originated [18]. Aedes aegypti females were allowed to imbibe blood directly from rabbits for colony maintenance, field cage equilibration, and experiments (UC Davis Animal Care and Use protocol 15653 and INSP Biosecurity permit #581).
Target and control populations were established in each of the 10 experimental cages during a 16-week period using GDLS2 mosquitoes. Shelves inside each cage held (i) a total of four plastic oviposition containers filled with ∼600 ml sterilized water and were lined with filter paper, (ii) four plastic larval development containers filled with ∼600 ml sterilized water and enclosed in a mesh-covered dome to prevent oviposition, and (iii) four plastic plates containing raisins as a source of sugar for adults (Figure S3). Each cage also contained two humid, adult resting sites; i.e., 15 L black plastic buckets that were lined with black fabric and contained a mesh-covered water container.
GDLS2 eggs were hatched in Centro Regional de Investigación en Salud Pública (CRISP) insectary and larvae transported to the field site after 48 hours. Establishment of the target population was initiated by adding 300 second-instar GDLS2 larvae weekly in each cage from week 0 to week 3. Larvae were fed dried brewer's yeast ad libitum and females were fed blood from restrained rabbits once a week for 30 minutes. From week 4 to week 16, the target and control GDLS2 populations were maintained by returning eggs laid by females in each cage to their respective cages as second-instar larvae at a rate of 200/week. Eggs produced in cages were collected twice a week and adults were sampled weekly to monitor population dynamics. Eggs were transported to the insectary, counted, dried and stored in a humidified chamber, and then hatched for the next generation. Adults were sampled using BG Sentinel Traps placed in each cage for 30 minutes each week, counted, sexed after sedation on a CO2 sedation device (Figure S6), and returned to their respective cage. Adult trapping and handling resulted in 0–8% mortality. Cages were inspected at least weekly for the presence of spiders, ants or other mosquito predators. When arthropod predators were found they were removed mechanically, without the use of insecticides. No vertebrate predators were found during the 33-week period that mosquitoes were in field cages.
The lack of a homozygous OX3604C strain prevented us from rearing and allowing OX3604C mosquitoes to emerge into treatment cages as was performed by Wise de Valdez et al. [18]. Instead, pupae were sexed by visual examination for size at the field laboratory insectary using 3 ml plastic droppers and only adult males were added into treatment cages to avoid introducing the few females lacking the transgene (∼0.5%) that could have interfered with population extinction. OX3604C eggs were hatched weekly without tetracycline in the CRISP insectary, second-instar larvae were transported to the field laboratory insectary, placed in 35 plastic trays (each tray contained 500 larvae in 1.5 L of water) and fed dry ground yeast ad libitum. Pupae were collected over the course of three days and sexed so that a total of 5,250 male pupae were introduced into the five-insectary cages (1,050 per 30×30×30 cm cage, taking into account 5% mortality), which were held in the field laboratory insectary. The sex of recently emerged adults was checked daily by visual examination of adults in cages, females were removed, and males were introduced into their respective treatment cages every 24 h over the course of 4 days each week. The release number remained constant. During the first OX3604C male release [week 0 post-release], the release ratio was approximately 10 times the weekly return rate of 200 second-instar/GDLS2 larvae/week (i.e., approximately 100 GDLS2 males). Because the mosquito population tended to decreased over time in the treatment cages, the release ratio correspondingly increased from 10⋮1 OX3604C⋮target males during Week 0 post-release to between 14⋮1 to 1,000⋮1 during Week 17 post-release (Figure S7). During Week 1 and 2 post-release, a total of 63 GDLS2 males from control cages and 50 OX3604C males that emerged from rearing trays at the field site were collected to compare size of the two strains. Right wing measurements were used to determine mosquito body size.
Starting on week 0 post-release, a 10% sample of eggs produced in each treatment cage was screened weekly for the DsRed2 marker to assess transgene introgression into target populations. When the first fluorescent larvae were detected in treatment cages confirming that OX3604C males were mating with GDLS2 females (week 3 post-release), the number of larvae added back to each cage was adjusted to reflect the impact of the OX3604C male release on egg production. Briefly, the number of larvae returned to control cages remained constant at 200 second-instar larvae/week, while the number of larvae added back to each treatment cage was changed to a proportion of the egg number in the respective paired control cage based on the procedures applied in Wise de Valdez et al. [18]. At the end of our experiment, week 18 post-release, all mosquitoes present in the field cages were collected with a backpack aspirator (John W. Hock Company, 23rd Ave, Gainesville, 32606 Florida, U.S.A.).
A total of six mating competition experiments were carried out between December 2010 and June 2011 with the aim of investigating the lower performance of OX3604C in the field cages trial versus the laboratory insectary trial [18].
Each of five caging units was partitioned into two paired cages (pair A consists of cages 1 and 2, pair B consists of cage 3 and 4, etc.) with dimensions of 6×3×2 meters (LxWxH) (Figure S1 and S2). Populations of the GDLS2 [18] were established in the ten cages for a period of 16 weeks (from April to August 2010) prior to the release of OX3604C males. Population densities stabilized in all cages by week 9 (Figure 1). One cage in each pair was assigned randomly during week 16 as a control or treatment cage (Figure S2). From week 16 to week 33 (from 16 August to 23 December 2010), corresponding to week 0 through week 17 post-release, ∼1,000 OX3604C males were introduced weekly into each treatment cage. This number corresponds to an approximate initial 10∶1 OX3604C∶GDLS2 male release ratio (Figure S7). The constant release number of OX3604C males, calculated to establish the initial 10∶1 ratio based on input rate by the average lifespan, was maintained from week 0 to week 17 post-release. When transgene introgression into the caged populations was first detected (presence of the DsRed2 marker gene in larvae), during week 3 post-release (Figure 2), the weekly number of larvae returned to each treatment cage was adjusted relative to the weekly return rate in the respective paired control cage (held constant at 200 second-instar larvae/week in all control cages). This was done to reflect any impact of OX3604C male release on egg production by females in each treatment cage. Based on the number of larvae returned to treatment cages from week 3 to week 17 post-release, release ratios of OX3604C∶target males increased in all treatment cages reaching the highest value of 1,000∶1 in cage 1 during week 17 post-release (Figure S7).
Temperatures during the trial ranged between 14.5°C and 41.8°C in all cages except cage 6 where a peak of 44.2°C was recorded on 27 April (Table S2). During the rest of the trial, temperatures in cage 6 were similar to those recorded in the other field cages. Daily temperature fluctuations ranged between 2.0°C and 20.7°C. Relative humidity (RH) in field cages ranged from 38.1% during the warmer hours of the day to 99.4% at dawn. Temperatures recorded outside of the cages were similar to those recorded in cages and ranged from 15.8°C to 40.7°C. RH outside of cages also was similar to that recorded inside cages, ranging from 42.8% to 100% with a mean of 89.2%±11.6% (SD) (Table S2).
Weekly adult sampling performed the day before starting the next OX3604C release into treatment cages confirmed the presence of a significantly higher number of males in treatment cages compared to their respective control cages (Mann-Whitney U test p<0.01 for all cage units) (Figure 3). The percentage of DsRed2 larvae produced in treatment cages fluctuated between 1 and 76% but never reached 100% (Figure 2).
ANCOVA indicated that the number of eggs produced in treatment cages decreased significantly subsequent to male OX3604C release compared to respective paired control cages in all treatment cages except cage 4, where covariates did not indicate an effect (Figure 4). Ratios of females collected in control vs. treatment cages per each cage unit matched results of ANCOVA (F) (Table S3; Figure 4), being highest in pair A and lowest in pair B, following the same ranking (i.e., pairs A, E, D, C, B from highest to lowest values).
No extinction, defined as two weeks without eggs collected in oviposition containers and no adult females collected with BG-Sentinel Mosquito Traps was detected in any treatment cage during the 17 weeks post-release. The low effectiveness of the OX3604C strain in all treatment cages is supported by high OX3604C fitness cost estimates calculated with a simulation model for each treatment cage (Table 1). Model predictions based on extrapolation from the 17 weeks of data indicate that mosquitoes in all but one cage had at most an 11% chance of extinction within 10 weeks from the day the field cage experiment was terminated (probability of extinction for cage 1 was 0.36). On average, predicted extinction times for cages ranged between week 23 and week 65 post-release (Table 2).
Results of mating competition experiments are summarized in Table 3. Mating competition experiments 1–5, which all used a 1∶1 OX3604C∶GDLS2 ratio, but varied in cage size and location, produced variable results as indicated by the significant heterogeneity statistic in the replicated G-test (Het. G = 38.04, p<0.01; see Table S4). However, even with this heterogeneity value, the pooled G-value (4.124) is significant (p = 0.042) and indicates a mean mating fitness cost of 10.8% for the OX3604C males (Table S4). Experiment 6 was designed to better reflect conditions in the field cage experiment with a 10∶1 OX3604C∶GDLS2 ratio and the same cages and initial density as the long-term experiments. Results of the replicates also were variable (Het. G = 20.72, p = 0.04) and the overall difference between the strains was significant (Pooled G = 43,83, p<0.01) with a mean mating fitness cost of 59.1% for OX3604C males (Table S4).
OX3604C males decreased, but did not eliminate target populations in the field-cage experiment. Experimental protocols were similar between the previous laboratory experiment [18] and our field cage experiment, including high 10∶1 release ratios that in both cases were expected, based on modeling, to favor population extinction. Over the course of the field cage study, the OX3604C∶target male ratio increased in all cages and reached the highest value of 1,000∶1 in cage 1 during week 17 post-release, but this did not result in extinction of the target population. These results also were consistent with the data on the frequency of the DsRed2 marker detected in larvae from treatment cages, which ranged from 20–54% during week 17 post-release when the trial was terminated. The high estimated OX3604C male fitness cost calculated from the field cage trial (97%), the lower than predicted population reduction, and the long estimated extinction times (average 23–65 weeks) lead to the expectation that male OX3604C may be less effective for population reduction under open field conditions than predicted from results of the laboratory cage experiment.
Output from a simulation model predicted that the lack of a homozygous OX3604C strain did not contribute significantly to the absence of population extinction, because of the high OX3604C∶target male release ratio. Presence of wild-type individuals in the transgenic population required that we manually sex pupae in order to avoid introducing wild-type females into treatment cages. This resulted in a difference in management of the transgenic strain relative to the target strain and relative to the handling of the transgenic strain by Wise de Valdez et al [18]. The mortality observed among the transgenic males before introduction into treatment cages was low (∼5%), suggesting that this additional handling did not cause substantial harm. Furthermore, transgenic male survival in the cages was high. As can be seen in Figure 3, there were on average about 5 times as many males in treatment cages as control cages on day six after each release of OX3604C. This is equivalent to a daily survival rate of ∼0.91, which is similar to published values for Ae. aegypti survival in houses in the field [24]. The potential effect of differential handling was not addressed directly by the mating competitiveness experiments because the need to separate males from females added an additional handling step for GDLS2 males and females, such that the treatment of the two types was matched more closely in these experiments than in the earlier field cage experiment. Because differences were not apparent in the field cage trial, male survival was not evaluated in mating competitiveness experiments.
Mosquito size is sometimes but not always [25] associated with fitness. For Ae. aegypti, Ponlawat and Harrington [26] reported greater mating success by larger than smaller males. Measurements made in the first two weeks of the field cage experiment indicated that OX3604C males (median wing length = 2.24) were slightly larger than GDLS2 males (median wing length = 2.17, Mann-Whitney U test p<0.01), so this is unlikely to have contributed to the field cage outcomes.
Results from mating competitiveness experiments with laboratory-reared mosquitoes in field cages have generally been found to underestimate fitness costs found when the same types of mosquitoes are released in the field [27]–[29]. While the mating disadvantage of OX3604C males observed in experiment 6 (59.1%) appears to be one factor explaining the lack of extinction in our field cage trial, if there had only been a fitness cost of 59%, some extinctions would have been expected (Figures S4, S5). Therefore, although short-term mating competitiveness experiments are useful in assessing one major component of fitness, they are not designed to measure as many aspects of fitness as are measured in long-term studies.
Adaptation of genetically-engineered mosquitoes and target populations to laboratory and field cages always needs to be taken into account when moving from the laboratory to the field. In the field cage trial, OX3604C was derived from introgression of the OX3604C construct into a GDLS1 genetic background, reared in laboratory conditions (i.e., stable temperature, relative humidity, and photoperiod) and mated in small, crowded laboratory cages for more than 20 generations, all of which potentially selected for capacity to mate in a small spaces, and other adaptations for increased fitness in a laboratory environment. Conversely, GDLS2 originated from mosquitoes collected from the same locations 2 years after those used to create GDLS1, and GDLS2 target populations were maintained in large outdoor field cages for 16 generations before the start of the experiment. During the prerelease period, GDLS2 populations experienced natural variation in daily temperature and relative humidity and mated successfully for ∼4 months in their large outdoor enclosures. Adaptation to field cages may have been an advantage for GDLS2 males when competing with OX3604C males for GDLS2 females.
Our results are consistent with those from previous mosquito studies [27], [28], [30] indicating that colony maintenance and mass rearing should be planned prior to field-cage or open-field trials. Rearing large numbers of transgenic mosquitoes in large outdoor, semi-field enclosures for several generations may help avoid undesirable laboratory adaptation and reduce fitness differences between transgenic mosquitoes and conspecifics in their natural, target populations. Short-term mating competition experiments in large field cages could be an efficient way to gather preliminary information on genetically-engineered mosquito fitness relative to local wild-type mosquitoes, but they only measure one important fitness component while field cage trials include additional components.
We emphasize the potential impact of differential strain adaptation to the field or laboratory, but it is also possible that the fitness difference was due to the transgenesis process. Although insertion of the transgene did not affect the ability of the OX3604C to cause extinction in the laboratory system, it is feasible that some negative pleiotropic effect of the gene insertion was manifested only under field cage conditions. Precautions were taken to avoid some negative effects that are often associated with transgenesis. Most importantly, the originally engineered strain was backcrossed for five generations to a strain for the local area where the experiment was conducted. This was expected to replace over 96% of the genes from the engineered strain with local strain genes, except for genes linked to the transgene. If the transgene had been inserted within a transcribed gene, it could have disrupted gene function that affected fitness in outdoor field cages, but not in the laboratory. Attempts to fine-scale map the location of the transgene indicated that the insertion was in a genomic area with repetitive DNA, indicating the transgene was not inserted within a transcribed gene.
Although an argument can be made for not pursuing an open field evaluation of OX3604C males based on our field cage results, the best way to resolve the discrepancy between laboratory and field cage results would be to assess them under uncontained, open field conditions. Because this has not been done, data do not exist to determine whether laboratory or field cage experiments are most informative about how this strain will perform under natural conditions. A different genetic background, different chromosomal location of the transgene or different rearing procedures could separately or in combination affect the competitiveness of transgenic mosquitoes. Evaluation of other Ae. aegypti strains carrying the female-flightless transgene would help determine if results observed in this trial apply to this genetic modification in general or are specific to the OX3604C strain we studied.
Our results support inclusion of large outdoor field cage experiments in the systematic, phased evaluation of GE Ae. aegypti, including those with transgenes like OX3604C that are self-limiting. Details of field cage construction and the level of containment needed will depend on the nature of genetic modification in the strain being evaluated as well as general requirements of the relevant regulatory authorities. If genetic modifications include the potential of elevated pathogen transmission or non-Mendelian inheritance (i.e., genetic drive systems), strain evaluation will require higher security caging than those used in our experiments [20]. We stress that short-term mating competition experiments in large field cages could be used to obtain predictive information on mating competitiveness and fitness costs, but it is not clear that by themselves these would be sufficient substitutes for longer-term field cage tests. Results of appropriately planned, executed and analyzed open-field releases of the OX3604C would be useful in addressing this issue.
All of the work described here was conducted within ethical, social and cultural guidelines for community engagement activities [16]. We found that this approach helped us to develop respect and trust, basic ingredients for strong working relationships with local residents living near the field site, and for appropriate dialogue with state and national health and environmental authorities, scientists, and local and international press. Although the containment measures and communication activities taken in this work were greater than expected for research with natural strains of mosquitoes, we feel that this precautionary approach could have long-term benefits by decreasing suspicion that transgenic mosquito technology is being applied carelessly [31].
|
10.1371/journal.pbio.0060178 | Proteomics Reveals Novel Drosophila Seminal Fluid Proteins Transferred at Mating | Across diverse taxa, seminal fluid proteins (Sfps) transferred at mating affect the reproductive success of both sexes. Such reproductive proteins often evolve under positive selection between species; because of this rapid divergence, Sfps are hypothesized to play a role in speciation by contributing to reproductive isolation between populations. In Drosophila, individual Sfps have been characterized and are known to alter male sperm competitive ability and female post-mating behavior, but a proteomic-scale view of the transferred Sfps has been missing. Here we describe a novel proteomic method that uses whole-organism isotopic labeling to detect transferred Sfps in mated female D. melanogaster. We identified 63 proteins, which were previously unknown to function in reproduction, and confirmed the transfer of dozens of predicted Sfps. Relative quantification of protein abundance revealed that several of these novel Sfps are abundant in seminal fluid. Positive selection and tandem gene duplication are the prevailing forces of Sfp evolution, and comparative proteomics with additional species revealed lineage-specific changes in seminal fluid content. We also report a proteomic-based gene discovery method that uncovered 19 previously unannotated genes in D. melanogaster. Our results demonstrate an experimental method to identify transferred proteins in any system that is amenable to isotopic labeling, and they underscore the power of combining proteomic and evolutionary analyses to shed light on the complex process of Drosophila reproduction.
| Across many species, males transfer both sperm and seminal proteins to their mates. These proteins increase male reproductive success by improving sperm competitive ability and modifying female behavior. In Drosophila, seminal proteins increase female rates of egg-laying and sperm storage and reduce a female's willingness to mate with subsequent suitors. Several male seminal proteins have been extensively characterized, and others have been predicted based on gene expression patterns, yet the full set of proteins that is transferred to females has not been defined. Here we introduce a new proteomic method that identifies transferred seminal proteins in recently mated females and quantifies their relative abundance. We confirm many of the predicted seminal proteins and discover a number of novel seminal fluid components. Some of these proteins show elevated rates of evolution, consistent with their involvement in sexual selection or sexual conflict, and many have arisen by tandem gene duplication. By using this method in three species of Drosophila, we identified lineage-specific components of seminal fluid. Additionally, we developed and validated a method to identify completely new genes in the D. melanogaster genome. These transferred proteins are now targets for follow-up genetic, biochemical, and evolutionary analysis.
| In addition to sperm, males of many internally fertilizing species transfer seminal fluid proteins (Sfps) to their mates during copulation. These proteins function in a variety of reproductive processes, including sperm capacitation, sperm storage and competition, and fertilization, and in some cases they affect female behavior and physiology [1]. Like other reproductive proteins, Sfps often evolve rapidly between species, underscoring their relevance to reproductive success [2]. Sfps are thought to interact with several classes of molecules, including other Sfps (which may originate from seminal fluid of the same male or from a competitor), proteins native to the female reproductive tract, and pathogens that may be introduced during the course of mating. These interactions create opportunities for coevolution, leading to speculation that sperm competition, sexual conflict, sexual selection, and/or host–pathogen interactions could drive the rapid, adaptive evolution of many Sfps [3]. Because of their rapid evolution and their critical importance to reproductive fitness, Sfps may also be involved in the formation of new species [3–5]. As such, researchers have sought to identify and characterize Sfps in such diverse taxa as mosquitoes, crickets, honeybees, rodents, and primates [6–10].
Although Sfps are being studied in many species, they are best characterized in Drosophila melanogaster. Because the Drosophila mating system features multiple matings by females and sperm competition between males [11–14], Sfps are thought to be especially important for reproductive success and for mediating conflict and competition. Previous studies have focused on three areas: (a) the effects of the full set of Sfps (and especially of accessory gland proteins, or Acps) on male and female fitness; (b) identification of putative Acps by expressed sequence tag (EST) sequencing, comparative genomics, and proteomics; and (c) functional analysis of specific Acps. Whole-organism work revealed that Acps mediate a “cost of mating” to females. The repeated receipt of Acps through multiple matings lowers female reproductive output by reducing female life span without a corresponding increase in egg production [15–17]. Furthermore, when a population of males harboring natural genetic variation was allowed to adapt to a static female genotype, male sperm competitive ability, mating success, and harm caused to females increased in 30–40 generations [18]. These dramatic evolutionary outcomes sparked much interest in identifying the specific proteins of seminal fluid. Screens for genes expressed specifically in the male accessory glands (and encoding proteins predicted to be secreted) identified ∼70 putative Acps [19,20], several of which have been genetically or biochemically characterized (reviewed in [21]). Work in related Drosophila species has revealed that many predicted Acps are subject to lineage-specific gene gain, gene loss, and/or copy number variation [22–24]. Additionally, several proteomic studies have examined both Acps and sperm proteins found in males [25,26], and whole-genome, tissue-specific microarray analysis has increased the number of predicted Acps to 112 [21,27].
In spite of this considerable progress, less than one-third of the predicted D. melanogaster Sfps have been detected in mated females [21,28]. Furthermore, prior work to predict Sfps has often required that candidate genes show tissue-specific expression in the male reproductive tract. Identifying the set of transferred Sfps in an unbiased fashion is of critical importance, since it is these proteins that are the most likely to influence post-mating processes like sperm competition and sexual conflict. We have developed a mass spectrometry (MS) method that specifically detects male Sfps in mated female D. melanogaster. In addition to confirming the transfer of many predicted Sfps, we identified dozens of new seminal fluid components, including completely new classes of proteins. Evolutionary analyses show that positive selection and tandem gene duplication drive the evolution of seminal fluid between species, and comparative proteomics with additional species identified lineage-specific Sfps. We also used our MS data to estimate the relative abundance of each Sfp in seminal fluid and to discover previously unannotated genes encoding additional Sfps. Taken together, our experiments illustrate the power of combining proteomics with evolutionary biology to address fundamental questions about reproduction.
To distinguish between transferred Sfps and proteins native to the female reproductive tract, we metabolically labeled female flies using a diet enriched in 15N isotopes to create an isotopically “heavy” form of the female proteins [29]. Females were reared on yeast that was grown in media enriched in 15N. After one full generation of labeling, the 15N enrichment in detected fly peptides was ∼98 atom percent excess, and no peptides from whole female flies were identified with natural abundance nitrogen isotopes. These data confirm that isotopic labeling can be readily achieved in D. melanogaster and other drosophilids (see below). Therefore, we reasoned that by mating unlabeled males to labeled females and then analyzing proteins found in the female reproductive tract by MS, transferred male Sfps could be identified by those peptides that showed natural abundance isotope distributions. We chose to label females instead of males, because MS resolution is best for unlabeled peptides, and we were interested in identifying male Sfps.
We performed multiple biological replicates of mating experiments with different strains of males: Canton S (a standard lab strain), sons of tudor females (spermless males) [30], and, as a negative control, DTA-E males (which are spermless and do not produce main cell accessory gland proteins) [31]. In two DTA-E experiments, we detected 11 transferred proteins, including several known to be produced outside of the accessory glands (Table S1). Six total experiments with Canton S and tudor males identified a set of 138 high-confidence Sfps (Table 1 and Table S2), using a peptide-level q-value ≤ 0.01 within each experiment [32]. Just over half (75/138) of the transferred Sfps were previously predicted through tissue-specific expression profiling or other experimental or comparative genomic methods [21], but only 19 were confirmed previously to be transferred at mating. Notably, we found only five previously documented sperm proteins [25], confirming that our protein preparation protocol effectively selected for soluble, extracellular proteins. We did not detect 49 predicted Sfps [21]. These proteins may be transferred at low levels, immediately cleaved or degraded in the female, or have certain peptide sequences or post-translational modifications that complicate detection by shotgun proteomics. Alternatively, some may not be transferred at mating.
We identified 63 novel Sfps, 45 of which were found in at least two biological replicates. Many of these proteins fell into the same functional categories as the previously predicted set, including proteases, protease inhibitors, mediators of an immune response, and proteins involved in lipid metabolism (Table 1). We discovered several new classes of proteins among the transferred Sfps. Most intriguing were six members of the odorant binding protein (Obp) family [33]. Obps are thought to shuttle small molecules through aqueous solutions by binding them in a small, hydrophobic pocket; they are traditionally associated with the olfactory nervous system [34]. We confirmed that these Obps are transferred in seminal fluid by performing MS on protein digests from dissected accessory glands and by confirming each gene's expression in the male reproductive tract with FlyAtlas [27].
Reproductive proteins of diverse species often evolve under positive Darwinian selection, which may indicate involvement in a coevolutionary process such as sexual selection, sexual conflict, or host–pathogen recognition [2]. We used coding sequence alignments from the 12 Drosophila genomes project [35,36] to calculate the rates of nonsynonymous substitution (dN) and synonymous substitution (dS) for all Sfps for which an ortholog was identified (116 of the total 138). For each Sfp, we determined the whole-gene, pairwise dN/dS (ω) ratio between the D. melanogaster gene and an ortholog from a closely related species (Figure 1). By this conservative test, five Sfps showed evidence of adaptive evolution (ω > 1). However, prior studies have shown that when the whole-gene pairwise ω ratio exceeds 0.5, or when the nonsynonymous substitution rate (dN) is elevated, there are often specific sites within the protein for which adaptive evolution can be detected with more sensitive methods [7,37]. Therefore, we used multiple species alignments to search for specific residues under selection for all genes with pairwise ω > 0.5 and/or pairwise dN > 0.05. (We did not test all Sfps, in order to minimize the number of statistical tests.) We found evidence for adaptive evolution at specific sites for 16 of 36 proteins (Figure 1 and Table S3), including four proteins that were unidentified previously as Sfps. Nine of these tests for selection remained significant after applying a strict Bonferroni correction for multiple tests. These rapidly evolving proteins are attractive targets for future study.
Previous studies found that some predicted Sfps are clustered throughout the genome [21,24]. We examined the chromosomal locations of the transferred Sfps and found similar patterns. We defined a cluster as genes with start codons located within 10 kb of each other. We identified 19 clusters of 2–5 transferred Sfps, which contain one-third (46/138) of the detected Sfps (Figure 2A). For 17 clusters, all member genes are transcribed in the same direction, and 15 clusters contain genes that encode proteins with full-length homology to one another. Thus, most of the observed clustering can be attributed to tandem gene duplication. Four paralogous clusters contain at least one gene that was under selection in the sites analysis above. Previous work also found a dearth of Acps on the X chromosome. Consistent with this finding, the 13 transferred Sfps on the X chromosome were significantly fewer than would be expected by chance (χ2 = 4.68, 1 degree of freedom [df], p = 0.03), given the proportion of annotated genes on the X.
One example of rapidly evolving tandem duplicates is the gene pair CG17472 and CG31680. Across five species, CG17472 has evolved adaptively, with 21.3% of sites predicted to be under positive selection (estimated ω = 3.36, PAML M8 versus M8a comparison: χ2 = 15.38, 1 df, p < 0.0001). These duplicates flank a transposition hot spot and a third, pseudogenized copy of the locus (Figure 2B). Additionally, CG17472 has duplicated along the lineage leading to D. simulans and D. sechellia (Figure 2C). Examining the ω ratio on each branch of the phylogeny reveals a burst of positive selection on the CG31680 lineage immediately after duplication. Indeed, a branch model allowing for variable selective pressures along each branch (shown in Figure 2C) fit the data significantly better than a model with a uniform ω for all branches (χ2 = 29.04, 14 df, p = 0.01).
We used our MS data to estimate the relative abundance of each Sfp in seminal fluid. By counting the number of spectra associated with each Sfp in a given experiment and standardizing by the length of the protein and the total number of Sfp spectra detected in the experiment, we calculated a normalized spectral abundance factor (NSAF) [38,39] for each protein, which could then be averaged across all experiments (Figure 3 and Table S2). Notably, NSAF values were positively associated with the number of biological replicates in which a protein was found (Figure 3). Several of the most abundant proteins were previously characterized Sfps, such as Acp62F (a protease inhibitor) and Acp70A (the sex peptide). However, several novel proteins were also in the top quartile for abundance, including Obp56f, Obp56g, and the tandem duplicate CG17472. Although these NSAF measurements are only approximate, these data provide the first proteomic-scale view of the relative amount of each transferred Sfp, which may be useful for selecting candidates for further investigation.
To examine the cross-species evolution of seminal fluid content, we used the predicted protein annotations of D. simulans and D. yakuba [35,36] to repeat our mating experiments with a wild-type strain of each species (Figure S1). Of the 63 Sfps detected in all three species, 19 were not reported previously as seminal fluid components. For Sfps that were detected in only one or two species, we investigated whether these proteins could be called as either lineage-specific gene gain or loss events. Most of the proteins had identifiable orthologs in the other species; our failure to detect these proteins may be due to changes in expression patterns, sequence substitutions that render MS identification more difficult, changes in the amounts of proteins transferred at mating, or the lower number of replicates (two per species) performed for D. simulans and D. yakuba. However, our data identify 13 lineage-specific Sfps across the three species (Table S5). For example, in D. melanogaster, CG6289 (a predicted serine protease inhibitor) has duplicated to form the lineage-specific gene CG6663. Also, in D. yakuba, Acp76A (another serine protease inhibitor) has duplicated, and several other proteins appear to be either lineage-specific to D. yakuba or rendered nonfunctional in other species (Table S5).
Some proteins detected for D. simulans and D. yakuba lacked annotated orthologs in D. melanogaster. For seven such proteins, we identified the syntenic region in D. melanogaster and performed reverse-transcriptase PCR (RT-PCR) to determine whether transcripts of the region were made. In five cases, we detected a transcript in D. melanogaster (see Table S5), and three of these putative loci were detected as proteins in D. melanogaster when searching for unannotated proteins in the D. melanogaster genome (see below). Curiously, one of these genes, which we have annotated as Sfp53D, showed male-specific expression in D. yakuba and male-biased expression in D. simulans, but no sex expression bias in D. melanogaster (data not shown). Sfp53D is therefore an example of the type of protein that would have been omitted from previous sets of predicted Sfps due to its lack of sex-specific expression.
Based on these results, we reasoned that other Sfps may not be annotated as genes in D. melanogaster, which would make them impossible to detect by searching mass spectra against the annotated proteome. To detect additional unannotated Sfps, we first constructed a six–reading frame translation of the D. melanogaster euchromatic genome, which produced >5.8 million potential open reading frames (ORFs). Then, to reduce computational search time, we applied the Hardklör algorithm [40] to predict which MS2 spectra from a tudor experiment came from male peptides containing only natural abundance isotopes. These spectra were searched against the six-frame database, and those that matched an ORF corresponding to an annotated protein were discarded. This procedure left 23 novel, putative ORFs that did not match any D. melanogaster gene annotation in FlyBase. For each putative ORF, we used rapid amplification of cDNA ends (RACE) and RT-PCR to confirm transcription of the region encoding the peptide and to define the full-length transcript. Through this method, we discovered 19 unannotated genes (Table S6; GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) accession numbers EU755332–EU755350), most of which showed no significant identity to the predicted proteins of the other sequenced Drosophila species.
All 19 proteins have predicted signal sequences for secretion; many are encoded by only one or two exons, and all produce short polypeptides (median length: 93 residues). Consistent with our clustering analysis, half of the genes were found in regions of the genome containing other annotated Sfps. Most of the novel proteins had no recognizable domains based on BLAST and structural homology searches, but we identified one C-type lectin and three enzyme inhibitors, including a putative protease inhibitor, Sfp24Ba. This protein was identified by three peptides, one of which is indicated in Figure 4A. Sfp24Ba is adjacent to another previously unannotated gene, Sfp24Bb (an apparent tandem duplicate), and lies 25 kb upstream of the gene that encodes a transferred protease inhibitor, Acp24A4 (Figure 4B). Comparative structural modeling (Figure 4C) suggests that this protein is a Kunitz-type protease inhibitor. The discovery of these 19 new Sfp genes in a model system that has been studied for over a century and for which comparative genomic analysis is now straightforward underscores the limitations of both computational gene prediction programs and the “whole-proteome” databases that are routinely used during shotgun MS analyses.
Our study provides a proteomic-scale view of the transferred Sfps in D. melanogaster. While we confirmed that 75 predicted Sfps are truly transferred at mating, we also identified a total of 82 genes (63 already annotated, 19 newly discovered) previously unknown to encode seminal fluid products. By using data from the genome sequencing projects and by performing comparative experiments in D. simulans and D. yakuba, we identified many instances of positive selection, tandem gene duplication, and lineage-specific changes in seminal fluid content between species. Taken together, our experiments demonstrate how new proteomic methods can be combined with the vast amounts of genomic sequence data that are now available to gain considerable insight into the molecular players of a specific biological process.
The two methodological advances presented here—the use of isotopic labeling to distinguish between the sexes, and searching MS data against an entire translated genome—should be applicable to many taxa. For example, worms, plants, rodents, and microorganisms are all amenable to isotopic labeling [29,41,42]. In any of these systems, differential labeling should readily allow the detection of proteins transferred from one organism (during mating or another behavior, e.g., courtship). Thus, our approach allows transferred proteins in a pre-specified biological process to be identified. Furthermore, our MS- and RACE-based method to identifying novel genes should be applicable to other organisms with sequenced genomes, particularly if their genome sizes are no more than 1–2 orders of magnitude greater than the D. melanogaster genome. Rodents, Arabidopsis, and humans all fall within this range; indeed, recent work in A. thaliana has found new genes using a similar approach [43]. Our results confirm that searching MS data against an entire translated genome, rather than only an annotated or predicted proteome, can identify a considerable number of new genes. Admittedly, this process might have been particularly useful for identifying Drosophila Sfps. As shown here and in previous work, these proteins are short, rapidly evolving, and relatively free of codon bias [19,22], three features making them less likely to be detected by computational gene prediction programs. Nonetheless, our gene identification method was straightforward to perform, and because it was experimentally based, it offered automatic verification for the new genes and allowed us to immediately assign them to a specific biological process: male reproduction.
One striking result from these experiments is that seminal fluid content in D. melanogaster appears to be considerably more complex than was previously predicted. The Obp genes identified reproducibly and at high abundance by MS are particularly attractive targets for further characterization. One hypothesis for the origin of these reproductive Obps is based on the fact that Obps are a large, 51-member family in D. melanogaster [33]. If some members of this family were functionally redundant, selection on the regulatory and coding sequences of some Obps might have been relaxed, allowing them to be co-opted from an olfactory function into a male reproductive function. Indeed, several of the identified Obps show accessory gland–specific patterns of expression, while others are expressed in both the accessory glands and the head [27]. The function of these reproductive Obps remains to be determined; they may present odorants or pheromones to odorant receptors in the female reproductive tract or play some other role, such as transferring small molecules to the female to elicit a behavioral response. If some of the Obps interact with a receptor in the female tract, the Or10a odorant receptor is one possible target, since its expression is up-regulated in female reproductive tracts in response to the receipt of Sfps [44].
While the selective pressures driving the evolution of Sfps (and of reproductive proteins in general) remain unclear, the important roles of tandem gene duplication and positive selection in the evolution of Sfps are consistent with the predictions made by models of sexual selection/conflict [45]. If males are engaged in a coevolutionary chase with females, driven by sexual selection or sexual conflict, duplication of an Sfp locus could allow males to better adapt to a particular allele or paralog of a female receptor [45]. Indeed, gene duplication followed by positive selection has been observed previously in a well-characterized reproductive protein, lysin, which allows abalone sperm to penetrate the egg vitelline envelope [46]. If Drosophila Sfps are coevolving with receptors in the female reproductive tract—or with other Sfps with which they interact—then gene duplication may be an important evolutionary strategy for males to increase their reproductive success. Tests of this hypothesis will require both functional data on the newly identified male proteins and the identification of their interacting female and/or male partners [47–49].
The rapid divergence characteristic of many Sfps has generated considerable interest in their potential role in speciation [3,4]. If proteins mediating processes such as sperm storage, fertilization, or post-mating behavior diverged quickly between allopatric populations, driven continually by coevolutionary forces such as sexual selection or sexual conflict, between-population matings may become less productive than within-population matings. Such a difference could exert pressure to further differentiate the mating systems or mating behaviors of each group, which could eventually lead to the formation of distinct species. Determining the transferred Sfps, and subsequently identifying their functions and evolutionary patterns, could therefore be important steps in identifying potential “speciation genes.”
In conclusion, this set of transferred proteins provides a rich resource for investigating long-standing evolutionary questions and for identifying the specific molecules and functional allelic variants that affect both sperm competition and male-female coevolution and conflict. The challenge ahead will be to apply the combination of genetic, biochemical, and evolutionary methods that have already yielded many insights into Drosophila reproduction to this novel collection of transferred proteins. Functional tests of individual Sfps are essential for understanding the causes of the dramatic post-mating changes in female behaviors. For example, several studies have used gene knockouts or RNA interference to identify the post-mating effects of specific Sfps [50–55]. Other experiments have associated naturally occurring variants in several Sfps with different measures of sperm competition [12,56]. We expect that both of these approaches will become more effective in the future, since they can now be targeted to those transferred Sfps identified here.
Fly stocks were maintained on standard media at 25 °C, except during isotopic labeling (see below). D. melanogaster stocks included a wild-type lab strain, Canton S, and the strain used for genome sequencing, y; cn bw sp. To produce spermless males, homozygous tud1 bw sp females [30] were mated to either Canton S or y; cn bw sp males, and male progeny were retained for use in mating experiments. The DTA-E stock was used to produce males lacking both sperm and main-cell accessory gland proteins [31]. D. simulans strain W89 and D. yakuba strain Tai6 were used in additional mating experiments.
The isotopic labeling procedure followed a previously described method [29], with some modifications. Wild-type Saccarhomyces cerevisiae was grown to saturation in minimal media containing 2% glucose, yeast nitrogen base without amino acids and ammonium sulfate (Difco), and 15N-labeled ammonium sulfate (≥ 99% 15N-enrichment; Spectra Stable Isotopes). Yeast cells were pelletted, resuspended in a small volume of sterile water, and lyophilized. This dried yeast was then mixed with water to form a “heavy” (15N) yeast paste. Flies were isotopically labeled by allowing unlabeled females to lay eggs for 24–36 h onto an agar plate topped with a small amount of heavy yeast paste. Adults were then discarded, and eggs were allowed to develop to adulthood at 25 °C in a vial capped at the open end by the plate. Heavy yeast paste was added to the plate throughout development as the sole food and nitrogen source. Virgin females were collected over CO2 within 8 h of eclosion and stored in a separate vial with 15N yeast paste on an agar plate. Shotgun MS analysis of proteins from whole, first-generation 15N flies was used to confirm isotopic labeling. In parallel to 15N labeling, males of the strain to be tested were grown in standard vials. Males were collected while young (0–3 d old) and aged in isolation in a standard vial.
We performed 12 total mating experiments: three biological replicates each of Canton S and tudor males, and two biological replicates each of DTA-E, D. simulans, and D. yakuba males. For each experiment, males and females were aged to 2–5 d before mating. On the day before an experiment, approximately 40 labeled, virgin females were divided into three vials containing agar with a small amount of heavy yeast paste. Unlabeled males, in a ≥1.5-fold excess relative to females, were placed into three standard vials. The next day, males were transferred to the female vials without anesthesia. Mating was allowed to proceed for 2 h; vials were inspected several times during this period to confirm that copulations occurred. At the end of the mating period, flies were sexed over CO2: males were discarded, while females were kept on ice and immediately dissected in 50 mM ammonium bicarbonate. The lower female reproductive tracts were retained and stored in cold ammonium bicarbonate, while ovaries were excluded to prevent saturating the protein sample with the highly abundant egg yolk proteins. (If ovaries had been included, a greater fraction of peptides identified by MS would have arisen from these female proteins, making it more difficult to detect peptides from lower-abundance male Sfps.) It is unlikely that the removal of the ovaries diminished our ability to detect certain Sfps, as we identified all five Sfps (Acp26Aa, Acp36DE, Acp62F, msopa, and Spn2) that had been shown previously to localize to the ovaries [28].
Because we sought to identify soluble, extracellular male Sfps, proteins were prepared in such a way so as to select specifically for soluble proteins. We also sought to reduce cell lysis and thus protein content from male sperm cells and female reproductive tract epithelial cells, since releasing intracellular proteins from these cells would dilute the concentration of transferred Sfps and render their identification more difficult. Female reproductive tracts were homogenized in the ammonium bicarbonate dissection buffer, which lacks any type of detergent and thus minimized cell lysis. The mixture was then centrifuged for 5 min at 18,000g. This process was repeated once, and the supernatant was retained. Protein concentration was estimated using a BCA assay (Pierce). Proteins were prepared for tandem mass spectrometry and digested with trypsin as previously described [57].
Two samples each of Canton S and tudor, and one sample each of DTA-E, D. simulans, and D. yakuba, were analyzed by multi-dimensional protein identification technology (MudPIT) [58]. Protein digests (50 μg) were bomb-loaded overnight onto a tri-phasic 100-μm internal diameter capillary column packed with 15-cm reversed phase material (Jupiter C12, 4 μm, 90 Å; Phenomonex) at the tip of the column, then 4 cm of strong cation exchange material (Whatman), then 3 cm more of C12 material. The columns were then placed on-line with either an LTQ ion-trap mass spectrometer (ThermoElectron) or an LTQ-FT Ultra mass spectrometer (ThermoElectron) and eluted over a 12-step gradient with increasing salt concentration as described previously [59]. We also analyzed additional samples using a single reversed-phase HPLC method. One sample each of Canton S, tudor, DTA-E, D. simulans, and D. yakuba was analyzed with 75-μm internal diameter capillary columns packed with 40 cm of Jupiter C12 reversed phase material. For each sample analyzed by reversed phase, four or five technical replicates of ∼6 μg of protein were analyzed by injecting the sample directly into an on-line column and running four-hour gradients to acquire mass spectra using data-dependent acquisition.
Tandem mass spectra from each RAW mass spectrometry data file were extracted from the proprietary data format and stored in the MS2 file format [60] using in-house developed software. The charge-state of multiply charged MS/MS spectra were assigned a single +2 and +3 charge state using the charge-czar program [61] and searched against two databases using Sequest [62]. One database contained the annotated proteome of the appropriate species; the other database contained a set of “decoy” proteins, made by randomly shuffling the amino acids in each protein of the annotated database. Each database also included common contaminants (or their shuffled counterparts). For D. melanogaster samples, the proteome was taken from the version 4.3 release of the D. melanogaster genome (downloaded from NCBI; gene annotations and names were later updated to version 5.2). For D. simulans and D. yakuba, the GLEANR protein predictions from the 12-genome Drosophila sequencing project were used [35,36]. Because the GLEANR sets were likely imperfect, these species' databases were supplemented with the best hit (e-value cutoff = 0.01) obtained when the identified D. melanogaster proteins were searched using tblastn against the D. simulans or D. yakuba sequences in GenBank. After the database searches, the percolator program [32] was used to improve the discrimination between correct and incorrect peptide spectrum matches and to assign a q-value as a measure of the false discovery rate [63].
To determine the list of high-confidence Sfps in D. melanogaster, we used the following criteria. Proteins identified in at least two independent experiments were automatically included. For proteins identified in only one of the six Canton S and tudor experiments, we required additional evidence that the protein could plausibly be involved in reproduction. This criterion could be satisfied if a protein was included in the most recent and comprehensive set of predicted Sfps [21] and/or if the protein showed strong evidence of being expressed exclusively or predominantly in the male reproductive tract (accessory glands or testes) in the FlyAtlas dataset [27]. Because we performed fewer mating experiments (two per species) and had no genome-wide catalog of Sfps or expression data, it was necessary to use different criteria in defining the sets of transferred proteins in D. simulans and D. yakuba. For each species, all proteins found in both experiments were automatically included. Furthermore, we included proteins found in only one experiment if they met any of the following criteria: (a) at least two peptides were used to identify the protein in the experiments; (b) if a single peptide was used for identification, it was detected at least twice during the MS run; or (c) the protein was identified as a transferred or predicted Sfp in D. melanogaster. After determining the list of transferred proteins shown in Table S2, functional information was acquired by examining FlyBase and the primary literature and was used to classify proteins listed in Table 1. We classified a protein as a “sperm protein” if it was found at least twice in Canton S experiments but not in tudor experiments, and/or if it was previously documented as such [25]. We used BLAST and BLAT searches to determine whether any transferred proteins of each species could be called as lineage-specific (Table S5).
Genomic locations of Sfps were determined by downloading from FlyBase (release version 5.2) the chromosomal location of the first transcribed base of each gene, and recording the strand from which the gene was transcribed. Only euchromatic genes were considered and plotted, such that plots in Figure 2A do not indicate, for example, centromeric heterochromatin. Clusters were defined as genes that were within 10 kb of each other. For proteins encoded in a given cluster, we used pairwise BLASTP searches to determine whether the proteins showed evidence of paralogy. We used simulations to estimate a null distribution of the number of clusters that would be expected for a set of 138 genes distributed across the chromosomes in the same ratio as our Sfps. We extracted coding sequence annotations from http://www.flybase.org/ (version 5.2) and noted the location of the start codon for each gene (one isoform per locus). We then generated 1,000 sets of 138 genes by randomly selecting genes from each chromosome arm in the same ratio as the observed Sfps. The number of clusters in each set was counted; the median was 3, and the range was 0–9 clusters. Therefore, we judged our observed 19 clusters to be significantly more than would be expected by chance.
Several GLEANR-predicted proteins identified in D. simulans and D. yakuba lacked annotated orthologs in the version 4.3 release of the D. melanogaster genome (one has since been annotated as CG12828, and another is reported [64] in GenBank under accession number BK003861, but is not yet recorded in FlyBase). We thus tested whether these genes (GLEANR numbers: dsim_2617, dsim_3447, dsim_15012, dsim_10234/dyak_792, dsim_9514/dyak_14199, dyak_12348, and dyak_10591) were expressed in D. melanogaster and showed sex-specific expression. PCR primers were designed to amplify transcripts in both the species of identification and the syntenic region of D. melanogaster. Although several of these proteins were encoded by short, single-exon genes, primers were designed to span putative introns when possible. Total RNA was prepared from whole male and whole female flies of both species using the TRIzol reagent (Invitrogen) and subjected to rigorous DNase treatment using the Turbo DNase kit (Ambion). First-strand cDNA from each sex was synthesized using the SuperScript III kit (Invitrogen) according to the manufacturer's instructions. This cDNA was then diluted and used in PCR reactions. As a positive control, we assayed for transcription of ribosomal protein L32 (RpL32) using previously published primers [12], modified as needed for D. simulans and D. yakuba. Negative PCR controls were performed by using template from cDNA reactions that lacked reverse transcriptase.
For each D. melanogaster protein identified, we used coding-sequence alignments generated by the 12-species genome sequencing projects [35,36] to conduct molecular evolutionary analyses. We preferentially used the more recent Fuzzy Reciprocal BLAST-based alignments of D. melanogaster coding sequences with orthologs in any other species (ftp://ftp.flybase.net/genomes/12_species_analysis/clark_eisen/alignments); however, less than half of our Sfps were included in this set, so for the others we used the comparative assembly freeze 1 (CAF1) GeneMapper alignments produced by S. Chatterji and L. Pachter. From these sources, we were able to analyze 116 of the 138 annotated Sfps from D. melanogaster. We first made pairwise estimates of dN/dS with model M0 of codeml in the PAML package [65]. When available, we used the D. simulans ortholog; otherwise, the D. sechellia ortholog was used. Alignments were obtained from one of the above sources and checked by eye using MEGA 4.0 [66]. For genes with pairwise dN/dS ≥ 0.5 or dN ≥ 0.05, we expanded our PAML analysis to up to five species (melanogaster, simulans, sechellia, yakuba, and erecta) in order to search for specific sites likely to have evolved under positive selection. For each gene, we used only those species for which alignments were reliable, and coding sequence alignments were checked by eye and edited in MEGA 4.0. We then tested for positive selection by comparing the likelihoods of codeml models M8 and M8a with a likelihood ratio test [67]. In model M8a, each codon is assigned to one of 11 classes, ten of which have an ω (dN/dS) value between 0 and 1 that is estimated from the data using maximum likelihood, and the 11th of which has ω = 1, representative of neutral evolution. Model M8 differs in that the 11th class of codons can take any ω value; this value is estimated from the data and can be greater than 1 (which indicates adaptive evolution). We corrected for multiple testing with a strict Bonferroni correction, though we note that among the 36 tests performed, only ∼2 would be expected to be false positives at a critical p-value of 0.05. As shown in Figure 2B and 2C, we re-analyzed CG17472 and CG31680 by including both paralogs from D. simulans and D. sechellia. The M8 versus M8a test of CG17472 in the results section contains all orthologs (including both duplicate copies in D. simulans and D. sechellia), but not CG31680 and its D. sechellia ortholog. For Figure 2C, we used the dnaml program in PHYLIP [68] to construct a phylogeny and to simulate 1,000 bootstrap replicates. We then used PAML to estimate ω for each branch of the phylogeny and performed a likelihood ratio test to compare the likelihoods of a model that allowed for ω to vary on each branch of the phylogeny versus a null model in which a uniform ω was estimated across all branches [69].
Relative protein abundance was estimated from D. melanogaster MS data by counting the number of spectra that positively identified each protein in a given MS run. This spectral count was normalized for the length of each protein and divided by the sum of all normalized counts for the entire MS run to produce an NSAF for each protein, as previously described [38,39]. This value was then averaged across all experiments in which a protein was detected, and identified proteins were ranked by their mean NSAFs. This rank should be interpreted as how common it was to identify ionizable and detectable spectra for a given protein, relative to the other unlabeled proteins.
To identify unannotated Sfps, we first used nr6frame (D. States, unpublished program) to make a six–reading-frame translation of the entire Berkeley Drosophila Genome Project D. melanogaster genome, version 5 (downloaded from ftp://hgdownload.cse.ucsc.edu/goldenPath/dm3/chromosomes). This program translates genomic DNA in all six reading frames; each reported ORF ends with a stop codon (but does not necessarily start with a methionine). Across the four Drosophila chromosomes, over 7.6 million ORFs were generated. We filtered these ORFs to exclude those that contained only one type of amino acid (mono-residue repeats), those that were too short to be confidently used in MS spectrum identification (<11 residues), or those that could not produce a tryptic peptide due to a lack of a K or R residue. This filtering reduced the data set to >5.8 million ORFs. For searching this large database, it was computationally advantageous to filter the MS2 files in order to reduce the search time. We used data from three technical replicates of a tudor sample, collected with a 40-cm reversed phase column on an LTQ-FT Ultra instrument. We used Hardklör [40] to predict the isotope distributions that resulted from 15N-enriched peptides and removed their corresponding MS/MS spectra from the analysis. Because of the excess of labeled peptides within the sample, this filtering reduced the number of spectra that needed to be searched by ∼86%. The remaining spectra were then searched against the six-frame translation database using Sequest [62], and identifications were filtered by DTASelect. Identified peptides matching annotated protein coding genes were discarded, leaving 23 ORFs that did not match a genome annotation. We designed primers in the genomic regions matching the identified peptides and performed 5′ and 3′ RACE to amplify transcripts from these regions (SMART RACE Kit, Clontech-Takara). This method identified 19 unannotated genes, which were then confirmed with RT-PCR and sequencing of cDNA from whole males. SignalP was used to predict whether each novel protein is secreted [70], and we used BLAST and PHYRE [71] searches to determine whether any protein had sequence or structural homology to other proteins.
Sequence data has been deposited in GenBank under accession numbers EU755332–EU755350.
|
10.1371/journal.pntd.0000127 | Quantitative High-Throughput Screen Identifies Inhibitors of the Schistosoma mansoni Redox Cascade | Schistosomiasis is a tropical disease associated with high morbidity and mortality, currently affecting over 200 million people worldwide. Praziquantel is the only drug used to treat the disease, and with its increased use the probability of developing drug resistance has grown significantly. The Schistosoma parasites can survive for up to decades in the human host due in part to a unique set of antioxidant enzymes that continuously degrade the reactive oxygen species produced by the host's innate immune response. Two principal components of this defense system have been recently identified in S. mansoni as thioredoxin/glutathione reductase (TGR) and peroxiredoxin (Prx) and as such these enzymes present attractive new targets for anti-schistosomiasis drug development. Inhibition of TGR/Prx activity was screened in a dual-enzyme format with reducing equivalents being transferred from NADPH to glutathione via a TGR-catalyzed reaction and then to hydrogen peroxide via a Prx-catalyzed step. A fully automated quantitative high-throughput (qHTS) experiment was performed against a collection of 71,028 compounds tested as 7- to 15-point concentration series at 5 µL reaction volume in 1536-well plate format. In order to generate a robust data set and to minimize the effect of compound autofluorescence, apparent reaction rates derived from a kinetic read were utilized instead of end-point measurements. Actives identified from the screen, along with previously untested analogues, were subjected to confirmatory experiments using the screening assay and subsequently against the individual targets in secondary assays. Several novel active series were identified which inhibited TGR at a range of potencies, with IC50s ranging from micromolar to the assay response limit (∼25 nM). This is, to our knowledge, the first report of a large-scale HTS to identify lead compounds for a helminthic disease, and provides a paradigm that can be used to jump-start development of novel therapeutics for other neglected tropical diseases.
| Schistosomiasis, also known as bilharzia, is a tropical disease associated with high morbidity and mortality, currently affecting over 200 million people worldwide. Praziquantel is the only drug used to treat the disease, and with its increased use the probability of developing resistance has grown significantly. The Schistosoma parasites can survive for up to decades in the human host due in part to a unique set of antioxidant enzymes that continuously degrade the reactive oxygen species produced by the host's innate immune response. Two principal components of this defense system, thioredoxin/glutathione reductase (TGR) and peroxiredoxin (Prx2), have been recently identified and validated as targets for anti-schistosomiasis drug development. In search of inhibitors of this critical redox cascade, we optimized and performed a highly miniaturized automated screen of 71,028 compounds arrayed as 7- to 15-point dilution sets. We identified novel structural series of TGR inhibitors, several of which are highly potent and should serve both as mechanistic tools for probing redox pathways in S. mansoni and as starting points for developing much-needed new treatments for schistosomiasis. The paradigm presented here effectively bridges the gap between academic target identification and the first steps of drug development, and should be applicable to a variety of other important neglected diseases.
| Schistosomiasis, also known as bilharzia, a debilitating disease resulting from the infection by the trematode parasite Schistosoma ssp. (S. mansoni, S. mekongi, S. japonicum, S. haematobium, and S. intercalatum) currently affects over 200 million people worldwide, mostly in developing countries [1]. A large percentage of those infected exhibit severe morbidity manifested as growth stunting, lassitude, and cognitive impairment [2], and an estimated 280,000 people die annually from the disease in sub-Saharan Africa alone [3]. The primary route of infection is via unsafe river and lake water, which is widely used in sub-Saharan Africa and Southeast Asia, among other regions, for irrigation, drinking, cooking, and bathing. Larval parasite forms (residing in and released by snails) can penetrate the skin of people contacting the water. The lifecycle of Schistosoma is exceedingly complex, with the parasite going through a number of stages both outside and inside the human host. Once inside humans, it can survive for years, even decades [4].
The need to control schistosomiasis is acute and efforts have been ongoing for years on three main fronts: prevention (via establishment and maintenance of sources of safe potable water), development of a vaccine, and use of drugs to treat the infection [1]. Although the number of schistosomiasis cases worldwide is indeed stunning, the number of drugs available to treat the disease is surprisingly small. Earlier in the 20th century, schistosomiasis was treated with highly toxic antimonial compounds, of which the most common was potassium antimonyl tartrate (PAT, tartar emetic). During the past three decades the only drug used against the infection is praziquantel, which is administered orally, is stable, effective against all major schistosome species in a single dose, and relatively inexpensive [5],[6]. However, because of high reinfection rates, praziquantel must be administered on an annual or semi-annual basis. While its exact mechanism of action is unclear, praziquantel is thought to affect the parasites by disrupting calcium homeostasis [7],[8]. Preliminary reports of praziquantel-resistant cases, and the generation of praziquantel-resistant parasites in the laboratory [9]–[11] highlight the need for new drugs to treat the disease. Artemisinin has shown promise as a new drug for schistosomiasis [12] although its use for schistosomiasis may be restricted in areas of malaria transmission so that its use as an antimalarial is not put at risk. Simplified derivatives of artemisinin, the 1,2,4-trioxolanes, show promise and potential selectivity, but these, like the parent compound, are significantly less active against adult schistosome parasites [13]. Oxamniquine, used extensively in Brazil in the past, is effective only against S. mansoni and resistance has been reported further reducing its potential value in schistosomiasis control [10].
Studies of the schistosome life cycle have focused on the fact it can survive for decades in the blood stream of the human host without being severely affected by the immune system and the associated assault by various reactive oxygen species (ROS). Since schistosomes do not have catalase to degrade hydrogen peroxide [14], other mechanisms must exist within the parasite to degrade ROS. Two uniquely positioned S. mansoni enzymes have been recently described that seem to act in concert to provide an effective antioxidant “firewall”. Thioredoxin glutathione reductase (TGR) is a multifunctional selenocysteine-containing enzyme that catalyzes the interconversion between reduced and oxidized forms of both glutathione (GSH) and thioredoxin (Trx), which are major contributors to the maintenance of redox balance in eukaryotes [15]. In contrast, humans possess two distinct enzymes, glutathione reductase (GR) and thioredoxin reductase (TrxR), which specifically recognize GSH and Trx as substrates, respectively [16]. The apparent replacement of two human enzymes by one dual-specificity worm enzyme has created a metabolic and regulatory bottleneck in which the inactivation of a single target, TGR, might have an enhanced deleterious effect on both the maintenance of parasite's redox balance and on its “antioxidant firewall”. Indeed, recent small molecule inhibition and RNA interference experiments have shown that inactivation of TGR has profound effects on S. mansoni survival rates both in culture and in infected mice [17]. Another component of the S. mansoni “firewall” are the peroxiredoxins (Prx), which are responsible for catalyzing the electron transfer to the main ROS agent hydrogen peroxide and, uniquely for schistosomes, from both GSH and Trx [18]. Thus, when TGR and Prx operate in concert, NADPH reducing equivalents are essentially transferred via TGR-catalyzed reaction to the oxidized forms of either Trx or GSH, while Trx or GSH in turn transfer reducing equivalents to hydrogen peroxide via Prx-catalyzed reactions (Figure 1).
Such improved understanding of the organisms responsible for neglected tropical diseases (NTDs) presents opportunities for new drug development. However, private-sector biopharmaceutical interest in NTDs has traditionally been limited due to high risk and low expected return-on-investment of these projects, though this is beginning to change with the advent of increased philanthropic and public-private-government partnership funding [19]. A significant problem that remains, however, is the significant gap in technologies, expertise, and cultures between academic and biopharmaceutical organizations [20]. At the US National Institutes of Health (NIH), the NIH Roadmap Molecular Libraries Initiative (MLI) was started in 2004 in part to address this problem. The MLI provides academic investigators with the pharma-scale infrastructure and technologies necessary to discover both chemical probes of physiology, and starting points for development of novel therapeutics for the rare and neglected diseases that are of less interest to the pharmaceutical sector [21].
The TGR/Prx work described here is the result of the first project officially accepted by the MLI in 2005. Since inhibiting either TGR or Prx can potentially lead to schistosome death [22], we chose to screen both enzymes in one assay as a reconstituted redox cascade. While TGR and Prx2 can be assayed individually, the separate assays are relatively less robust. TGR can be assayed in a relatively simple colorimetric assay by following the catalytic reduction of DTNB (5,5′ dithiobis(2-nitrobenzoic acid), Ellman's reagent) by NADPH; Prx2 at present can only be assayed with thioredoxin as a substrate together with TGR or thioredoxin reductase, or in a coupled reaction involving yeast glutathione reductase, and HTS-compatible assays [23] have yet to be developed [17],[18],[24]. By performing the high-throughput screen against both enzymes (present at equivalent levels in the assay), we were able to address both novel targets simultaneously while also combining target deconvolution and confirmation at the post-screen stage. In this report, we describe the miniaturization to 1536-well density of a cuvette-based assay for the TGR/Prx2 cascade which utilizes as a quantitative measure the decrease in fluorescence of the consumed NADPH substrate, the performance of a quantitative high-throughput screen (qHTS) [25] against 71,028 discrete compounds, and the initial characterization of several novel series of inhibitors. The application of qHTS, in which each library compound is assayed at a range of concentrations to generate a dose-response profile, facilitated triaging of actives for the purpose of structure-activity relationship (SAR) analysis and lead expansion.
Nicotinamide adenine dinucleotide phosphate (NADPH), glutathione reduced form (GSH), hydrogen peroxide, Tween-20 and potassium antimonyl tartrate (PAT) were procured from Sigma-Aldrich. DMSO Certified ACS Grade was from Fisher. The screening assay was performed in 100 mM phosphate buffer pH 7.4 containing 10 mM EDTA and 0.01% Tween-20.
Recombinant TGR with a fused bacterial-type SECIS element was expressed in the Escherichia coli strain BL21(DE3) (Invitrogen) in the presence of pSUABC in LB medium supplemented with 20 µM flavin adenine dinucleotide following conditions for optimal selenoprotein expression as described [17]. TGR was purified to homogeneity on an adenosine 2′,5′-diphosphate agarose (Sigma) column equilibrated with TE buffer as described [17]. TGR concentration was determined from the flavin adenine dinucleotide absorption (ε463 = 11.3 mM−1cm−1). The pure protein was dialyzed against PBS and stored at −80°C.
Recombinant Prx2 in pRSETA was expressed in E. coli strain BLR(DE3)pLysS (Novagen) as described [18]. Briefly, after a 3-hr induction in 1 mM IPTG, cells were sonicated in 5% monothioglycerol (3-mercapto-1,2-propanediol) in 10 mM imidazole, 0.07 M Na2HPO4, 0.01 M NaH2PO4, and 0.15 M NaCl, pH 7.4. The supernatant was filtered and Prx2 was purified to homogeneity on a His Trap column (Amersham Biosciences). Protein purity was verified by SDS-PAGE. The purified protein was dialyzed against PBS and stored at −80°C until used.
The 71,028 member library comprised two main subsets: 59,692 compounds from the NIH Molecular Libraries Small Molecule Repository (www.mli.nih.gov), prepared as 10 mM stock solutions in 384-well plates and delivered by Biofocus DPI (South San Francisco, CA, http://mlsmr.glpg.com/MLSMR_HomePage/), and NCGC internal exploratory collection of approximately 11,336 compounds which consisted of several commercially available libraries of known bioactives (1280 compounds from Sigma-Aldrich (LOPAC1280 library), 1120 compounds from Prestwick Chemical Inc. (Washington, DC), 980 compounds from Tocris (Ellisville, Missouri), 280 purified natural products from TimTec (Newark, DE), 1980 compounds from the National Cancer Institute (the NCI Diversity Set)), 1408 National Institute of Environmental Health Sciences collection of known toxic compounds, as well as collections from other commercial and academic collaborators (three 1000-member combinatorial libraries from Pharmacopeia (Cranbury, NJ), 718 compounds from Boston University Center for Chemical Methodology and Library Development, 96-member peptide library from Prof. Sam Gelman's lab, University of Wisconsin, Madison, and 474 compounds from the University of Pittsburgh Center for Chemical Methodology and Library Development). The compound library (7 µL each in 1536-well Greiner polypropylene compound plate) was prepared as DMSO solutions at initial concentrations ranging between 2 and 10 mM. Plate-to-plate (vertical) dilutions and 384-to-1536 compressions were performed on Evolution P3 dispense system equipped with 384-tip pipetting head and two RapidStak units (Perkin-Elmer, Wellesley, MA). Additional details on the preparation of the compound library are provided in Inglese et al [25].
Titration of the known inhibitor PAT (PubChem CID6328158) was delivered via pin transfer from a separate plate to the lower half of column 2 of each assay plate. The starting concentration of the control, dissolved in 1∶1 DMSO∶water, was 1 mM, followed by five-fold dilution points in duplicate, for a total of eight concentrations.
Three µL of reagents (100 µM NADPH in columns 3 and 4 as negative control and 100 µM NADPH, 42 nM TGR, 700 µM GSH, 83 nM Prx2 mixture in columns 1, 2, 5–48) were dispensed into 1536-well Greiner black assay plates. Compounds and control (23 nL) were transferred via Kalypsys PinTool equipped with 1536-pin array (10 nL slotted pins, V&P Scientific, Palo Alto, CA). The plate was incubated for 15 min at room temperature, and then a 1 µL aliquot of 400 µM NADPH/700 µM GSH was added, immediately followed by a 1 µL aliquot of 2.5 mM H2O2 to start the reaction. The plate was transferred to ViewLux high-throughput CCD imager (Perkin-Elmer, Wellesley, MA) where kinetic measurements (16 reads, one read every 30 sec) of the NADPH fluorescence decrease were acquired using 365 nm excitation/450 nm emission filter set. During dispense, the reagent bottles were kept submerged into 4°C recirculating chiller bath to minimize degradation. All screening operations were performed on a fully integrated robotic system (Kalypsys, San Diego, CA) containing one RX-130 and two RX-90 anthropomorphic robotic arms (Staubli, Duncan, SC). Library plates were screened starting from the lowest and proceeding to the highest concentration. Vehicle-only plates, with DMSO being pin-transferred to the entire column 5–48 compound area, were inserted uniformly at the rate of approximately one plate for every 50 library plates in order to monitor for and record any shifts in the background.
Time course data were collected on per-assay plate basis and were processed using in-house developed software. For each sample and at each individual concentration, 16 time points were processed using ordinary least squares regression to determine slope and intercept of linear fit. Additionally, a difference (delta) of last and first time point was generated for each time course. For activity calculations, delta values were chosen while the calculated slope, intercept, and the raw time-course data were stored in the database. Screening data were corrected and normalized and concentration–effect relationships derived by using the GeneData Screener software package (Basel, Switzerland). Percent activity was computed from the median values of the uninhibited, or neutral, control (48 wells located in column 1 and one-half of column 2) and the no-enzyme, or 100% inhibited, control (64 wells, entire columns 3 and 4), respectively. For assignment of plate concentrations and sample identifiers, ActivityBase (ID Business Solutions Ltd, Guildford, UK) was used for compound and plate registrations. An in-house database was used to track sample concentrations across plates. Correction factors were generated from the DMSO plate data and applied to each assay plate to correct for systematic errors in assay signal potentially resulting from issues with reagent dispensers or decrease in enzyme specific activity. A four parameter Hill equation [26] was fitted to the concentration-response data by minimizing the residual error between the modeled and observed responses. Outliers could be identified and masked by modeling the Hill equation and asking if the differences exceeded those expected from the noise in the assay.
The curve classification used is the same as described in Inglese et al. (2006) [25]. Briefly, concentration-response curves are placed into four classes: Class 1 contains complete concentration-response curves showing both upper and lower asymptotes and r2 values >0.9. Class 2 contain incomplete concentration-response curves lacking the lower asymptote and show r2 values >0.9. Class 3 curves are of the lowest confidence as they are defined by a single concentration point where the minimal acceptable activity is set at 3 SD of mean. Curves are classified as negative or positive depending on whether they exhibit signal decrease (apparent inhibition) or increase (apparent activation). Finally, Class 4 contains compounds that do not show any concentration response curves and are therefore classified as inactive.
A workflow was developed to facilitate a systematic approach in providing exhaustive analysis of structure activity relationships (SAR). First, a set of criteria used to define rules of determining an active set of compounds for the assay. These include decisions on inhibitors and activators, selectivity and counter screen information, curve class ranges, background fluorescence, etc. For this assay, compounds that showed signal activation (positive curve classes) were regarded as active due to fluorescence and were thus filtered out. The criteria are implemented as filters that are applied to rapidly define a core active set of compounds. This process eliminates many positive series that appear to have SAR and show reasonable titration response curves, but are not of biological relevance to the targets [23]. Next, the range of curve classes was limited to −1 through −3 to select for compounds showing signal decrease. Once an active set of compounds was identified, hierarchical agglomerative clustering with a 0.7 Tanimoto cutoff was performed using Leadscope (Leadscope Inc., Columbus, OH) fingerprints, which are ideally suited for two-dimensional scaffold-based based clustering [27]. For each cluster, maximal common substructures (MCS) were extracted, and a manual step of trimming the MCSs was performed to create a list of scaffolds. This clustering step typically has overlapping compounds and thus can lead to overlapping MCSs. This list of trimmed scaffolds is abridged to a canonical set. Each scaffold is then represented as a precise definition to indicate number of attachments, ring size variability, etc. All filters were then relaxed to include all negative assay data. In the initial clustering, a set of singletons was found. These compounds were reported upon separately with their individual activity profiles. SAR series and singletons were finally ranked by their activity profile.
Screening actives and analogues sourced as powders from the respective original suppliers (Sigma-Aldrich, NCI, Asinex, Chem Bridge, Tocris, Ambinter, and ChemDiv) were dissolved in DMSO to produce 10 mM initial stock solutions. The samples were then serially diluted row-wise in 384-well plate in twofold steps for a total of 24 concentrations, from 10 mM to 1.2 nM. Upon completion of the 24-point dilution, solutions from two 384-well plates were transferred to duplicate wells of 1536-well compound plate. The last two rows of the 1536-well plate did not contain any test compound and were reserved for placement of positive and negative controls. The assay protocol for confirmation was essentially the same as that described in the qHTS protocol section. A Flying Reagent Dispenser (FRD, Aurora Discovery, presently Beckman-Coulter) [28] was used to dispense reagents into the assay plates.
Assays were performed at 25°C in 0.1 M potassium phosphate, pH 7.4, 10 mM EDTA using 100 µM NADPH. The Prx2 assay was based on the reduction of H2O2 by Prx2 in the presence of GSH measured by the reduction of the GSSG produced in a coupled assay with yeast glutathione reductase monitored by observing the decrease in A340 nm due to consumption of NADPH (ε340 nm = 6.22 mM−1 cm−1) during the first three minutes [18]. The activity of TGR was determined with 3 mM 5,5′ dithiobis(2-nitrobenzoic acid) (DTNB, Ellman's reagent) [29] following the increase in A410 nm due to the production of 2-nitro-5-thiobenzoic acid (ε412 nm = 13.6 mM−1 cm−1) [17],[18].
The assay was initially developed and optimized using a spectrophotometer by following the decrease in absorbance at 340 nm associated with the consumption of NADPH substrate. During these studies, the main parameters of the assay, such as buffer conditions, concentration of each enzyme and substrate, DMSO and detergent tolerance, were tested and optimized (data not shown). The optimized assay utilized TGR and Prx2 at final concentrations of 25 nM and 50 nM, respectively. The substrates' final concentrations were 200 µM NADPH, 700 µM GSH, and 500 µM H2O2. The assay was miniaturized to 1536-well format by volume reduction and appropriate adjustment of stock concentrations of enzymes and substrates to reflect the volumes being combined. For example, the assay was started by the dispense of the two enzymes at 5/3 of their final concentration to account for the well volume increasing from 3 to 5 µL, while the hydrogen peroxide substrate was delivered as 5× solution to account for its dilution (1 µL to 5 µL final volume, see Materials and Methods, Table 1). All three types of reagents (enzymes, second aliquot of NADPH, and hydrogen peroxide) were tested and were shown to be stable overnight at 4°C, a requirement for the execution of an uninterrupted fully automated screen on the Kalypsys robotic system (Figure 2). In addition, the signal being monitored was changed from absorbance to fluorescence in order to: 1) improve the signal strength, as UV-shifted absorbance assays are generally difficult to scale to a 1536-well density (due to the combination of low extinction coefficient and short path length), and 2) minimize the quenching and inner-filter effects of a multitude of compounds which absorb light in this wavelength region [23].
In a typical uninhibited reaction in 1536-well plate, the well fluorescence changed from 370 relative fluorescence units (RFU) to 180 RFU within an eight-minute window. This resulting outcome, if registered as an end-point reading, and assuming zero change in the background, would yield a signal-to-background (S∶B) ratio of only approximately 2.1. Therefore the assay was further improved by modification of the type of signal collected from single end-point to multipoint kinetic read. Thus, for each plate the reaction progress was recorded for 8 minutes at the rate of one fluorescence read every 30 seconds. Such kinetic mode data acquisition not only secured a more robust assay signal but minimized the interfering effects of dust and mildly fluorescent compounds (see Discussion).
In total, 453 assay plates were screened in one uninterrupted robotic run lasting approximately 75 hours. Dispenser malfunction resulted in deteriorated signal in two plates and since the issue was noted in real time, the two 7-point libraries containing the problematic plates, plus two DMSO-only plates, were scheduled for re-screening immediately after the end of the main run. In this manner, the re-screened series were tested using the same batch of reagents as last series of the main screen. The assay performed robustly, yielding an average Z′ value of 0.76 [30]. Overall, the Z′ factor remained flat with the screen progression, with minor shifts tracking the introduction of new batches of the two enzymes (Figure 3A). The intraplate PAT control titration was stable throughout the screen progression, resulting in average IC50 of 14±8 nM and minimum significant ratio of 4.03 (Figure 3B) [31]. Each library compound was tested at a minimum of seven concentrations, ranging from 57 µM to 2.9 nM, and for each well, 16 time points were collected for a total of 9,562,432 data points. The screen and the preceding optimizations and validations consumed approximately 4.7 mg of TGR and 10 mg of Prx2. The overall materials cost of the screen (not including the cost of protein production) was approximately $5,200, or 0.85 cent per sample well, with approximately 80% of the costs associated with the assay microtiter plates and 15% attributed to NADPH.
Unlike traditional HTS, qHTS provides concentration responses for all the compounds screened and allows determination of an AC50 value, defined as the half-maximal activity concentration, for each compound in the primary screen. In qHTS concentration response curves are classified as belonging to one of four groups based on efficacy (response magnitude), presence of asymptotes, and goodness of fit of the curve to the data (r2). For the present screen, the activity associated with each well was computed from the change in fluorescence intensity over the time-course measurement period, normalized against control wells. In addition, the y-intercept of the reaction progress plot, typically equal to fluorescence at the first time point, was stored in the database and was used to further scrutinize purported actives. Compounds which showed activity but also had elevated y-intercept values were flagged as potential fluorescent artifacts.
Analysis of the qHTS results revealed 39 actives characterized by full concentration-response curves and IC50 values of better than 10 µM. After exclusion of antimony-containing compounds, as well as various mercury- and other heavy metal-containing molecules, the following series and singletons were selected for further studies after SAR analysis (Figure 4): oxadiazole 2-oxides (5 actives out of 29 analogues in collection, IC50 potency range between 8 µM and inactive), phosphinic amides (two compounds in the collection, one inactive and one active at 37 nM), phosphoramidite (singleton active, IC50 of 560 nM), and isoxazolone (singleton active, IC50 of 530 nM). In addition, a weaker series, quinolinyl sulfonamides (8 actives out of 47 total analogues in collection, IC50 potency range between 0.6 µM and inactive), was identified but noted to contain a number of both active and inactive members which were strongly fluorescent as judged by the extreme intercept values recorded during the screen. In contrast, neither the oxadiazole series nor the singleton actives exhibited any detectable autofluorescence (Figure 5).
In addition to inhibitors, the screen yielded a number of apparent activators, that is, compounds for which the increase in concentration led to a fluorescence intensity change greater than that of the neutral control. Upon examination of the time-course plots associated with these activators it became evident that the signal enhancement originated from high starting fluorescence which decreased during the observation window and in many cases entirely obscured the assay-driven NADPH fluorescence change (Figure 5C). While some of these compounds might be fluorescent substrates for either TGR or Prx2, which get converted to non-fluorescent products, a large number might simply be reactive towards any one or more components of the assay milieu (GSH, NADPH, and/or H2O2). As such, their confirmation and mode of action is subject of separate study.
In order to further confirm the qHTS actives and to expand the actives series, especially around the otherwise attractive singletons, powder samples were purchased from the original compound vendors and processed as described in the Methods. In addition to qHTS-identified compounds, untested analogues of the singletons were also procured in an attempt to support the singleton findings by the generation of small SAR series. The comparison of qHTS results, where applicable, and re-test results from independently acquired powder samples are shown in Figure 4, first and second data columns, respectively. The overall confirmation rate was excellent with the exception of the sulfonamide series of actives, which showed wide shifts between qHTS and confirmatory assay (results not shown). The apparent lack of confirmation for this series was consistent with the aberrant fluorescent values associated with many of its members. An analogue series built around a lower-potency benzoindolone singleton (NCGC00038549, IC50 of 3.9 µM) failed to yield activity against the screening assay and against both TGR and Prx2 individual assays. The original activity of that singleton was therefore deemed an artifact.
Gratifyingly, the previously-untested analogues of the phosphinic amides (compounds 1–4), phosphoramidite (13–16), and isoxazolone (10–12) actives all showed activity with various degrees of potency, supporting and expanding the qHTS findings. Specifically, the “gap” in potency between the highly active 3 (NCGC00042730, qHTS IC50 of 37 nM, confirmed at 25 nM on re-test) and the inactive distant analogue 4 (NCGC00064648) was filled partially by the newly-acquired analogues 1 (NCGC00093512, IC50 of 247 nM) and 2 (NCGC00093512, IC50 of 23 µM). Similarly, increased potency was achieved by the addition of analogues to the phosphoramidite (from a singleton IC50 of 0.5 µM to a range of 0.2–2 µM) and the isoxazolone (from a singleton IC50 of 0.5 µM to a range of 0.1 µM to 9 µM).
After establishing the activities of primary hits and new analogues against the screened dual-enzyme system, the compounds were further subjected to target deconvolution experiments. When tested against Prx2 in a GR-coupled hydrogen peroxide reduction assay none of the selected compounds showed activity up to the 50 µM top concentration tested (and by extension, none were active against GR, an enzyme related to TGR). Prx2 was therefore ruled out as the target of any of the actives identified in the screen. Results from the TGR assay are shown in the last column of Figure 4. The majority of active compounds demonstrated approximately the same, and in some instances improved (most notably with 7 and 8), potency against the isolated TGR as they did against the dual-enzyme system. These results not only confirm the initial findings from the screen, but also further support the hypothesis of TGR being the sole target of these actives. Additionally, all members of the sulfonamide series were inactive against both Prx2 and TGR, further strengthening the argument that their initial classification as actives was due to fluorescence interference originating from either the compound, impurities, or product(s) of its breakdown.
The stability, relatively low cost, and effectiveness of praziquantel has practically created a dependency on this single drug to treat schistosomiasis. Both the success of praziquantel and the general lack of incentives for large pharmaceutical companies to embark on research and development in the area of tropical diseases have led to a fairly dry pipeline for both drugs to treat schistosomiasis and basic research tools to study the lifecycle of this important parasite. To this end, we implemented a highly-miniaturized automated screen of the NCGC small molecule collection in an attempt to identify novel inhibitors of S. mansoni TGR or Prx2, both of which have been recently validated as crucial S. mansoni enzymes and have been proposed as targets for drug development. Prior to HTS adoption, the assay employed monitoring NADPH absorbance. While such a format is very convenient, offering fast access to kinetic data via the use of standard spectrophotometers, measuring absorbance in the UV region in 1536-well density is rarely practical. A significant fraction of organic molecules, as well as dust and buffer components, absorb in the 350 nm range, thereby introducing unacceptably high levels of interference. Additionally, the relatively low extinction coefficient of NADPH coupled with the short optical path length of the plate well significantly reduces the signal available for detection. Because NADPH is naturally fluorescent, emitting at ∼450 nm, while its oxidized counterpart NADP is not, we switched the detection platform for the coupled reaction from absorbance to fluorescence, a step that parallels the application of profluorescent substrates in assays for phosphatases and proteases [23], with the main difference being the fluorescence change trending from high to low in this reaction.
Because of the anticipated fluorescence interference from compound library members in this blue-shifted detection region and because the output generated from NADPH is not very strong (due to the combination of low extinction coefficient and quantum yield), we further modified the detection format of the assay to measure the reaction progress in kinetic mode as opposed to collecting a single end-point read. While kinetic, or time-course, measurements are routinely performed during assay development in low-throughput settings, their practical implementation during automated large-collection screens is not trivial. Unless the reaction under study is slow, only a fast-scanning reader or whole-plate imager (such as the ViewLux) can allow positionally-unbiased and rapid repeated measurements of 1536-well plates without significantly slowing down the overall plate processing speed. The collection of at least a two-point time course allows the effects of dust and fluorescent but otherwise inert library members to be subtracted out to reveal the true reaction course. Because the first time-point values (when the enzymatic reaction has produced minimal amount of product) associated with each compound well are stored in the database, a further analysis can be performed to flag interfering fluorescent library members [32]. An added benefit is that the signal-to-background computed from kinetic measurements significantly improves relative to end-point data and thus allows screening under conditions of low substrate conversion [33]. While in this screen we collected a total of 16 points per well, further optimization of the assay conditions could have resulted in shortened read time without the loss of sensitivity.
The primary screen against the TGR/Prx2 cascade was performed in Quantitative High Throughput Screening (qHTS) format. In qHTS, every compound in the collection is tested over a range of concentrations, spanning from tens of micromolar to low nanomolar, to generate a complete concentration-response profile. As such, qHTS is best described as high throughput pharmacology, since as a result of its application, not only are potencies and efficacies assigned to each active compound but also false positives and negatives due to outliers associated with individual concentration responses are easily identified in the context of titration. Additionally, due to the built-in replicates in the testing of each compound, the need for laborious and infrastructure-intensive cherry-picking, original-result replication, and dose-response characterization are eliminated. In our present assay, the application of qHTS enabled us to not only skip the direct confirmation steps but also to combine the actives verification from independently-sourced powders with series expansion around limited SAR or singletons. In traditional single-concentration screening, singleton actives are necessarily treated with great caution given statistical uncertainties. In this study, the qHTS paradigm allowed us to confidently select the potent phosphinic amide, isoxazolone and phosphoramidite singletons 3, 10 and 14 (Figure 4) for further testing and that selection was later validated by the excellent confirmation of those actives and the successful expansion of the series. Separate, but equally important, is the aspect of reliability and robustness of screening data. qHTS, with the combination of dose-survey and replicate points, indeed offers uniquely rich and robust data sets for deposition in recently established public databases, such as PubChem. Additionally, in order to minimize the interfering effect of promiscuous inhibitors acting via colloidal aggregate formation [34],[35], we included detergent in the assay buffer.
Throughout the entire screen, the assay performed in a robust manner, yielding an average Z′ value of 0.76. Overall, the Z′ factor remained flat with the screen progression, with minor shifts tracking the introduction of fresh batches of enzyme. The availability of periodically computed Z′, signal-to-background, and activity heatmaps throughout the screen progression, made possible by the development of fast data-processing tools in-house, significantly improved our response time when screen complications arose. For example, the dip in Z′, also accompanied by noisy activity heatmap (not shown), was noted almost in real-time and this allowed the appropriate concentration series to be scheduled for re-run within the same screening session.
Figure 3B presents a cumulative plot of all intra-plate concentration-response curves of PAT throughout the entire screen. The narrow range of observed IC50 values serves as a further indication that the screen performed robustly from a standpoint of enzyme activity and responsiveness to inhibition. The titration curve displayed stability throughout the screen despite the fact that PAT is only partially soluble in DMSO and required formulation in high-percent water, leading to concerns about evaporation-related variability. Analyzing the trend in the intra-plate control as a function of screen progression allows one to ascertain the ‘health’ of the screening system as a whole, because the variations or dramatic shifts in potency of the control could be due to not only a deterioration in enzyme quality (which could otherwise be detected from an shift in the S∶B value) but also to problems with the pintool delivery of compounds. The absence of abrupt and significant shifts in the intra-plate control curve allows us therefore to conclude that the compound transfer remained uniform throughout the screen.
The screen identified numerous arsenic, antimony, mercury, and other heavy-metal containing compounds (data not shown; for complete set of actives, see PubChem, AID 448). The antimony-containing compounds were largely similar to PAT and were therefore expected to be identified by this assay. Likewise, Hg-derivatives inhibited the enzymes strongly, as expected. While PAT and the gold-containing drug auranofin had been shown to inhibit thioredoxin reductase and TGR [17],[36],[37], and while more recently arsenic trioxide has shown anticancer activity and its effect has been ascribed to Trx reductase-mediated apoptosis [38], we restricted our analysis to novel nonmetal-containing compounds, primarily due to the fact that heavy metal-containing compounds frequently exhibit non-selective inhibition against a broad panel of enzymes and because nonmetal novel chemotypes against these previously-unscreened targets appeared to offer the greatest promise for optimizing potency and specificity. In this regard, the screen was successful, having resulted in identification of several distinct series of TGR inhibitors. It is noteworthy that the top actives from some series, such as the phosphinic amide 3, oxadiazole 2-oxide 7, and isoxazolone 11, yielded IC50 values close to the final TGR assay concentration (25 nM in qHTS and confirmation and 15 nM in the TGR individual assay), thus approaching the limit of the detectable potency range. In terms of the lead actives there are several interesting points. For instance, the role of the benzothiazole heterocycle within the phosphinic amide series is apparently critical for inhibition as illustrated by the comparative values of analogues 3, active and containing a benzothiazole moiety, and 4, inactive and devoid of benzothiazole (Figure 4). The oxadiazole series contains several symmetric heterocyles (a function of their synthetic ease) and are known NO donors [39],[40]. The presence of two phosphorus based small molecules may well relate to the presence of a selenocysteine in TGR and the relative electrophilic nature of this functionality. Studies to expand upon selected lead actives and further understand their mechanism of action are currently underway. Furthermore, in the Prx2 deconvolution assay, none of the top actives were found to inhibit GR, an enzyme closely related to TGR. This suggests that selective activity against the parasite (which lacks GR) and less toxicity to humans (who have GR) can be achieved.
In summary, a kinetic-based qHTS against a pair of novel, validated targets from S. mansoni allowed fast and reliable identification of compounds active against this critical redox cascade. We have identified several novel structural series of TGR inhibitors, several of which are highly potent and should serve both as mechanistic tools for probing the redox balance in S. mansoni, and starting points for developing medicinal leads for much-needed new treatments for schistosomiasis. The work presented here effectively bridged the gap between academic target identification and the first steps of drug development for an important neglected disease [20]. Generalization of this paradigm to other neglected diseases could prove be a powerful approach to catalyzing new therapeutic development for NTDs. |
10.1371/journal.pmed.1002316 | Risk of hospitalization with neurodegenerative disease after moderate-to-severe traumatic brain injury in the working-age population: A retrospective cohort study using the Finnish national health registries | Previous epidemiological studies suggest that working-aged persons with a history of moderate-to-severe traumatic brain injury (TBI) may have an increased risk for developing neurodegenerative disease (NDD) while persons with a history of mild TBI do not. In this comprehensive nationwide study in Finland, we assessed the risk of NDD and history of moderate-to-severe TBI in the working-age population.
We performed a population-based follow-up study using the Finnish Care Register for Health Care to identify all persons between the ages of 18 and 65 years hospitalized during 1987–2014 due to TBI who did not have a baseline NDD diagnosis. We compared the risk of hospitalization with NDD between persons hospitalized due to moderate-to-severe TBI (intracranial lesions) and persons hospitalized due to mild TBI (no intracranial lesions). Follow-up NDD diagnoses were recorded from 1 year following the TBI to the end of 2014. NDD diagnoses included dementia, Parkinson disease, and amyotrophic lateral sclerosis. We used a Cox proportional hazards model, adjusting for age, sex, education, and socioeconomic group, to assess the association between TBI and NDD. In total, 19,936 and 20,703 persons with a history of moderate-to-severe TBI and mild TBI, respectively, were included. The overall time at risk was 453,079 person-years (median 10 years per person). In total, 3.5% (N = 696) persons in the moderate-to-severe TBI group developed NDD compared to 1.6% (N = 326) in the mild TBI group. After adjusting for covariates, moderate-to-severe TBI was associated with an increased risk for NDD, with a hazard ratio (HR) of 1.8 (95% CI 1.6–2.1) compared to mild TBI. Of the NDD subtypes, only moderate-to-severe TBI was associated with an increased risk for dementia (HR 1.9, 95% CI 1.6–2.2). Yet, this large-scale epidemiological study does not prove that there is a causal relationship between moderate-to-severe TBI and NDD. Further, the Care Register for Health Care includes only hospitalized persons; thus, patients diagnosed with NDD in the outpatient setting may have been missed. Additional limitations include the potential for miscoding and unmeasured confounds.
In working-aged persons, a history of moderate-to-severe TBI is associated with an increased risk for future dementia but not for Parkinson disease or amyotrophic lateral sclerosis.
| Traumatic brain injury is a leading cause of death and disability worldwide, especially among young adults.
Previous studies on the association between traumatic brain injury and neurodegenerative diseases have controversial findings. Furthermore, association studies with a long follow-up time are scarce.
We conducted a retrospective, register-based nationwide study, identifying 19,936 persons (18–65 years) with a history of moderate-to-severe traumatic brain injury and 20,703 persons (18–65 years) with a history of mild traumatic brain injury admitted to hospitals between 1987 and 2014 in Finland.
After adjusting for confounding factors, we found that persons with moderate-to-severe traumatic brain injury had a 90% increased probability of developing dementia compared to persons with mild traumatic brain injury.
The risk was increased only for dementia, not for Parkinson disease or amyotrophic lateral sclerosis.
Moderate-to-severe traumatic brain injury seems to be associated with a risk of future dementia in young and middle-aged adults.
Further studies are required to examine causal factors contributing to the found association.
| Traumatic brain injury (TBI) is a globally increasing healthcare problem, affecting persons of all ages [1]. Following the early phases of TBI, patients face a significant risk of long-term disability and neurological morbidity [2]. Previous epidemiological studies have found an association between history of TBI and risk for future neurodegenerative disease (NDD) (a concept including dementia, Parkinson disease [PD], and amyotrophic lateral sclerosis [ALS]), but the results have been conflicting and few studies have focused on the working-age population [3–7]. Gardner et al. showed that persons under 65 years with a history of mild TBI did not have an increased risk for dementia compared to non-TBI controls [8]. The development of NDD is supposedly most deleterious in the working-age population, as this not only causes significant morbidity but also has major socioeconomic consequences. Yet, to our knowledge, no previous studies have specifically looked at the overall risk for developing NDD in working-aged persons hospitalized due to TBI.
Our aim is to contrast the risk of NDD in working-aged persons hospitalized due to moderate-to-severe TBI to that of persons hospitalized due to mild TBI. Persons with a history of mild TBI and moderate-to-severe TBI are similar in regard to TBI-specific risk factors, such as alcohol use, which is why individuals with mild TBI serve as a suitable control group [9]. Further, since data on the possible effects of socioeconomic factors on the association between TBI and NDD are lacking, we adjust for education and socioeconomic group [10]. We hypothesized that working-aged persons with a history of moderate-to-severe TBI would have an increased risk for future NDD compared to persons with a history of mild TBI, after adjusting for covariates.
The National Institute for Health and Welfare (THL/1326/5.05.00/2015) approved of the study, in accordance with Finnish national legislation. Statistics Finland (Dnro TK-53-1179-16) and the Population Register Centre (Dnro 1873/410/16) granted us access to their databases. The Finnish Office of the Data Protection Ombudsman (Dnro 2794/402/2015) approved the data collection and combining of data registries. The study was conducted according to the Declaration of Helsinki of the World Medical Association.
We used the nationwide Finnish Care Register for Health Care to identify persons treated in a public hospital due to TBI in Finland during 1987–2014. The Care Register for Health Care (a continuation of the previous Hospital Discharge Register) was established by the National Institute for Health and Welfare and contains data on persons discharged from every public hospital in Finland from 1969 onwards. The Finnish healthcare system is tax-funded by local municipalities and by the state. In practice, acute care of TBIs is provided solely by public nonprofit healthcare providers and not by private institutions. Thus, the Care Register for Health Care comprehensively includes persons hospitalized due to TBI. The same register can also be used to identify persons hospitalized with NDD. Previous studies have verified the diagnostic accuracy of the registers [11,12].
The study population consisted of working-aged persons (18–65 years) hospitalized due to moderate-to-severe TBI or mild TBI between 1 January 1987 and 31 December 2014. Mild TBI was defined as a discharge diagnosis indicating no traumatic intracranial lesion (ICD-9 850; ICD-10 S06.0) according to US Centers for Disease Control and Prevention (CDC) criteria, with the exception of isolated skull fractures [13]. Moderate-to-severe TBI was defined as a discharge diagnosis indicating traumatic intracranial lesion (ICD-9 851–854; ICD-10 S06.1–S06.9).
To diminish the likelihood of persons in the mild TBI group having significant traumatic intracranial lesions, we included only those hospitalized for less than 1 day. To diminish the likelihood of persons in the moderate-to-severe TBI group having no significant traumatic intracranial lesion, we included only persons hospitalized for 3 days or longer. In the case of several hospitalizations due to moderate-to-severe TBI or mild TBI, we used the first date of the most severe head injury (i.e., if the person first had a mild TBI and later a moderate-to-severe TBI, the latter was used).
The study population was prospectively followed up from time of TBI until diagnosis of NDD, death, emigration, or end of follow-up on 31 December 2014. We identified persons hospitalized, for any reason, who were diagnosed with a new NDD diagnosis of dementia (ICD-9 290, 331, 797; ICD-10 G30, F00, F01, F02, F03), PD (ICD-9 332; ICD-10 G20), or ALS (ICD-9 335.2; ICD-10 G12.2) from the Care Register for Health Care. The admission date for the hospitalization including the new NDD diagnosis was defined as the date of diagnosis. We excluded persons with a NDD diagnosis prior to the TBI and persons permanently living outside of Finland. We further excluded persons who received a NDD diagnosis or died within 1 year of the TBI to diminish the possibility of reverse causality. Fig 1 shows a flow chart of the selection and follow-up protocol.
Statistics Finland classification of socioeconomic groups is based on the United Nation’s recommendations for the 1990 population censuses. The education classification system is based upon the International Standard Classification of Education 1997 and 2011 classifications. Data on socioeconomic group and highest level of education were obtained for the year closest to the end of follow-up. The classification systems are presented in S1 Text.
Data on mortality and date of emigration were obtained from the Population Register Centre of Finland. Information on date of birth, sex, hospitalization dates, and diagnoses came from the Care Register for Health Care. Data on socioeconomic group and education were obtained from Statistics Finland. We used the unique identification number assigned to all Finnish citizens to identify individuals and combine data from the registers.
Descriptive characteristics of the cohort are presented either as categorical data (N [percent] and compared using a two-sided χ2 test) or as continuous data (mean [standard deviation] and compared using a t test). We calculated the unadjusted rates of NDD per 100,000 person-years. This was done overall and in prespecified age groups, given the increase in risk of NDD with age. Data are presented as NDD rate per 100,000 person-years with 95% confidence intervals.
We used Cox proportional hazards models in Stata (version 14, StataCorp, College Station, TX) to estimate covariate-adjusted hazard ratios (HRs) with 95% confidence intervals. Date of NDD was set as date of study exit. Persons dying before the end of follow-up or moving outside of Finland before the end of follow-up were censored at the time of death or emigration. Date of study entry was the hospital admission date for the hospitalization due to TBI. Age (continuous variable) was the underlying time parameter in all analyses. Socioeconomic group and level of education were treated as categorical variables, using the group with the most individuals as the reference category.
In the primary analysis, we used NDD as a composite outcome variable. In the sensitivity analyses, we separately assessed the risk for dementia, PD, and ALS. Subgroup analyses were conducted by sex, by prespecified age groups (18 to 40 years, 41 to 50 years, 51 to 60 years, and 61 to 65 years), and by equally sized hospital length of stay quartiles (3–5 days, 6–10 days, 11–24 days, ≥25 days). Hospital length of stay served as a surrogate marker of TBI severity [14].
We repeated the analysis using the Cox proportional hazards model in a matched sample within the cohort. We matched persons with a history of moderate-to-severe TBI and mild TBI in a 1:1 fashion based on age, sex, education, and socioeconomic group using the “ccmatch” function in Stata. If there were multiple mild TBI cases that matched a moderate-to-severe TBI case with respect to these variables, the “ccmatch” function included all of them.
The results are presented as HRs with 95% confidence intervals. We used the group of persons with a history of mild TBI as the reference group in all analyses. p-Values < 0.05 were considered statistically significant. We derived log–log plots of survival curves of TBI to verify that the proportional hazards assumption was not violated.
The reporting of this study is in accordance to the REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement (S1 RECORD Checklist) [15].
The preplanned statistical plan is presented in S2 Text. One additional analysis assessing how the risk of dementia behaves as a function of time (in the matched cohort) when including persons diagnosed with dementia within the first year after TBI was performed in response to a reviewer comment.
A total of 19,936 persons with a history of moderate-to-severe TBI and 20,703 persons with a history of mild TBI were identified (Fig 1). Baseline characteristics of the two groups are shown in Table 1. Persons in the moderate-to-severe TBI group were on average 7 years older at the time of injury than persons in the mild TBI group (46 versus 39 years). The male to female ratio was higher in the moderate-to-severe TBI group than in the mild TBI group, although men predominated in both groups. There were no major differences in educational level between the groups. Nearly half of all persons had an upper-secondary level of education, while the higher education levels contained between 4% and 9% of individuals. Twenty-nine percent of persons in the moderate-to-severe TBI group died during follow-up, compared to 12% in the mild TBI group.
Overall time at risk was 453,079 person-years (mean 11 years [SD 8], median 10 years [IQR 4–17], per person). During the follow-up, 696 (3.5%) persons with a history of moderate-to-severe TBI developed NDD compared to 326 (1.6%) of those with a history of mild TBI (p < 0.001). Dementia was the most frequent NDD, followed by PD and ALS, in both groups. Persons in the moderate-to-severe TBI group were on average 4 years younger than persons in the mild TBI group at the time of NDD diagnosis (67 versus 71 years). Significantly more persons in the moderate-to-severe TBI group were diagnosed with NDD before the age of 65 years compared to the mild TBI group (40% versus 26% of all NDD cases in the respective groups, p < 0.001). Persons with a history of moderate-to-severe TBI who went on to develop NDD tended to have a longer hospital length of stay than those who did not develop NDD, suggesting higher TBI severity.
There were no significant differences between persons in the moderate-to-severe TBI group and persons in the mild TBI group who developed NDD in terms of age, socioeconomic group, or level of education (S1 Table). The male to female ratio was higher among individuals in the moderate-to-severe TBI group who developed NDD than among individuals in the mild TBI group who developed NDD.
The unadjusted rate of NDD was 331 per 100,000 person-years in the moderate-to-severe TBI group (318 per 100,000 in men and 373 per 100,000 in women) and 134 per 100,000 person-years in the mild-TBI group (115 per 100,000 in men and 162 per 100,000 in women). The unadjusted rates for all three NDD subtypes were notably higher in the moderate-to-severe TBI group than in the mild TBI group (Table 2). The incidence of NDD increased with age. In the two youngest age groups (18–40 and 41–50 years), the rate of NDD was three to five times higher in persons with moderate-to-severe TBI compared to mild TBI, whereas in the two older age groups (51–60 and 61–65 years), the incidence was approximately one and a half times higher in the moderate-to-severe TBI group.
In the primary analysis, adjusting for age, sex, level of education, and socioeconomic group, moderate-to-severe TBI was associated with an increased risk of NDD, with a HR of 1.8 (95% CI 1.6–2.1, p < 0.001) compared to mild TBI (Fig 2). When analyzing all persons, female sex was associated with a decreased risk of NDD (HR 0.8, 95% CI 0.7–0.9, p < 0.001).
The sensitivity (for NDD subtypes) and subgroup analyses (by sex, age group, and hospital length of stay) are shown in Table 2. In the sensitivity analyses, moderate-to-severe TBI was associated with an increased risk for dementia compared to mild TBI (HR 1.9) but not for PD or ALS. When analyzing women separately, moderate-to-severe TBI was associated with an increased risk for NDD compared to mild TBI, with a HR of 1.9. When analyzing men separately, moderate-to-severe TBI was associated with an increased risk for NDD with a HR of 1.7 (Fig 3). The relative risk of NDD among those with a history of moderate-to-severe TBI was highest among those aged 41 to 50 years and 51 to 60 years of age at baseline (HR 2.7 and 2.0, respectively), though the differences between age groups were not significant (overlapping 95% CIs). Increasing TBI severity, as reflected by duration of hospitalization, was associated with an increased risk for NDD (HR 1.3 to 2.1) in a dose–response pattern.
We identified a total of 25,747 exposure–control matched persons within the follow-up cohort, of which 13,470 were persons with a history of mild TBI, and 12,277 persons with a history of moderate-to-severe TBI. There were no major differences in age, sex, level of education, or socioeconomic group distribution between the matched groups (S2 Table). In the matched sample, 3.2% of persons with a history of moderate-to-severe TBI developed NDD compared to 2.3% of persons with a history of mild TBI. In the Cox proportional hazards model, TBI was associated with an increased risk for NDD, with a HR of 1.8 (95% CI 1.6–2.1, p < 0.001), providing additional support for the results from the primary analysis.
A reviewer requested additional analysis to show how the risk of dementia behaved as a function of time (in the matched sample cohort). This analysis did not exclude persons diagnosed with dementia within the first year after the TBI. The risk for dementia was continually higher for persons with a history of moderate-to-severe TBI compared those with a history of mild TBI (S1 Fig). Furthermore, the lines diverge with time, providing further support for the association between moderate-to-severe TBI and dementia.
In this nationwide study in Finland investigating the association between TBI and NDD, we found that persons with a history of moderate-to-severe TBI had an 80% increased probability of future NDD compared to persons with a history of mild TBI. The matched sample analysis strengthened our results. The risk of future NDD increased with TBI severity (length of hospital stay) in a dose–response pattern. However, when the three NDD subtypes were analyzed separately, TBI was associated with increased risk only for dementia (90% increased probability).
Similar to our findings, three previous meta-analyses found TBI to be associated with an increased risk for dementia [3–5]. We further found that the associated risk between NDD and dementia increased with TBI severity. Thus, the evidence for the association between TBI and dementia seems to be compelling. Conversely, the association of TBI with PD and ALS is not as clear. Some studies suggest a significant association, and some no association [6,16–19]. In the present study, no association between moderate-to-severe TBI and PD or ALS could be established. The absolute numbers of persons developing PD and ALS were limited. Thus, the negative finding might be the consequence of a type II error. Yet, considering that this was an almost 30-year-long nationwide follow-up study including all persons hospitalized for TBI in Finland, ALS and PD do not seem to be a significant long-term neurological problem in individuals with a history of moderate-to-severe TBI. It should be noted that we were only able to identify persons diagnosed with ALS and PD who had been hospitalized. Thus, it is possible that we missed persons diagnosed with ALS or PD in the primary care setting or by private sector specialists.
In the present study, we could not study the risk of dementia in persons without a history of mild or moderate-to-severe TBI. However, based on previous studies, 14,500 persons are diagnosed with dementia annually in Finland. [20]. For a Finnish adult population (18 years or older) of 4,385,426, this translates to an incidence of 331 per 100,000 person-years. In comparison, the unadjusted incidence of dementia was 293 per 100,000 person-years for persons with a history of moderate-to-severe TBI and 114 per 100,000 person-years for persons with a history of mild TBI. Thus, working-aged persons with a moderate-to-severe injury at a young age have a similar incidence of dementia as the general population, where most cases are among elderly individuals [21]. The incidence of dementia in persons with a history of mild TBI is somewhat lower than the general incidence, most probably because our cohort included persons under 65 years of age at the time of injury.
A major limitation for comparing previous studies of the association between TBI and NDD is the large variation in TBI and NDD definitions [5]. Some studies rely upon self-reported diagnoses for TBI (and thus suffer from recall bias), while others use ICD diagnoses (eight, ninth, or tenth revisions). For NDD diagnosis, some studies use DSM criteria, ICD diagnoses, or NINCDS-ADRDA diagnostic criteria. Thus, not surprisingly, with such a wide spectrum of the basic definitions, the results have been conflicting. In this study, we defined mild TBI according to the CDC criteria [13]. The same definition was used by Gardner et al. [8], who showed that persons under the age of 65 years with mild TBI do not have an increased risk for dementia. Thus, as we specifically investigated persons under 65 years of age, the mild TBI population formed a suitable control group for individuals with moderate-to-severe TBI. Persons treated for mild and moderate-to-severe TBI are likely to be similar in TBI-specific risk factors, such as age, gender, alcohol use, and socioeconomic factors [22]. Yet, the included mild TBI population might differ somewhat from the most common form of mild TBI, when hospital admission is not required. Identifying a control group of individuals with a history mild TBI without hospitalization is, however, impossible in large epidemiological studies such as the present study.
The role of sex differences in risk of developing dementia after TBI has been widely discussed. Both Mortimer et al. [4] and Fleminger et al. [3] found men, but not women, with a history of TBI to have an increased risk of dementia after TBI. Speculated theories include estrogen- and progesterone-induced neuroprotection [23,24]. In our study, both men and women with a history of moderate-to-severe TBI had an increased risk for dementia compared to those with a history of mild TBI. Yet, female sex was associated with a reduced risk for dementia in comparison to male sex, both in the mild TBI and moderate-to-severe TBI groups. Population studies generally do not find differences in dementia incidence between men and women [25,26]. Our study cannot establish any causation between sex, TBI, and dementia, although our results imply that sex may play a role in TBI-related dementia. The underlying causes of TBI may be different in men and women, resulting in a differential capacity to recover from brain injury and hence differential risk of dementia.
Poor socioeconomic factors (such as low education level and socioeconomic group) increase the risk for sustaining a TBI [9]. However, they also serve as a risk factor future NDD [27]. Thus, the association between TBI and NDD in previous studies may have reflected the underlying association of these socioeconomic factors, rather than of TBI itself, with NDD. Yet, we found that even after adjusting for socioeconomic group and level of education, TBI was significantly associated with an increased risk for NDD.
This is, to our knowledge, the first nationwide study on the subject (including over 40,000 persons). The nationwide coverage and the high data quality of the registries strengthen the study’s generalizability [11,12]. Only a few previous studies match the present one in size. One of the larger studies (with a maximum follow-up time of 5 to 7 years) was by Gardner et al. [8], in which the authors identified approximately 50,000 persons with a history of TBI and found results like ours, i.e., moderate-to-severe TBI increased the risk for dementia, with a HR of 1.7 (compared to 1.9 in our study) [8]. Yet, to date, the present study has one of the longest follow-up times (mean time at risk 11 years, or 453,079 person-years), something that is essential when investigating long-term neurological morbidity after TBI. By using persons with a hospitalization due to mild TBI as controls, we diminished the likelihood of detecting an effect that is not present (i.e., type I error), as it is unlikely that persons with a history of mild TBI have an increased risk of dementia compared to persons with a history of non-brain trauma [8]. Yet, there are studies suggesting that mild TBI itself might increase the risk for NDD, and therefore it is possible that our results underestimate the effect of moderate-to-severe TBI in the development of NDD [28]. On the other hand, a recent systematic review found no association between mild TBI and dementia [29].
Despite the high quality of the registries used, all register-based studies include diagnostic inaccuracies, coding errors, and other confounding factors that cannot be controlled for. First, as the Care Register for Health Care includes only hospitalized persons; it is possible that we missed persons being diagnosed with NDD in the outpatient setting (e.g., milder forms of PD and dementia). Second, although the median follow-up time was 10 years, persons hospitalized for TBI during the more recent years inevitably had shorter follow-up times and may not yet have been diagnosed with a NDD. Third, a notably higher proportion of persons in the moderate-to-severe TBI group died during the follow-up period compared to the mild TBI group (29% versus 12%), decreasing the moderate-to-mild TBI cases’ exposure time. Thus, it is likely that the risk of NDD in persons with a history of moderate-to-severe TBI is even higher than presented in this study. Furthermore, the diagnosis of NDD, especially dementia, is prone to error if it occurs too soon after TBI. Such diagnosis may be a residual effect of delirium, medication, or other complications following TBI. To avoid the possibility of reverse causality, we recorded NDD diagnoses starting 1 year following the TBI. Yet, even after including persons diagnosed with dementia within the first year after the TBI, moderate-to-severe TBI was associated with a significantly higher risk for dementia than mild TBI, strengthening the association (S1 Fig).
As in many register-based studies, we used ICD-9 and ICD-10 discharge diagnoses to identify persons with a history of moderate-to-severe TBI and mild TBI [30]. For mild TBI we used diagnoses indicating no structural intracranial injury, and for moderate-to-severe TBI, we used diagnoses indicating an objective intracranial injury [31]. We further excluded persons with mild TBI hospitalized for longer than 1 day, as these may have had an undiagnosed intracranial injury, and persons with moderate-to-severe TBI hospitalized for shorter than 3 days, as these may represent cases of either rapid death or a milder form of intracranial injury not requiring hospitalization. It is possible that in the mild TBI group there were persons with clinically silent diffuse axonal injuries that passed undetected. How such traumatic microlesions affect the risk of future NDD is unknown.
The most evident limitation, which is shared by most large-scale epidemiological studies, is that the study setup does not allow to us analyze any causative factors. It has been hypothesized that TBI does not itself cause NDD but rather accelerates an underlying process of developing NDD in persons with predisposing factors [32]. Such predisposing factors may include comorbidities; genetic variations, such as APOE ε4 allele expression and neprilysin polymorphism; and lifestyle factors, such as cognitive reserve, physical activity, obesity, alcohol, and smoking [33]. For example, a substantial proportion of TBIs in Finland relate to alcohol use [34]. Alcohol-related TBIs are much more common in men than in women, especially among less educated people, who also have a higher baseline risk for NDD, which may confound our results [35,36]. Furthermore, comorbidities, such as hypertension [37], stroke [38], and diabetes [39], have been found to significantly associate with risk of NDD, especially with dementia. Therefore, differences in comorbidities between the mild and moderate-to-severe TBI groups may potentially have affected our results. The aspect of physical activity after TBI is interesting. Decreased physical activity is probably more likely to happen after moderate-to-severe TBI than after mild TBI. Decreased physical activity is associated with an increased risk for dementia [40,41]. Thus, increasing physical activity in persons after moderate-to-severe TBI, in combination with aggressive treatment of cardiovascular comorbidities, might decrease the risk of subsequent dementia. Yet, further studies investigating the causative relationship between TBI, other environmental risk factors, and genetics are needed.
Our results suggest that in working-aged persons, moderate-to-severe TBI increases the risk for developing NDD later in life. In our study, the risk seemed to increase with TBI severity in a dose–response pattern, and the risk was higher in men. With regard to the NDD subtypes, moderate-to-severe TBI was associated with increased risk for dementia but not for PD and ALS. The effect of covariates, such as comorbidities, lifestyle factors, and genetic factors, should be accounted for in future etiological studies, as well as studies to improve diagnostics and prevention of dementia after TBI.
|
10.1371/journal.pntd.0005420 | Single locus genotyping to track Leishmania donovani in the Indian subcontinent: Application in Nepal | We designed a straightforward method for discriminating circulating Leishmania populations in the Indian subcontinent (ISC). Research on transmission dynamics of visceral leishmaniasis (VL, or Kala-azar) was recently identified as one of the key research priorities for elimination of the disease in the ISC. VL in Bangladesh, India, and Nepal is caused by genetically homogeneous populations of Leishmania donovani parasites, transmitted by female sandflies. Classical methods to study diversity of these protozoa in other regions of the world, such as microsatellite typing, have proven of little use in the area, as they are not able to discriminate most genotypes. Recently, whole genome sequencing (WGS) so far identified 10 different populations termed ISC001-ISC010.
As an alternative to WGS for epidemiological or clinical studies, we designed assays based on PCR amplification followed by dideoxynucleotide sequencing for identification of the non-recombinant genotypes ISC001 up to ISC007. These assays were applied on 106 parasite isolates collected in Nepal between 2011 and 2014. Combined with data from WGS on strains collected in the period 2002–2011, we provide a proof-of-principle for the application of genotyping to study treatment outcome, and differential geographic distribution.
Our method can aid in epidemiological follow-up of visceral leishmaniasis in the Indian subcontinent, a necessity in the frame of the Kala-azar elimination initiative in the region.
| Visceral Leishmaniasis (VL) or Kala-azar is a life-threatening neglected tropical disease that annually affects half a million people worldwide. In the Indian subcontinent (India, Nepal, Bangladesh), the disease is caused by infection with the protozoan parasite Leishmania donovani, which is transmitted by female sand flies. Currently, the Kala-azar elimination program aims at reducing the number of VL cases in the region to less than 1 in 10.000 at upazila, sub-district and district level in Bangladesh, India, and Nepal respectively. In support of this program, tools for tracking L. donovani populations are essential, because these allow monitoring geographic spread over time. However, the parasite populations in the region are highly homogeneous, requiring sequencing of the entire genome to gather sufficient information for discriminating them. Because whole genome sequencing (WGS) is impractical for large-scale use, we designed a simple alternative to identify the WGS-genotypes. Our method is based on PCR amplification followed by sequencing of one particular locus, diagnostic of each population. We provide proof-of-principle that our method can be used to track parasite populations over time, and to correlate them with clinical parameters. We believe that our assay can support the Kala-azar control efforts in the Indian subcontinent.
| Visceral Leishmaniasis (VL) is a neglected tropical disease caused by parasites of the Leishmania donovani species complex, which are transmitted by the bite of phlebotomine sand flies. WHO estimates that annually, 300 million people worldwide are at risk of VL [1]. In the Indian subcontinent (ISC), including Bangladesh, India, and Nepal, each year an estimated 237,500 new cases occurred, of which 4,450 in Nepal [2]. Together this accounts for 80% of all new VL cases worldwide. However, recently VL incidence is declining in the Indian subcontinent, and the numbers of new reported cases have dropped to 735, 9241, and 335 respectively in the aforementioned countries in 2014 [3]. Monitoring/tracking the genetic diversity of the parasites in these endemic regions over time is essential to understand transmission dynamics. This can help to evaluate the effect of both preventive and curative intervention programs, to monitor progress towards elimination, and to study the dynamics of variants associated with specific phenotypes.
Several molecular assays for identifying populations of L. donovani have been developed. Among these are sequencing of ribosomal loci [4], multi-locus microsatellite typing [5], multilocus sequence typing [6], amplified fragment length polymorphism analysis [7], and kinetoplast minicircle DNA RFLP (restriction fragment length polymorphism) analysis [8,9]. These methods are capable of discriminating a number of genotypes, but are less effective in recently evolving epidemic parasite populations which are relatively homogeneous, such as L. donovani in the ISC. A phylogenomic study, based on whole genome sequences of over 200 clinical ISC parasite isolates, allowed documenting genetic diversity on a much finer scale [10]. Based on single point mutations (SNPs), ten major ISC populations (named ISC001-ISC010) could be defined. Seven of them (ISC001-ISC007) represented congruent monophyletic groups, while the remaining three (ISC008-ISC010) contained mixed signatures of these.
Because whole genome sequencing (WGS) cannot currently be applied in all laboratories, our goal was to develop simpler molecular assays for discriminating ISC001-ISC007 in individual isolates, based on PCR amplification followed by amplicon sequencing, further called ISC single locus genotyping (ISC-SLG). This will allow application in epidemiological surveys and transmission studies. In case of genotype ISC005, such assay could also have a clinical relevance, as this genotype was shown to correlate with antimonial drug resistance and treatment failure [10]. After development of the assay and evaluation of its analytical performance, we applied it to Nepalese isolates that were collected in the years following the WGS survey, and analyzed the spread of genotypes over time and space.
Ethical clearance was obtained from the institutional review boards of the Nepal Health Research Council, Kathmandu, Nepal and the corresponding body of the Institute of Tropical Medicine, Antwerp, Belgium.
The sample collection consisted of a total of 204 Leishmania isolates from 195 confirmed VL patients who presented between 2002 and July 2014 at the B. P. Koirala Institute of Health Sciences (BPKIHS), a tertiary care medical center in Dharan, Nepal. The clinical criteria were fever for more than 2 weeks, combined with hepatomegaly and/or splenomegaly. The laboratory criteria were a positive rK39 rapid diagnostic test (InBiOS, Cat. nr. INS015) [11,12], and bone marrow smear positivity. Written informed consent was obtained from patients, or from parents or guardians in case of children.
Detailed geographical and clinical information of each patient is provided in S1 Database. Among the 195 patients, 9 were treated twice because they either did not respond to the first treatment, or relapsed. The period during which the second treatment was administered is referred to as second episode, hence our study included 204 disease episodes. Of these, treatment history of 26 was not traceable. The remaining episodes were treated with antimonials (Sodium stibogluconate, SSG), Miltefosine (MIL), or Amphotericin B (AmB) in 33, 85, and 57 cases, respectively, while in 3 cases treatment was not completed. Treatment failure was recorded for 42 episodes, meaning that patients either did not respond to treatment (persisting clinical signs and symptoms of VL, positive bone marrow smear after treatment), or they relapsed after initial cure (reappearance of disease symptoms and/or positive bone marrow smear during 12 months follow-up). Of these 42 episodes, SSG was used in 10, MIL in 29, and AmB in 3 cases.
Parasite promastigotes were derived from bone marrow aspirates of VL patients by culturing in Tobie’s blood agar medium with Locke’s overlay [13], with 200 IU/ml penicillin and 200 μg/ml streptomycin. Once the parasites were fully grown from the clinical material, they were transferred to M199 (Sigma-Aldrich, cat. nr. 2520) with 20% fetal calf serum (Invitrogen, cat. nr. 10270). The parasite cultures were grown to late logarithmic growth phase and cryopreserved at -80˚C with 10% sterile glycerol.
A total of 204 parasite cultures were isolated at BPKIHS, one for each disease episode. Out of these, 98 had been previously analyzed by WGS [10], while 106 were genotyped in this study with ISC-SLG. DNA was extracted from parasite cultures using the QiaAmp DNA mini kit (Qiagen, www.qiagen.com). Parasites in late logarithmic growth phase were washed thrice with sterile PBS solution and DNA was eluted in 200 μL AE buffer. DNA concentration and purity was verified by spectrophotometric measurement with the NanoDrop 2000 (Thermo Scientific, Waltham, MA, USA). The species of L. donovani was confirmed using PCR-RFLP analysis of the heat-shock protein 70 gene (hsp70). The fragments referred to as HSP70-N [14] were digested with restriction enzymes HincII (the isoschizomer of HindII) [15], and MluI.
Previous WGS analysis of L. donovani from ISC identified 10 populations [10], seven of which (ISC001-ISC007) being characterized by a unique combination of apomorphic homozygous SNPs or INDELs (insertions/deletions). The remaining three (ISC008-ISC010) represented composite genotypes. For each of the genotypes ISC001-ISC007, a unique apomorphic homozygous SNP or INDEL was selected for designing a specific assay. PCR primers were chosen to amplify about 500 nucleotides flanking each SNP/insertion at both sides. The same primers were used for the amplicon sequencing.
PCR assays were done in 50 μl final reaction volume which contained 1x PCR buffer with a total of 2 mM MgCl2, 200 μM of each dNTP, 1 μM of each primer, and 1.5 units of HotStarTaq Plus DNA polymerase (Qiagen, Cat. nr. 203605). Finally, 0.1 to 1 ng of L. donovani DNA was added. Since SNPs of ISC001 and ISC002 were located in a high GC% locus, 1x Q-solution was used in both these PCRs to decrease secondary structures. The thermo-cycling program was (i) initial denaturation at 95°C for 5 minutes; (ii) 36 cycles of denaturation at 94 °C for 1 minute, annealing at 60°C for 30 seconds and extension at 72°C for 45 seconds; and (iii) final extension at 72°C for 10 minutes. In addition, two positive controls (1 ng and 0.1 pg parasite DNA of MHOM/NP/2003/BPK282/0cl4) and two no-template controls were included in each experiment. The amplified PCR-products were verified on a 2% agarose gel prior to sequencing.
The analytical sensitivity was determined for each PCR using reference strain MHOM/NP/2003/BPK282/0cl4. Ten-fold DNA dilution series in water were examined, ranging from 20 ng to 0.2 pg added as template. The concentration of the reference DNA was determined using spectrophotometric measurements with the Nanodrop machine (www.nanodrop.com).
PCR amplicons were shipped to MACROGEN (www.macrogen.com, Seoul, South Korea) for capillary sequencing (ABI3730XL DNA Analyzer, Applied Biosystems) with either the forward or reverse PCR primer. Chromatogram trace files were aligned with the corresponding reference sequences from WGS data in order to identify the SNPs or insertion in the amplicon. Because we sequenced only one strand, we ensured having an excellent quality read at the nucleotide position in question, and no insertions or deletions outside the query area when aligning to the reference sequence.
GPS coordinates of the villages where patients were living were taken from Google maps (www.google.com.np/maps) [16] and elevations were extracted from the R3.0.3 software (www.R-project.org) [17] using the “rgbif” package. Bubble plots were made with R3.0.3 using the “ggplot2” package. GPS plots were generated with QGIS version 2.8.7-Wien (www.qgis.org) [18] and Google earth (www.google.com/earth) [19].
For each of the WGS-defined populations ISC001-ISC004 and ISC006-ISC007, a unique homozygous SNP was selected: each SNP was thus an apomorphic character for a given population, also not found in ISC005 or in any of the composite genotypes ISC008-ISC010. In the case of ISC005, an apomorphic two-nucleotide insertion (GA) in the Aquaglyceroporin-1 gene was selected. Table 1 lists all selected SNPs and INDELs, with their characteristics and chromosome location.
The PCR primers that were designed to amplify the specific SNP/INDEL positions with their flanking regions are given in Table 2. Also indicated are the PCR primer that was used for sequencing, and the analytical sensitivity as determined on dilution series of parasite DNA. The analytical sensitivity is illustrated in Fig 1 and ranges between 2 pg and 0.2 pg.
A total of 204 clinical parasite isolates of L. donovani obtained in Nepal were included in this study (detailed information in S1 Database), 98 of which were collected between 2002 and 2011 and were previously analyzed with WGS [10]. For the present study, we used ISC-SLG to classify the remaining 106 isolates collected between mid-2011 and 2014 in genotypes ISC001 (n = 21), ISC003 (n = 9), ISC004 (n = 8), ISC005 (n = 8), ISC006 (n = 16), and ISC007 (n = 0), while 44 could not be classified. This was done in a sequential manner, whereby the assays for the most common Nepalese genotypes were performed first. Isolates that could not be assigned to a genotype at this stage were further analyzed for less common genotypes, and so on, till all six assays were used. If an isolate could not be categorized after running the six assays, it was listed as “unclassified genotype”. Even though we also developed ISC-SLG for ISC002, we did not test for this genotype as it was so far only reported from Bangladesh [10], and migration between both countries is restricted by travel visa requirements. Detailed results of the WGS and ISC-SLG analyses are provided in S1 Database. Of the 204 isolates, 7 originated from patients who were treated a second time, and that showed the same genotype as found in the first disease episode. When these 7 are not taken into account, as they do not represent independent data points, we found 33 ISC001, 16 ISC003, 33 ISC004, 15 ISC005, 43 ISC006, and 8 ISC009 isolates, while 49 could not be classified (6/93 analyzed with WGS, 43/104 analyzed with ISC-SLG).
Fig 2 illustrates the distribution of genotypes over the period of sampling. Genotypes ISC001 and ISC003-ISC006 were present in most sampled years, with some exceptions. But even when they were not found in a particular year, they reappeared later on, proving that all of them circulated continuously in the region. For a given year, the most abundant classified genotype varied: ISC004 in 2002–2003, ISC006 from 2004 to 2012, and ISC001 in 2013–2014. ISC009 was the rarest genotype and was not identified after 2011 because no ISC-SLG test is available. ISC007 was not found in any sample.
The distribution of genotypes according to the treatment outcome of patients with three different drugs (SSG, MIL and AmB) is shown in Fig 3, whereby only episodes where treatment was completed are included. With respect to SSG, treatment failure was observed only in ISC004-ISC006, but most patients infected with ISC004 and ISC006 were cured. In contrast, none of the patients infected with ISC005 cured after SSG treatment. For MIL, relapse was observed for the 6 identified genotypes but never for the unclassified ones. Half of four MIL treated relapsing patients for which genotyping was done at relapse stage, showed the same genotype before and after relapse. Two of them however did not: BPK519 and BPK676 were initially infected with ISC003, but at the time of relapse respectively ISC001 and ISC006 were isolated. The three cases of AmB relapse were associated with ISC004, ISC006, and an unclassified genotype.
VL patients included in this study were from 15 administrative districts. Fig 4 shows the presumed geographical origin of typed parasites over time, with the assumption that patients got infected in their home village. An overview of the numbers per year and per district is given in S1 Table. Seven patients for which identical genotypes were recovered before and after relapse were included only once in this analysis. Genotype ISC001 was mostly present in hill districts such as Bhojpur, Dhankuta, Khotang, Okhaldhunga, Sankhuwasabha, and Udayapur. Other genotypes were unevenly distributed in the lowland area (Terai). The highest number of different genotypes was found in the three districts Morang, Sunsari and Saptari.
Several Leishmania populations have been documented to circulate in the Indian Subcontinent [10]. Keeping track of their distribution over time is of particular importance for parasite elimination by control and treatment measures. Indeed, tracking the spread of populations allows tracing parasite movements from one area to another, assisting in monitoring control efforts. In addition, parasites from different populations could react differently to treatment, be carried by different vectors, or differ in response to vaccines should they become available. Identifying L. donovani populations in the ISC is however challenged by the fact that they are genetically highly homogeneous: WGS revealed no more than 2418 SNPs in 191 isolates of the core population [10]. Previously used methods such as microsatellites, which are effective in other areas in the world, have shown limited value in this region [5]. Data from WGS offered the additional resolution needed for distinguishing different populations based on single nucleotide polymorphisms [10]. Although WGS could be used for large-scale epidemiological surveys or clinical studies, it remains currently too expensive, has limited availability, and requires too much bio-informatics analysis time. The ISC-SLG method presented in this paper partly solves these limitations, as it requires only PCR and classic Sanger sequencing methods, which are far easier to perform and interpret than WGS, at a significantly smaller cost.
As a proof-of-principle, we investigated the geographic spread of different parasite genotypes over a period of more than 10 years (2002–2014), combining data from WGS and ISC-SLG of L. donovani in Nepal. Based on these data we could show that particularly ISC001 has a different geographic distribution as compared to the other genotypes, and is present predominantly in hilly areas. This could be related to a local epidemic, which is corroborated by a recent outbreak investigation showing that transmission does take place at higher altitudes [20]. ISC001 is genetically very different from parasites belonging to the main population endemic in the Terai lowland, and shows different phenotypic features, hence its epidemiological follow-up is highly relevant. No obvious pattern could be seen for the other genotypes, which are found in the Terai (lowland). As this analysis was merely a proof-of-concept, no firm conclusions should be drawn. The data collection was not set up for such precise monitoring, and only confirmed VL cases visiting the BPKIHS medical services were included. In addition, the sampling should be extended to other countries of the ISC.
An assessment was presented on differential treatment response from the different genotypes. Apart from ISC005, which systematically did not react to SSG, no obvious patterns were seen. The treatment failure of ISC005 infected patients does not come as a surprise, because all parasites of this population have a defective aquaglyceroporin transporter [21]. This transporter is needed for uptake of trivalent antimonials, which are the active derivatives of pentavalent antimonial drugs such as SSG. It is exactly the underlying 2-nucleotide insertion responsible for the inactivating frameshift that was used for identification of ISC005. As for the other treatments, AmB was effective in all genotypes, and MIL frequently resulted in relapse, also across genotypes. Regarding the latter, no relapse was seen in unclassified genotypes, which remains to be elucidated: identification of these samples is needed, as well as a sufficiently powered study.
In two patients, different genotypes were detected at the onset of disease and at relapse stage after MIL treatment. Several scenarios could explain this. First, the patients might have been simultaneously infected with 2 different genotypes initially, of which one outgrew the other during culturing. After treatment, the first detected parasite could have been eliminated, allowing the other to be isolated at the time of relapse. Alternatively, the patients might have been re-infected with a different genotype, causing the relapse.
As WGS-based methods can be considered the gold standard for identifying parasite populations because they use nearly all nuclear sequence information, our study can assess the power of alternative methods based on single or few genes. A previous report, including 31 strains from our analysis, classified Nepalese genotypes with kinetoplast DNA mini-circles, microsatellites, cysteine proteinase B, glycoprotein 63, and the hydrophilic surface protein B [9]. This analysis identified 14 kDNA and 8 nuclear genotypes, which in general correlated well with WGS-defined genotypes (data provided in S1 Database), even though there was no perfect one-to-one relationship. This could be explained by the fact that microsatellites tend to suffer from homoplasy [22], mini-circles are highly variable and difficult to analyze in a reproducible manner, and genetic homogeneity calls for analysis of many loci in parallel.
We applied ISC-SLG on parasite cultures, but so far not on clinical samples. Nevertheless, the analytical sensitivity of the assays was between 0.2 and 2 pg, or roughly 1 to 10 parasite genomes, which makes the assay at least theoretically usable on VL samples with sufficient parasite load, such as bone marrow aspirates, spleen/liver biopsies, and even blood [23]. In addition, also tissue from Post-Kala-azar dermal leishmaniasis (PKDL) patients was shown to surpass this threshold [24]. Also, currently efforts are ongoing to specifically select or amplify low abundant parasite DNA from clinical samples, which may provide new opportunities to extract useful genetic information from these [25].
Improvement of ISC-SLG could involve the use of multiplex and real-time PCR. Currently, each genotype requires a different PCR, and these could be combined in a single tube by multiplexing to increase speed and reduce cost. Moreover, as illustrated in this paper, analysis can be done in a sequential manner, whereby samples are first tested for the most abundant genotype(s) instead of running all assays simultaneously. Further, to enhance throughput and compatibility with existing infrastructure in low-resource settings, sequencing could be avoided by using fluorescent real-time PCR probes as shown by Feehery et al. and Tyagi et al. [26,27].
It should be noted that ISC-SLG can detect documented genotypes/populations, but does not allow identification of new ones or alternative genotypes showing the same polymorphism as used in the assay. Indeed, 42% of tested parasite isolates could not be classified. These could either represent hitherto undetected genotypes, or populations for which no ISC-SLG test was developed or used. Nevertheless, highly useful information can be obtained even by tracking half of the circulating genotypes, as it can provide insight into the way of spreading, and hint to invasion of new genotypes in a particular area. Hence, WGS remains complementary to our method, allowing to investigate new genotypes and serving as a basis for developing additional ISC-SLG assays.
In conclusion, ISC-SLG was successfully applied to cultured isolated parasites from various locations in Nepal. The method needs further improvement, but is already in its current version a practical tool for use in clinical studies and epidemiological surveys. As such it contributes to “research on epidemiology and transmission dynamics of VL”, one of the key research priorities for visceral leishmaniasis elimination recently suggested by Singh et al. [28]. Applying our technique in country-wide systematic epidemiological surveys would help to better understand and control leishmaniasis in Nepal. The method could easily be extended to additional genotypes circulating in other regions of the ISC, in support of the Kala-azar elimination program.
|
10.1371/journal.ppat.1000855 | A Timescale for Evolution, Population Expansion, and Spatial Spread of an Emerging Clone of Methicillin-Resistant Staphylococcus aureus | Due to the lack of fossil evidence, the timescales of bacterial evolution are largely unknown. The speed with which genetic change accumulates in populations of pathogenic bacteria, however, is a key parameter that is crucial for understanding the emergence of traits such as increased virulence or antibiotic resistance, together with the forces driving pathogen spread. Methicillin-resistant Staphylococcus aureus (MRSA) is a common cause of hospital-acquired infections. We have investigated an MRSA strain (ST225) that is highly prevalent in hospitals in Central Europe. By using mutation discovery at 269 genetic loci (118,804 basepairs) within an international isolate collection, we ascertained extremely low diversity among European ST225 isolates, indicating that a recent population bottleneck had preceded the expansion of this clone. In contrast, US isolates were more divergent, suggesting they represent the ancestral population. While diversity was low, however, our results demonstrate that the short-term evolutionary rate in this natural population of MRSA resulted in the accumulation of measurable DNA sequence variation within two decades, which we could exploit to reconstruct its recent demographic history and the spatiotemporal dynamics of spread. By applying Bayesian coalescent methods on DNA sequences serially sampled through time, we estimated that ST225 had diverged since approximately 1990 (1987 to 1994), and that expansion of the European clade began in 1995 (1991 to 1999), several years before the new clone was recognized. Demographic analysis based on DNA sequence variation indicated a sharp increase of bacterial population size from 2001 to 2004, which is concordant with the reported prevalence of this strain in several European countries. A detailed ancestry-based reconstruction of the spatiotemporal dispersal dynamics suggested a pattern of frequent transmission of the ST225 clone among hospitals within Central Europe. In addition, comparative genomics indicated complex bacteriophage dynamics.
| Because fossils of bacteria do not exist or are morphologically indeterminate, the timescales of bacterial evolution are widely unknown. We have investigated the short-term evolution of a particular strain of methicillin-resistant Staphylococcus aureus (MRSA), a notorious cause of hospital-associated infections. By comparing 118 kilobases of DNA from MRSA isolates that had been collected at different points in time, we demonstrate that this strain has accumulated measurable DNA sequence variation within two decades. Further, we exploited this sequence diversity to estimate the short-term evolutionary rate and to date divergence times without paleontological calibration, and to reconstruct the recent demographic expansion and spatial spread of this MRSA.
| Clinical microbiologists have frequently been astonished by the impressive capability of pathogenic bacteria to acquire novel traits such as antimicrobial resistance. However, the actual speed at which nucleotide substitutions, entire genes, or complex mobile genetic elements are gained and lost in bacterial populations has rarely been determined [1],[2],[3],[4]. A measure of the real-time nucleotide substitution rate in natural populations of pathogenic bacteria would enable the dating of evolutionary events and the reconstruction of a pathogen's demographic history based on DNA sequence variation, which ultimately could provide fundamental insights into the forces driving pathogen emergence and spread [2],[5].
Methicillin-resistant Staphylococcus aureus (MRSA) are a common cause of hospital-acquired infections, imposing a heavy burden on patients and health care resources [6]. The prevention and treatment of such infections has become increasingly difficult due to this bacterium's ability to acquire resistance against all classes of antibiotics. Staphylococcus aureus has long been known to cause local outbreaks and regional epidemics of hospital infections, where the causative strains – identified through bacterial typing – may spread both within and across hospital wards, and among different hospitals [7]. Contemporary typing of S. aureus is performed by using molecular techniques, including DNA macrorestriction (pulsed field gel electrophoresis) and DNA sequence-based methods. Among the latter, multilocus sequence typing (MLST), which indexes variation at seven slowly evolving genetic loci, has been extremely useful to gain a basic understanding of the population structure of S. aureus [8]. While more than 1,400 MLST-based sequence types (ST) have been reported for S. aureus to date, most of this diversity is clustered in a limited number of clonal complexes [8]. The worldwide predominance of a few clonal lineages among MRSA has resulted in the conception that MRSA strains may spread globally very rapidly [9],[10]. However, by investigating the diversity and phylogeography of one such clone (ST5) in greater detail, we have recently detected considerable spatial subdivision among populations from different localities, indicating that the dispersal of this clone over long distances happens rarely in comparison to the frequency at which novel MRSA arise through acquisition of the genetic methicillin-resistance island SCCmec [11].
In the present study, we have investigated the evolutionary history of an MRSA strain that recently emerged in Central Europe. By MLST, this strain is identified as sequence type ST225 (allelic profile, 1-4-1-4-12-25-10), which is a single locus variant of ST5, the presumed ancestor of clonal complex CC5 [8]. While ST225 had been discovered first among isolates collected during the 1990s in the USA [8],[12], it was not detected in any European country before the year 2000 [13],[14],[15],[16],[17],[18]. Since 2001, however, its reported proportional abundance in Germany increased very rapidly [14], and it was also reported from hospitals in neighboring countries [19],[20]. Hence, this strain has a demonstrated ability to spread rapidly and to become predominant in the hospital environment, thereby replacing other MRSA strains that heretofore had been established for years [14]. At the same time, ST225 seems almost entirely restricted to the hospital environment, since it has not been reported from asymptomatic S. aureus carriage outside of hospitals and it is very rarely found among isolates from community-associated MRSA infections; in the latter, sporadic cases, close contacts to hospital patients or staff could not be excluded [21],[22].
We analyzed an international sample of MRSA type ST225 sequenced at 118,804 basepairs per isolate. Based on serial, time-structured samples of DNA sequences, we observed the accumulation of genetic diversity over a few years. By using coalescent (i. e., genealogy-based) methods, we calculated divergence times and reconstructed the pathogen's past demography. Our results are consistent with a scenario of a recent reduction in population size that has caused losses of genetic variation, and a subsequent population expansion of ST225 within Central Europe.
Isolates affiliated to ST225 – including both, MRSA and methicillin-susceptible S. aureus (MSSA) – display very limited genotypic and phenotypic variability based on contemporary, molecular typing techniques and antimicrobial resistance (Table S2). We used denaturing high-perfomance liquid chromatography (dHPLC) to screen for sequence polymorphisms at 269 genetic loci (predominantly randomly chosen housekeeping genes) from each of 73 S. aureus isolates (Tables S2, S3). Genome fragments investigated were scattered along the S. aureus chromosome and altogether comprised 4.2% (118,804 basepairs) of the genome (Table S3). Polymorphisms were ascertained through subsequent sequence analysis (Table S4a). All isolates belonged to sequence type ST225 or a single locus variant thereof (ST710) and had been isolated between 1994 and 2007 in the USA, the Czech Republic, Denmark, Switzerland, and Germany (Table S2). These analyses revealed 48 bi-allelic polymorphisms (BiPs; i. e., polymorphic sites at which exactly two alleles were observed), including 11 synonymous base substitutions in protein-coding regions, 26 non-synonymous substitutions, 10 substitutions in intergenic regions, and one insertion of a single nucleotide (Tables S4a, S5). The nucleotide diversity, π (the average number of nucleotide differences per site between sequences from two isolates), was 0.00001 for coding regions and 0.00003 for non-coding regions (Table S1). This level of diversity is extremely low; in a similar study on a global sample of S. aureus sequence type ST5 (the founder of clonal complex CC5), we recently discovered ten-fold higher diversity in both, protein-coding and intergenic regions [11]. In 70 ST225 isolates from Europe, we found 41 BiPs, which corresponds to 0.6 BiPs per isolate or 28 differences between any two 2.8 Mbp genomes. A similar level of divergence was recently reported for community-associated MRSA strain ‘USA300’, which, on average, displayed 35 differences between any two out of eight re-sequenced genomes [23]. The dN/dS value (the ratio of changes at non-synonymous sites to changes at synonymous sites) for protein-coding genes in ST225 was 0.77, hence, similar to the value found for ST5 [11]. This high proportion of non-synonymous substitutions is unlikely to represent a signal of selective pressures, but is a consequence of the dynamics of short-term evolution (i. e., evolution which occurs within a few years, see below) [11],[24],[25].
The 48 BiPs enabled the discrimination of 36 haplotypes (i. e., unique combinations of BiP alleles) among the 73 isolates investigated (Table S2). There were only five parsimony informative sites (where derived alleles occurred in >1 haplotype), and four of these were found in isolates from the USA. Consequently, most of the variation was unique to individual haplotypes, and little phylogenetic structure was discerned among European ST225 isolates (Figure 1). The minimum spanning tree based on these BiPs shows a star-like radiation that is rooted at a hypothetical node representing the most recent common ancestor of ST225 and the JH strain (ST105; Figure 1a). This ancestor is affiliated to lineage ST5-K within the ST5 radiation (Figure 1b). It carries a number of derived alleles (listed in Table S4b) that distinguish it from ST5 haplotypes, in agreement with the previous presumption that ST5 was the ancestral genotype within the clonal complex CC5 [8].
All ST225 MRSA isolates that we have investigated, including those from the US, carry a unique 997-basepair deletion in their SCCmec cassettes, which encompasses a 0.3-kb open reading frame (N315-SA0035) and the adjacent direct repeat unit (dru) locus. Deletions of the dru locus have rarely been reported [26],[27]. The presence of this characteristic feature in SCCmec indicates that the most recent common ancestor of the ST225 radiation had already been methicillin-resistant, which suggests that the entire radiation is younger than a few decades. The same dru deletion was present in the genome of the closely related JH strain (ST105, represented by isolates JH1 and JH9 [28], Figure 1), indicating it also existed in the common ancestor of ST225 and ST105, which, hence, already was methicillin-resistant. In addition, we found identical recombinase (ccrB) and helicase (cch) gene sequences in SCCmec from all ST225 MRSA isolates and from the JH genome (not shown), supporting the notion of a common origin. The dru deletion in international isolates also indicates a history of long-distance dissemination of MRSA, since sequence identity in this region would be unlikely if SCCmec elements had been imported repeatedly into locally endemic, methicillin-susceptible ST225 strains. Notably, our methicillin-susceptible isolates could not be distinguished from MRSA based on BiPs (Table S2), lending support to the presumption that they represent strains that have lost methicillin resistance together with parts of their SCCmec elements. Three of these MSSA carried SCCmec remnants in their chromosomes which we detected by PCR and sequencing, including the region with the dru deletion (Table S2). Even those isolates with no detectable traces of SCCmec may be secondary MSSA, however, since spontaneous, precise excision of SCCmec from the staphylococcal chromosome has been reported [29],[30].
There are several arguments why our American isolates of ST225 represent the ancestral population of the European clade. First, US ST225 isolates have been observed as early as 1994 (Table S2), whereas this clone was not encountered before 2000 in Europe. Second, considerable genetic diversity is observed among US isolates even from a single federal state (Wisconsin), with seven SNPs including four parsimony informative sites observed in only three isolates (Figure 1a). This is in stark contrast with the extremely low genetic diversity in European isolates, which suggests a recent population bottleneck (i. e., a brief reduction in population size) associated with the introduction of ST225 into Europe. A population bottleneck occurs, for example, when a small number of individuals founds a new population (‘founder effect’), and may result in a significant loss of genetic variation. Third, American ST225 carry a spa sequence (spa type t002) that is presumably ancestral to spa from European ST225 (t003, t045, t456, t1107; Tables S2a, S2b); the latter spa sequences may have arisen from t002 through deletions of individual repeat units, a frequent phenomenon during DNA replication, whereas the opposite (regain of unique repeats) appears less likely. Spa type t002 was also previously considered ancestral to other spa types based on the presence of a large number of single-repeat variants [31]. Finally, the ST225 radiation branches off from the ST5-K lineage (Figure 1b), to which the majority of ST5 isolates from the USA had been affiliated as reported in our previous study [11].
Taken together, we conclude that ST225 evolved from an MRSA that already carried the dru deletion in its SCCmec element. The novel clone spread to Europe somewhat later, where it rapidly became highly prevalent. The hypothesis of a single transmission event from the US is further supported by the low diversity and the monophyletic structure of the European ST225 radiation (Figure 1). However, current data do not preclude the existence of an ancestral ST225 population outside the US, although no such isolate has been observed so far.
A plot of genetic distance from a common ancestor against sampling time gave a first indication of a measurable accumulation of DNA sequence variation over the sampling time interval (Figures 2a, 2b). Such sets of temporally spaced molecular sequences with a statistically significant number of genetic differences can be used to simultaneously estimate divergence times, temporal changes of population size, and nucleotide substitution rates by applying suitable statistical methods [32]. Based on the sequence variation ascertained, we calculated the age (divergence time) of ST225 by applying a Bayesian coalescent method of phylogenetic inference that incorporated a strict molecular clock model [33]. The relaxed molecular clock model was ruled out as it yielded a posterior distribution of clock rates showing negligible variation (with the standard deviation abutting zero), and was not statistically supported (likelihood ratio test, P = 0.99). Based on our dataset of 73 sequences, the most recent common ancestor of ST225 was estimated to 1990 (95% confidence intervals, 1987 to 1994) (Table 1). The age of the American ST225 clade coincides with the age of the entire ST225 radiation, and the European clade was estimated to have diverged since 1995 (95% confidence intervals, 1991 to 1999) (Table 1). Alternative tree priors (i. e., prior probability distributions) for the Bayesian analysis resulted in very similar time spans (Table 1). Sampling from the prior distribution, in contrast, resulted in hugely inflated divergence times (Table 1), suggesting our results are not mere artefacts reflecting the priors. While it may seem surprising that the little sequence variation discovered may suffice to calculate divergence times with such tight confidence intervals, a test based on random permutation of sampling times across isolates resulted in much older dates and much larger credible intervals (Figure 3), indicating our age calculations were based on a genuine signal in the data [34].
The Bayesian skyline plot indicates a very sharp increase of the effective population size starting in 2001, with strong growth continuing for about three years and levelling off thereafter (Figure 4a). This demographic expansion, including the timing of events, is in full agreement with our observation of ST225 abundance in Central Europe (Figure 4b). This scenario is also consistent with a rampant expansion of the clone after its trans-Atlantic spread. The skyline plot (Figure 4a) was not unduly affected by heterogeneity in sample size per year, as indicated by the analyses of ten random subsamples of sequences from each year (Figure S1). However, we cannot exclude that population growth may have been more stochastic during the 1990s than is suggested by the current skyline plot (Figure 4a). To gain more detailed insights into the population structure during this time period, it would be particularly useful to investigate additional American ST225 isolates collected between 1990 and today, which are unfortunately not available at present. The composition of our sample seems to reflect the worldwide population structure of ST225 quite well, since many thousands of MRSA isolates have been genotyped to date in many countries, but no ST225 has ever been found outside Central Europe or the US. In a recent survey based on MLST typing of over 2,000 MRSA isolates sampled from Wisconsin, we did not find a single additional ST225 isolate (unpublished results of SKS). To probe the abundance of ST225 in Germany during the 1990s, we randomly chose 200 isolates from 1997 from the culture archive of the German national reference center for staphylococci and characterized them by spa typing and MLST. None of them was affiliated to ST225, suggesting that, at the time, the strain had been either absent or very rare in Germany.
The mean nucleotide substitution rate within ST225 was estimated at 2.0×10−6 substitutions per nucleotide site and year (95% confidence intervals, 1.2×10−6 to 2.9×10−6) (Table 1). This short-term evolutionary rate varied only slightly depending on clock model and choice of priors (Table 1), and was also largely confirmed by an alternative method based on a full likelihood model assuming a perfect star genealogy, which gave a rate of 1.1×10−6 (95% confidence intervals, 7.5×10−7 to 1.4×10−6). Even higher upper limits of substitution rates in bacteria have previously been estimated for Neisseria gonorrhoeae (4.6×10−5; [2]), Helicobacter pylori (4.1×10−5; [4]), and Campylobacter jejuni (6.6×10−5; [3]). In contrast to S. aureus, however, these three species are characterized by extremely high rates of homologous recombination, and, hence, part of the polymorphisms observed might have resulted from recombination rather than mutation [2],[3],[4]. Therefore, those reported rates had been considered maximal estimates; in the case of H. pylori, 100-fold lower rates were equally likely [2],[4].
Our rate for MRSA ST225 exceeds an evolutionary rate estimate that had been proposed for Escherichia coli in the past (3×10−8 substitutions per nucleotide site and year) by almost two orders of magnitude [35]. That previous estimate had been based on a laboratory mutation rate of 10−10 per nucleotide site and generation, and the assumption of approximately 300 generations elapsing per year [35]. Mutation frequencies measured in vitro (i. e., the average fraction of individuals carrying a particular resistance mutation in a laboratory culture) are very similar in E. coli and S. aureus [36],[37], suggesting comparable underlying mutation rates (the probability of a mutation to occur in each generation). While ‘mutator’ strains with elevated mutation frequencies have been described, they seem to be uncommon among clinical isolates [38]. A mutator phenotype for ST225 is also not supported by a comparison of whole genome sequences from 04-02981 (ST225, accession number CP001844, see below) and related isolates, including N315, JH1 and JH9 (Figure S2), and additional isolates (our unpublished data). In the genome from 04-02981, we detected no inactivating mutations in any genes involved in DNA replication fidelity, DNA repair mechanisms, or recombination, which are commonly associated with mutator phenotypes [38],[39]. Instead, it seems likely that the massive clonal expansion of ST225 was associated with short bacterial generation times and frequent transmission to new hosts. During rapid demographic expansions, both genetic drift and natural selection will be reduced, thus leading to an increase in the number of mutations segregating in a population, at least transiently [40].
Our results pointing to a rapid clonal evolution of S. aureus suggest that other bacteria may evolve faster than previously acknowledged. It must be considered, however, that observed molecular clock rates are time-dependent [41]. Generally, clock rates decline from initial mutation rates to long-term substitution rates, because the majority of mutations get eliminated with time due to genetic drift and selection [41]. Such rate curves have not yet been determined for bacteria. However, our results imply that recent divergence times of bacteria were possibly overestimated with dating based on the molecular clock rate suggested by Guttman and Dykhuizen [35],[42],[43]. It will be interesting to investigate short-term evolutionary rates in additional clones of S. aureus and other bacterial species. The time dependency of these rates may be established by comparing radiations at different levels of divergence.
Interestingly, the high rates of evolutionary change we found in MRSA caused the accumulation of DNA sequence variation within a few years, a feature that heretofore had been found only in highly recombinant (panmictic) gonococcus [2] and in rapidly evolving viruses [44]. Importantly, the time-structured sampling of DNA sequences within evolutionary timescales enables the application of sophisticated analytical methods, which opens up exciting prospects for investigations of the recent evolutionary history of bacterial pathogens, together with the forces that have shaped their spatial distribution.
We have investigated ST225 isolates from four European countries (Table S2, Figure 1a) by reconstructing the most likely ancestry path between isolates to reveal the spatiotemporal dynamic of ST225 spread by applying the SeqTrack algorithm [45]. Interestingly, our results indicate that multiple haplotypes have been introduced into several countries (Figure 1a). Figure 5 represents the cumulative number of isolates from any location (bubbles) and the inferred ancestries (arrows) for successive time windows. Note that while Figure 5 represents the best-supported ancestry path given the sampled isolates, some ancestries might not correspond to actual transmission events, as the true ancestral population might not have been sampled. To avoid any overinterpretation of the results, we restrict our interpretation to the global pattern and some specific unambiguous features of the inferred ancestries.
After initial seeding into Europe, ST225 was transmitted to other locations in Germany and to additional European countries (the Czech Republic, Switzerland and Denmark). Some local ancestries (i. e., within the same city) are characterized by a relatively large genetic differentiation (Figure 5, colored dots) suggesting long-term persistence of ST225 within the same location. Another interesting feature of the reconstruction of the spatiotemporal dynamics of ST225 lies in the repeated transmission events between countries. For instance, one isolate from Denmark is assigned an ancestor from Germany with high likelihood (same genotype) at least three years after the first transmission from Germany to Denmark. The first transmission (Figure 5, time window 2004/01/22–2004/12/17) could be traced back epidemiologically to an index patient that had been transferred from a hospital in Germany into a hospital in Copenhagen, Denmark, in 2004, where the carried MRSA strain (haplotype H225-07; Table S2, Figure 1a) later caused an outbreak involving multiple patients and staff. Two additional isolates collected from the same hospital in 2006 and 2007 were affiliated to the same haplotype (Table S2), indicating the clone was still present three years after the initial outbreak. However, a second haplotype (H225-01) was indicated to have been introduced from Germany into Denmark (Figure 5, time window 2007/02/08–2007/11/20), and this is unlikely to be an artefact due to insufficient sampling within Denmark, as several local ancestry events were identified earlier within Denmark.
The SeqTrack results indicated 15 transfers among different countries within Europe (Figure 5). Considering the low informative diversity discovered, the limited number of isolates and countries investigated, and the short time span since emergence of ST225 has started in Europe, this number of detected international transfers of clones is very high. It indicates that cross-border spread of MRSA between the countries considered must have occurred frequently, and, more generally, that the turnover of hospital-associated MRSA is quite rapid even within a larger geographic region (Central Europe). Hence, the question arises how efficient geographic dissemination may be mediated. Abundant international travel will result in occasional hospitalization outside the country of residency, and potential subsequent cross-border patient transfers into the respective home countries. This route is exemplified by the introduction of haplotype H225-07 from Germany into Denmark, with the subsequent establishment of this clone in the hospital for several years. In addition, it is well documented that colonized health-care personnel may promote the spread of MRSA [46]. It is also possible that some spread of ST225 occurs outside of hospitals, even though the lack of community-associated isolates suggests the prevalence to be low [47]. Efficient containment of MRSA spread requires pro-active surveillance and eradication of colonization [46],[48].
It is unclear at present, if the success of particular MRSA strains such as ST225 may be due to fortuitous stochastic events or adaptive genetic changes. To reveal any genetic traits that distinguish ST225 from other strains of MRSA and may enable its massive expansion within short time, we sequenced the genome from one representative isolate, MRSA 04-02981 (haplotype H225-01, sequence accession number CP001844). We used both 454 (Roche) and Solexa (Illumina) technology, and closed the genome sequence by using long-PCR and Sanger-sequencing. The final genome sequence likely contains very few sequencing errors, if any, since the application of two independent sequencing approaches resulted in only six conflicting SNP calls. The genome from isolate 04-02981 was found to be co-linear with previously sequenced genomes from related isolates N315 (ST5) and JH1 (ST105) [28],[49]. There was no indication for the presence of any plasmids in isolate 04-02981.
Base substitutions were distributed evenly among genes of different functional categories (not shown). The effects that individual missense mutations may have on protein function are hard to predict in most cases. In the genomes from both, 04-02981 and the JH strain (including isolates JH1 and JH9), two open reading frames were truncated, one of which encodes an unknown, hypothetical protein (N315-SAS092) and another (N315-SA1092) encodes Smf, a protein that has been suspected to be associated with transformation competence. In addition, two open reading frames were uniquely truncated in the genome from 04-02981, encoding an adhesion factor (N315-SA1267) and the transcription regulator norG (N315-SA0104). The latter pseudogene initially appeared particularly interesting, because experimental disruption of this gene had been shown previously to result in a fourfold increase of in vitro resistance to beta-lactam antibiotics [50]. However, after applying a deletion-specific PCR (Table S6), we found that none of the other ST225 isolates in our collection had this deletion. Hence, truncation of norG is not a common trait of ST225, but rather is an idiosyncrasy of isolate 04-02981, which just happened to be the one we had chosen for genome sequencing.
The genome of isolate 04-02981 contains a stretch of 44 kilobases of DNA that is inserted in a non-coding region downstream of the sufB gene (N315-SA0778), resulting in a duplication of the 67-basepair sequence upstream of the integration site. The inserted sequence is highly similar (sequence identity, 99.5%) to an as yet unnamed prophage previously found in the JH strain at the same genomic position [28]. It shares 50% or less overall sequence similarity to other phage genomes sequenced previously, including Φ11 from S. aureus NCTC8325 [28]. The prophage contains 68 predicted open reading frames, 19 of which encode proteins for basic phage functionality, and 49 of which have unknown functions. None of them has similarities to any known or presumed virulence factors.
By using PCRs targeting five specific regions (Table S6), we detected the presence of this prophage in all European ST225 isolates investigated and in other isolates affiliated to lineage ST5-K, but not in any other ST5 strains (Figure 1b). Thus, this particular prophage is specific to lineage ST5-K and its descendants, and we thus named it ΦSaST5K. Of note, prophage ΦSaST5K was not detected in any of our three ST225 isolates from the US, and, hence, it must have been lost by their common ancestor. There is a second phage – ΦN315 – in the genome of 04-02981, which it shares with isolate N315, an MRSA from Japan that is affiliated to lineage ST5-G [11]. In the JH strain, however, ΦN315 has been replaced apparently by another, dissimilar phage [28], and JH1 and JH9 harbor two additional prophages that have as yet not been seen in any other sequenced S. aureus genomes (Figure S2, Table S8b). This comparison of only three closely related MRSA genomes already points to the existence of complex phage dynamics, with varying apparent half-lives of prophages in their respective bacterial host chromosomes.
Our data indicates that several phages are associated to ST225 and its ancestral lineage, and may have played a role for its evolution. Bacteriophages have been suspected to promote the spread of pathogenic bacteria, by using various potential mechanisms. For example, phage genes may be directly implicated in immune evasion or virulence [51], or indirectly by affecting in trans the activity of bacterial genes outside the prophage, which in turn may enhance transmission or affect other fitness-related traits [52]. Alternatively, phages may possibly impact on competition between strains of staphylococci by driving lysis of bacterial cells that do not carry a related lysogenic phage.
We have shown that a strain of MRSA has accumulated measurable genetic change within an epidemiological timescale. The high short-term evolutionary rate in this MRSA enabled the estimation of divergence times and analyses of past changes in population size based on time-structured, serial DNA sequence samples, which heretofore had been possible only for highly recombinant gonococci and viruses. Moreover, ancestry reconstruction revealed the history of geographic spread of this MRSA at unprecedented detail. Confirmation of higher than expected short-term substitution rates in a wider range of bacterial pathogens, together with the tangible prospect of whole-genome sequences for large numbers of related isolates [53],[54] could prefigure a golden age for bacterial epidemiology. Presumably, bacterial pathogens will soon be amenable to detailed investigation of their recent evolutionary history and spread. At the same time, abundant polymorphisms will be discovered that will be useful for bacterial typing in epidemiological surveillance [55],[56],[57].
Sources and properties of 73 isolates of S. aureus are listed in Table S2a. Susceptibilty to antibiotics was tested by using the broth microdilution method according to the DIN58940 instructions [58] and bacterial typing was performed as described previously [31].
Draft genome sequences were generated and assembled commercially. 454 sequencing was performed on a GS FLX machine at 454/Roche in Branford, CT, USA, providing 32-fold average coverage of the staphylococcal chromosome and resulting in 42 initial contigs with >500 basepairs. Solexa sequencing was performed on a Genome Analyzer System at GATC in Konstanz, Germany, generating paired-end reads that were mapped onto the N315 genome sequence at 49-fold average coverage. Remaining gaps between contigs were closed by PCR using Hot Taq DNA polymerase (Peqlab, Germany) or long PCR using the Expand Long Template PCR System (Roche), respectively, and subsequent Sanger sequencing (primers in Table S7). Comparisons of contigs and genomes were performed by using Kodon software (Applied Maths, Belgium). After correcting sequences at contig ends and within repetitive elements, there were 468 sequence differences to N315, including base substitutions, insertions, and deletions (Tables S8A–S8D, Figure S2). Sequence differences to N315 that were shared between ST225 and the JH strain were considered correct since matching data had been generated in an independent study [28]. For insertions in the sequenced genome, we relied on 454 data, since they could not be detected among Solexa reads mapped against the N315 genome (Tables S8A–S8D). Gene annotation was performed automatically using the RAST server [59] and corrected manually using Kodon and Artemis software [60]. The annotated genome sequence from isolate 04-02981 was submitted to GenBank (accession number CP001844).
Mutation discovery was performed as described previously [11]. PCR primers used for amplification and sequencing are listed in Table S3. A minimum spanning tree based on BiPs was constructed with Bionumerics 5.1. The ancestral node was determined by comparison to genome sequences from isolates N315 and JH1.
PCR amplification of regions including the dru deletion, the four-basepair deletion within norG, SCCmec remnants, and prophage-specific fragments, respectively, were performed by using Hot Taq DNA polymerase (Peqlab, Germany) according to the manufacturer's instructions and by using the primers listed in Table S6.
Based on an alignment of polymorphic sites in protein-coding sequences, a maximum likelihood tree was calculated by using Treefinder software (available at www.treefinder.de), applying the HKY model of DNA substitution. Rooting of the tree and linear regression of root-to-tip distances against dates of first haplotype appearance was performed by using Path-O-Gen software (available at http://tree.bio.ed.ac.uk/software/pathogen/), and the significance of the correlation was determined with SigmaPlot 11.0 (SPSS).
To assess whether nucleotide substitution rates in protein-coding sequences departed significantly from expectations under a strict molecular clock, we used a likelihood ratio test, based on a comparison of likelihood scores for maximum-likelihood trees calculated by using PAUP, with and without a molecular clock enforced. The statistical significance of the difference between likelihood scores was determined by assuming a chi-square distribution and s-2 degrees of freedom, where s was the number of sequences [61].
Evolutionary rates, divergence times, and Bayesian skyline plots were computed with the BEAST software (available at http://beast.bio.ed.ac.uk/) [62], using the HKY model of nucleotide substitution and a strict clock model (unless stated otherwise), with concatenated protein-coding sequences (108,261 basepairs) dated based on the year of isolate sampling, and with 108 iterations after a burn-in phase of 106 iterations. Markov chain Monte Carlo samples from three independent analyses were combined for estimation of posteriors, resulting in effective sample size values greater than 1,000 for all parameters. Various prior sets were used as indicated (Table 1). To test if date estimates were unduly influenced by prior assumptions, analyses were re-run (5×107 iterations) on each of five datasets generated by randomly switching sampling dates across isolates. To sample from the prior distributions, analyses were run on an empty alignment. Further, to test if the resulting Bayesian skyline plot was confounded by temporal variation in sample size, we generated and analysed (107 iterations) a series of datasets by subsampling from time classes and randomly drawing four isolates from each year.
For an alternative rate estimate, we used a full likelihood model assuming that demographic expansion was strong enough to result in a perfect star genealogy (i.e., without any coalescent events). To avoid violation of this assumption, we analysed protein-encoding loci (108,261 basepairs) from 58 European isolates exclusively, including only one isolate from each haplotype, except for the ancestral haplotype H225-01. Likelihood of the model for each locus was then given by the binomial probability of the number of mutations observed in all isolates, given the sum of the genealogical branch lengths for all isolates (i. e., date of isolate collection - date of expansion start) and a substitution rate parameter per locus and per year. A point multilocus substitution rate estimate (per nucleotide site and per year) and its 95% confidence interval were inferred based on the product of the above-described likelihood function for all loci, considering that all loci had a specific number of sites, were independent, and had a single, constant mutation rate. The procedure was written in R [63] and is available upon request to R. Leblois.
The SeqTrack algorithm [45] was used to reconstruct the most plausible scenario for the spatiotemporal spread of the ST225 clone. This new method has been developed to study the dispersal and transmission of emerging pathogens during disease outbreaks, such as the 2009 swine-origin influenza A/H1N1 pandemic [45]. SeqTrack reconstructs the most likely ancestries among sampled strains using their genotype and sampling dates. This method differs fundamentally from phylogenetics in that it does not attempt to infer hypothetical (and unobserved) common ancestors, but rather seeks to reconstruct ancestries directly from the sampled isolates. Because of the low level of genetic variability in ST225 (most strains differ by a single nucleotide from each other), we used a maximum parsimony approach to infer ancestries. Thus, the most likely ancestry path was searched for by minimizing the number of mutations between ancestors and descendents. Whenever several strains were equally likely ancestors of the isolate under consideration, we retained the one that was geographically closest. All analyses were performed using the R software [63]. Raw genetic distances between isolates (in terms of number of point mutations) were computed using the ape package [64]. SeqTrack analysis was then run using the seqtrack function implemented in the adegenet package [65].
|
10.1371/journal.pntd.0001909 | Application of PCR in Serum Samples for Diagnosis of Paracoccidioidomycosis in the Southern Bahia-Brazil | Paracoccidioidomycosis (PCM) cannot always be diagnosed by conventional means such as direct examination of histopathology or clinical samples, and serological methods, used as an alternative, still have many cases of cross-reactivity. In this scenario, molecular techniques seem to arise as a rapid approach, specific and direct that could be used in the diagnosis of this mycosis. In this study we analyzed 76 serum samples from patients in southern Bahia suspected of having paracoccidioidomycosis using a conventional PCR with primers for the ITS1 ribosomal DNA of P. brasiliensis. Of these 76 patients, 5 were positive for PCM by double immunodiffusion and/or direct examination and histopathology. To test specificity of PCR, we used human DNA and three isolates of P. lutzii (1578, 01 and ED01). Additionally, we analyzed by serial dilutions of DNA the limit of detection of the assay. The test of PCR proved specific, as only a 144 bp fragment of the three isolates of P. lutzii and no human DNA was amplified. Detection limit was 1.1 pg/µL of DNA. Despite the high detection limit and specificity of PCR none of the 76 serum samples were found positive by PCR, but a biopsy specimen obtained from one of the patients with PCM was positive. These results, albeit limited, show that PCR is not effective in detecting DNA of P. brasiliensis or P. lutzii in serum, but could perhaps be used with other types of clinical samples, especially in those instances in which conventional methods fail.
| Paracoccidioidomycosis (PCM) has been included in such diseases neglected since this impact on public health have not been measured. Except in the South and Southeast of Brazil, there are no government programs for this mycosis. After 100 years of discovery of the disease there is the need to deploy an effective and continuous program for the prevention, diagnosis and treatment. PCM requires a prolonged treatment, generally greater than 1 year. Moreover, the patients with PCM are associated with other co morbidities such as alcoholism, smoking and malnutrition.
To aggravate this scenario the serological diagnosis of PCM in many cases results in cross reactivity with other mycoses. Thus, this study aimed to develop a PCR which can be used in association with the serological techniques to improve the diagnostic these individuals. We work with plasma samples from patients in northeastern Brazil, in an area where the disease has not been reported. None of the 76 plasma samples were positive by PCR, but a biopsy specimen obtained from a patient was positive. These results reinforce that PCR have the limitation when serum or plasma is used. However, PCR can be an important diagnostic tool when conventional diagnostic methods are not successful.
| Paracoccidioidomycosis (PCM) is a deep mycosis caused by the thermo-dimorphic fungus Paracoccidioides brasiliensis, endemic in some countries of Latin America, mainly in Brazil [1]. In this disease, the fungus can remain confined in the lungs, the primary focus of infection, or spread to other organs and tissues, resulting in different clinical manifestations. The form acute and sub acute occurs in young people of both sexes, mainly affecting the reticuloendothelial system. The chronic form of PCM predominates in adult males and is characterized by presence of the fungus restricted in the lungs and/or disseminated to the mucosa, skin and lymph nodes [2].
The conventional diagnosis of PCM is based on viewing and/or isolation of the fungus in clinical specimens. However, the specimen may not always be viewed and microbiological culture is time-consuming and mostly negative [3]. Serological techniques have been employed, but there are still many cases of cross-reactivity with other fungal species [4]. In addition, false-negative results can be obtained in cases where the patient has some type of immunodeficiency [5]. Nevertheless, the scene of the PCM have been changed since the discovery of a new species, P. lutzii, which has very distinct behaviors of P. brasiliensis [6].
Therefore, a molecular approach appears to be an excellent alternative in the diagnosis of PCM. However, so far this technique has not been used routinely in the diagnosis of PCM. To this end specific primers for conserved regions have been developed, i.e., for detection of 18S, 5.8S and 28S rDNA and their regions ITS1 and ITS2, of the gp43 gene and an antigenic protein of 27 kDa [7], [8], [9], [10]. Although the gp43 gene is one of genes used in the classification of cryptic species [11] and some studies show a relationship between polymorphisms of these gene and pathogenicity [9], we believe that perhaps not all isolates of P. brasiliensis or P. lutzii have the gp43 gene.
On the other hand, ribosomal DNA is present in all isolates and there are conserved regions within this structure, thus primers for these regions would be more appropriate in terms of sensitivity and specificity [12]. Thus, we aim to develop a conventional PCR using a pair of primers specific for the known ITS1 region of ribosomal DNA of P. brasiliensis [13], and know what is the viability of this PCR in serum samples. We also made an effort to determine whether this PCR assay could be used to detect DNA of P. lutzii.
We use three isolates of P. lutzii (1578, 01 and ED01) gently provided by Professor Carlos Taborda from the Laboratory of Pathogenic Dimorphic Fungi from the Institute of Biomedical Sciences II of Universidade de São Paulo (Sao Paulo, Brazil).
DNA of the three isolates was extracted as described by Kennedy et al. [14], with some modifications. Briefly, 0.4 g yeast was macerated with liquid nitrogen and the resulting powder was transferred to a 2 mL microtube. Then 1.7 mL of extraction buffer (100 mM EDTA, 100 mM Tris, 1.5 M NaCl, 1% SDS, 2% CTAB) were added and the mixture was incubated for 20 min at 65°C (inverting the microtube every 5 min) before centrifugation (20 min at 4500× g, Centrifuge MiniSpin, Eppendorf-AG, Germany). The supernatant was collected and transferred to a new microtube. An equal volume of phenol-chloroform-isoamyl alcohol (25∶24∶1) was added, and after homogenization the tube's content was centrifuged at 4500× g for 10 min. The supernatant was collected and 0.7 volume of isopropanol (100%) and 0.1 volume of 3 M sodium acetate were added and the content mixed by gently inverting the microtube 10 times. After overnight storage at −20°C the sample was centrifuged for 10 min at 4500× g and the resulting pellet washed twice with 70% ethanol. The pellet was dried, resuspended in 200 µL MilliQ water and analyzed by electrophoresis in 2% agarose gel in TBE buffer (40 mM Tris base, 20 mM boric acid, 1 mM EDTA) at 100 V for 30 min and GelRed staining (Uniscience of Brazil, Brazil). Quantification was performed in GeneQuant pro (Amersham Bioscience, USA).
We used the primer pair described by Buitrago et al. [13] in a trial of Real-time PCR, but adapted for a classic PCR and using P. lutzii not P. brasiliensis. The direct primer (OliPbMB1) was 5′-ACCCTTGTCTATTCTACC-3′ and reverse primer (OliPbMB2) was 5′-TTACTGATTATGATAGGTCTC-3′, which generated a 144 bp fragment amplified from the region ITS1 of the rDNA of P. brasiliensis. Primers were synthesized by the Bioneer Corporation (CA, USA).
The reaction was performed in 0.2 mL sterile microtube containing 10 ng DNA, 75 mM tris-HCl (pH 8,8), 20 mM (NH4)2SO4, 1.5 mM MgCl2, 1.5 mM dNTP (Fermentas Inc., MA, USA), 0.4 pM of each primer, 1 U Taq polymerase (Fermentas Inc., MA, USA) and sterile milli-Q water, obtaining a final volume of 25 µL. The reactions were processed in the Veritas thermal cycler (Applied Biosystems, CA, USA) programmed as follows: 96°C for 5 min; 40 cycles of 55°C for 35 s, 72°C for 35 s and 96°C for 35 s; 72°C for 7 min. Positive control was DNA extracted from P. lutzii (isolate 01) and negative control was sterile water milli-Q. To visualize the reaction, 5 µL of the amplified product was applied in agarose gel 2% in TBE buffer, at 100 V for 1 h and stained with GelRed.
Detection limit of the PCR assay was established by using serial dilutions of P. lutzii DNA (from 120 ng/µL to 0.5 pg/µL) and specificity was evaluated with DNA from humans and from the isolates of P. lutzii (1578,01 e ED01).
We analyzed serum samples from 76 patients (citizens from Itabuna and Ilhéus), suspected to have PCM and no conclusive diagnosis to another fungal disease/non-fungal. Five of these patients had been diagnosed with PCM in our laboratory by the method of double immunodiffusion alone (5) or in combination with direct examination of sputum and/or histopathology (3). Furthermore, the diagnosis was endorsed by a clinical experience of the physician responsible for the patient. These five patients had the chronic form of the disease and had not started treatment before blood collection. Only one of these had associated tuberculosis/paracoccidioidomycosis. The care and clinical monitoring of these patients was conducted in city of Itabuna at the Center for Health José Maria de Magalhaes or Santa Casa de Misericordia or at the Specialized Care Center III, located in Ilhéus.
All individuals/patients were informed about the methodology and signed an informed consent according to ethical standards. The project was also approved by the Ethics in Research Committee of the Universidade Estadual de Santa Cruz (Text S1).
DNA extraction from serum samples used QAamp DNA Mini Kit (Qiagen, Hilden, Germany). We used 200 µL of sample and smaller volumes were adjusted with phosphate buffer (PBS, pH 7.4). Purified DNA was eluted in 25 µL of elution buffer and analyzed by electrophoresis in 2% agarose gel in TBE buffer at 100 V for 30 min (stained with GelRed). Quantification of extracted DNA was performed in GeneQuant pro. We also extracted DNA from a biopsy sample of one the patients with confirmed PCM diagnosis. All DNA samples were then stored at −20°C until use in assay of PCR.
The test was performed in mesh citrate-agarose (1% agarose, 0.4% sodium citrate, 0.9% sodium chloride and 7.5% glycine). In the central well apply 10 µL of exoantigen of P. brasiliensis, 339 isolated, and the side wells 10 µL of serum from patients or healthy subjects. Slides were incubated in a humid chamber at room temperature for 48 hours, with readings every 24 hours. Subsequently, they were washed with sodium citrate to 5% and 0.85% physiological solution, dried at 60°C for 24 hours and stained with coomasie brilliant blue R-250 0.15% (Vetec, Rio de Janeiro, Brazil) in water-methanol-acetic acid (4∶4∶1) for 5 minutes. Positive samples were serially diluted (1/2 to 1/256) and re-subjected to the test ID for semiquantitative analysis. As a positive control we used the reaction hyperimmune serum of rabbit anti-exoantigens of P. brasiliensis, obtained from the Immunodiagnostic Laboratory of Mycoses of the Adolfo Lutz Institute (São Paulo) and kindly provided by Dr. Adriana Vicentini Pardini.
The functionality and specificity of PCR standardized by us can be proven by visualization of the 144 pb amplicon generated when using DNA from isolates of P. lutzii (Fig. 1). The reaction with P. brasiliensis (isolate 03) also generated the same amplicon (Fig. S1). Human DNA served as specificity control and was always negative.
Our standardized PCR test was able to detect until 1.1 pg of DNA of P. lutzii (Fig. 2). The range of detection of the test was from 1.1 pg/µL to 60 ng/µL of DNA.
We analyzed seventy-six serum samples of patients with suspected PCM. Five of these patients has been diagnosed with PCM by double immunodifusion (Fig. S2), and three of them were confirmed by direct examination of sputum or histopathology (Fig. S3). In the PCR, none of the seventy-six serum specimens were positive. However, the single biopsy specimen that was tested was positive, revealing the expected 144 pb fragment (Fig. 3).
In those cases where the microscopic observation of P. brasiliensis is not successful and when the levels of antigen or antibody are very low, the use of molecular techniques such as PCR has been shown to be effective [15]. There is much antigenic variation among isolates of P. brasiliensis from the different parts of Brazil and Latin America, and as consequence differences in serologic reactivity have been observed, showing the need to work with antigens specific for isolates from a defined region [16], [17].
Considering these aspects and the fact that we use antigens of isolates from other regions in the double immunodiffusion test to detect cases of PCM in Southern Bahia and that we have not yet succeeded in isolating the fungus, the use of PCR in detection of PCM seemed more appropriate.
For these purpose, the primers pairs specific for the ITS1 region of rDNA of P. brasiliensis described by Buitrago et al. [13] seemed more suitable for application of PCR as a diagnostic tool. Therefore, this primes pairs allowed the detection of both DNA P. lutzii as P. brasiliensis, being generated the same 144 pb amplicon. However, because to this characteristic, it is not possible to distinguish which of the two species is the DNA amplified from single biopsy sample tested.
Our results reinforce the fact that PCR and PCR-based molecular techniques have a limitation when serum or plasma is used [18], [19]. This can be explained because of the fungus is rarely in the blood stream [20]. Another possibility is that in this body part the yeast and DNA of P. brasiliensis is readily phagocytosed by leukocytes, reducing the chances of finding these elements there [21].
Despite these issues, we conducted an experiment where we evaluated the possibility of directly using the serum/plasma, without extraction of DNA, in the PCR reaction (data not shown). In this case, serum/plasma could contain the yeast P. brasiliensis or P. lutzii and/or its DNA dispersed, or none at all of these elements. Considering the case of the yeast to be integrated, we mixed a plasma sample a predetermined amount of the yeast P. brasiliensis and submitted to the PCR reaction but the result was negative.
Interestingly, when conducting another test with a plasma sample received 10 ng of DNA of P. brasiliensis, the reaction was positive. This experiment leads us to two hypothesis: that in plasma not exist inhibitory elements of the PCR reaction or that these elements exist but are at low levels and are further diluted when mixed with other components of the PCR reaction mix. More tests are needed to confirm this finding to be applied in the PCR reactions in general.
In summary, we include some information about the limitations (Text S2) and experimental design of this study (Fig. S4). We believe that PCR can be used as a diagnostic tool in diagnosis of PCM, especially when conventional diagnostic methods are not successful, but much remains to be done to make this molecular test secure. Moreover, we know that quantitative molecular methods, such as real-time PCR, would be more appropriate to predict disease or infection. However, this is an expensive technique and would not be performed at some locations. Thus, more studies are needed that aim to refine and increase the discriminatory power of more feasible molecular techniques as conventional PCR or nested-PCR.
|
10.1371/journal.pbio.1001493 | Mechanistic Insights Revealed by the Crystal Structure of a Histidine Kinase with Signal Transducer and Sensor Domains | Two-component systems (TCSs) are important for the adaptation and survival of bacteria and fungi under stress conditions. A TCS is often composed of a membrane-bound sensor histidine kinase (SK) and a response regulator (RR), which are relayed through sequential phosphorylation steps. However, the mechanism for how an SK is switched on in response to environmental stimuli remains obscure. Here, we report the crystal structure of a complete cytoplasmic portion of an SK, VicK from Streptococcus mutans. The overall structure of VicK is a long-rod dimer that anchors four connected domains: HAMP, Per-ARNT-SIM (PAS), DHp, and catalytic and ATP binding domain (CA). The HAMP, a signal transducer, and the PAS domain, major sensor, adopt canonical folds with dyad symmetry. In contrast, the dimer of the DHp and CA domains is asymmetric because of different helical bends in the DHp domain and spatial positions of the CA domains. Moreover, a conserved proline, which is adjacent to the phosphoryl acceptor histidine, contributes to helical bending, which is essential for the autokinase and phosphatase activities. Together, the elegant architecture of VicK with a signal transducer and sensor domain suggests a model where DHp helical bending and a CA swing movement are likely coordinated for autokinase activation.
| Two-component signal transduction systems (TCSs) are promising targets for new antimicrobial research because they help bacteria and fungi adapt and survive. One of the main components of TCSs is a sensor histidine kinase (SK), which relays extracellular signals to intracellular pathways. Despite intensive research, a full-length structure of an SK has yet to be solved. In this study, we report the first crystal structure of the complete cytoplasmic region of VicK, an important SK in the tooth decay pathogen S. mutans. VicK is composed of several domains (HAMP, PAS, DHp, and catalytic and ATP binding domain [CA]) in addition to a short transmembrane domain. We find that the dimeric VicK protein has an elegant rod-shaped structure with the domains linearly connected like beads on a string. The structure suggests that VicK kinase activates itself by helical bending of the DHp domain and coordinated swinging around of the catalytic CA domain to engage with the target histidine. Structure-based mutagenesis experiments also helped us to identify key residues that are required for VicK's opposing phosphatase activity. Our studies of the multi-modular VicK protein suggest a sequential kinase activation model that may involve helical bending of the DHp domain and repositioning of the CA domains.
| Protein phosphorylation is an essential signal carrier. Bacteria respond to transient living environments through transmembrane-integrated sensor histidine kinases (SKs), which act in concert with their intracellular cognate response regulators (RRs) to elicit necessary adaptive responses that are critical for their survival and virulence. The SKs and RRs have evolved into a two-component signal transduction system (TCS), whereby stimulation of the SK autophosphorylates at a conserved histidine residue to initiate a signaling cascade [1]. The phosphoryl group is transferred from the SKs to their cognate RRs, some of which lead to quickly reprogram bacteria by altering the transcriptional level of specific downstream target genes [2]. Because of the wide prevalence in bacteria and fungi, TCSs have been considered attractive targets for the development of potential therapeutics to control bacterial infections [3],[4].
Sensor domains are key modulators for SKs [5]–[7]. PAS domains (acronym for Per, ARNT, and SIM from Drosophila) are sensors in a majority of SKs, which respond to alterations in the redox potential, oxygen content, light, and small molecules in their environments [8],[9]. Because of their broad involvement in biological processes, the structure and function of the PAS domains in interactions with a variety of ligands have been extensively studied [10],[11]. The oligomeric dynamics of PAS domains in cooperation with local conformational changes can affect the stability of the entire SK, which is thought to be part of the mechanism of signal sensing and transduction [10],[12].
The enzymatic activities of SKs are modulated by HAMP domains, which are commonly found in histidine kinase, adenylyl cyclase, methyl-accepting chemotaxis, and phosphatase proteins [6]. Structural dynamics of the HAMP domain are believed to mediate transmembrane signal transductions [13]. The NMR structure of a HAMP domain from the putative transmembrane receptor Af1503 in Archaeoglobus fulgidus revealed an unusual knobs-to-knobs interhelical structure, which suggests a coordinated helical rotation model for the HAMP domain in signal transmission [14]. This model was further supported by a series of experimental structures of the HAMP mutants and detailed bioinformatics analyses [15],[16]. Several other laboratories reached a consensus conclusion using homologous HAMP domains from different TCSs that the dynamic properties of the HAMP domains are essential in mediating signal transduction [17]–[22].
The C-terminal catalytic and ATP-binding domain (CA) of the SKs, also called HATPase_c, in addition to the DHp domain (dimerization and histidine phosphorylation domain), phosphorylates a conserved histidine residue in the middle of DHp helices [2],[23]. The active site of an SK is assembled with the CA and DHp domains [23],[24]. The plasticity of the DHp domain and CA positioning is implicated in the on/off switch of the temperature sensor DesK kinase from Bacillus subtilis [25].
VicRK is a well-characterized TCS that is highly conserved and essential for survival and virulence in a wide range of firmicute bacteria, including Streptococci, Bacilli, and Staphylococci [26]. Because of its critical role in cell wall synthesis, VicRK is also referred to as WalRK [27]. The VicRK in S. mutans, which is an important pathogen in caries etiology, regulates acid production and tolerance conducive to dental caries, including proton expulsion (F1F0-ATPase) [28]–[31]. VicK belongs to the type IA family based on sequence conservation of the DHp and CA domains [32]. In addition, VicK has one HAMP domain and one PAS domain. However, the function of the PAS domain is not clear because no ligand has been identified [26]. Recently, a deletion experiment indicated that the HAMP and PAS domains are essential for VicK phosphatase activity [33].
Given the vital role of TCSs in bacterial adaptation and pathogenicity, decades of research have focused on the molecular mechanisms of TCS signaling cascades [34]–[36]. Although the structure and functions of individual domains are well known, the signal transduction mechanisms remain largely unknown. Toward this end, we determined a crystal structure for a streptococcal VicK that harbors HAMP transducer and PAS sensor domains. Our crystal structure of the nearly full-length VicK comprises an elegant construction of multiple domains and reveals novel insights into the molecular mechanisms of the VicK histidine kinase.
The S. mutans VicK has one transmembrane domain (TM, aa 9–30) that anchors itself to the cytoplasmic membrane (Figure 1A). Following the TM domain, a HAMP signal transducer domain and PAS sensor domain are directly connected to the CA domain through a DHp domain. As several attempts to express full-length VicK (aa 1–450) resulted in insoluble protein, the entire intracellular region (TVicK, aa 31–450) was successfully purified and crystallized. TVicK had a Km of 44.5 µM and a Kcat of 0.413 min−1 (Table S1). Its kinetic parameters were similar to the TM-truncated VicK homologue of S. pneumoniae [33]. Static light scattering revealed a single species of VicK at a molecular weight of 103.2 kDa, indicating that the holoenzyme exists as a stable dimer in solution (Figure 1B).
VicK crystals diffracted slightly better than 3 Å; however, because of strong anisotropy, the final structure was refined up to only 3.3 Å (Table S2). The overall structure of VicK comprises a dimer in the shape of a long slim rod (Figure 1C). The longest dimension of this molecule is nearly 150 Å (Cα distance). Each monomer contains a series of helices (α1–α11). One asymmetric unit contains two VicK dimers. Remarkably, the total buried surface area is 7590.8 Å2 upon VicK dimerization. The dimer interface consists of three tight hydrophobic contact patches (Figure 1D, I–III). In addition, the HAMP, PAS, and DHp domains are organized as dimers, which are connected by long, straight coiled-coils of helices α2 and α6 (Figure 1C). The C-terminal ends of the VicK dimer harbor two monomeric CA domains (Figure 1C). The N-terminal end of monomer A (Na, aa 31–37) and both C-terminal tails (Ca or Cb, aa 433–450) are disordered.
The HAMP domain (aa 36–86) is located at the uppermost position within the N-terminal region of the VicK structure (Figure 1C). Helices α1 and α2 of each VicK monomer form a parallel four-helical coiled-coil that is connected with loops L1 (Figure S1A). S. mutans HAMP shares approximately 45% identity to other Streptococci; however, it shows little (∼5%) identity to A. fulgidus Af1503, although the critical residues at positions a and d are mostly conserved (Figure 2A).
The helical interactions within the HAMP domain can be grouped into three shells (Figure S1B). The outer shell is formed by hydrophilic and polar residues at positions b, e, and g of the coiled-coil in addition to two residues from loop L1 (Figure 2A). The residues at the b and g positions in helix α1 are rich in basic residues with long side chains, whereas the corresponding positions in α2 are rich in polar residues (Figure S1B, gold). The middle shell is formed with hydrophobic residues, including leucine, isoleucine, and valine (Figure S1B, green). These residues form the canonical knobs-into-holes packing of coiled-coils [37]. Central to the VicK HAMP bundle are the hydrophobic residues that are all in van der Waals contacts and display knobs-to-knobs or x-layer packing (Figures 2A and S1B, blue).
The HAMP outer shell is distinct with three bound rings visible on the electrostatic surface (Figure S1C). Within this bundled structure, there are three pairs of hydrophobic residues in the knobs-to-knobs packing solely from two α2 helices. Two Leu71 residues compose the inside core of the first ring (Figure 2B) and two Leu78 residues are inside the core of the second ring (Figure 2C). Strikingly, the Phe82 pair forms π-π stacking inside the third ring, which is further stabilized by two Tyr56 residues from two L1 loops (Figure 2D). Indeed, Tyr56 is completely conserved in Streptococci, whereas Phe82 can be replaced by leucine in Af1503 (Figure 2A). In other HAMPs, position 82 can be Ile, Val, or Leu, whereas position 56 is typically Leu, Phe, or Tyr [15]. These residues are also capable of making van der Waals contacts if placed into the VicK HAMP domain (unpublished data). Notably, Gly54 is absolutely conserved in these HAMP homologs (Figure 2A).
The PAS domain (aa 87–198) in VicK, which is located downstream of the HAMP domain, adopts a canonical fold (Figure 3A). Both PAS domains can be well aligned (root mean square deviation [rmsd] of 1.2 Å) except for the large shift of loop L5 (Figure S2A). The five β-strands form a core of β-sheets, which are sandwiched by two α2 helices on one side and helices α3–5 on the other. The loop L3 and the connected helix α5 form a surface layer on the top of the β-sheet with two flexible loops, L4 and L5, on both edges. Three substantial binding pockets (S1–3) with a large cavity and inside tunnel are observed on this surface (Figure 3B). The tunnel is lined by mostly hydrophobic residues, including Leu108, Ile116, Leu127, Ile142, Phe178, and Leu195.
The VicK PAS domains form a unique dimer, which is mediated by a leucine-zipper (Figure 3C). This leucine-zipper further forms hydrophobic networks with the canonical PAS domain, which suggests that the leucine-zipper is an integral part of the VicK PAS domains (Figure S2B). The residues Leu89, Leu96, Leu100, and Lys93 are in key positions to make van der Waals contacts (Figure 3C). The substitution of Leu100 with arginine was previously shown to disrupt VicK autokinase activity in S. pneumoniae [38]. It is likely that the L100R mutation destabilizes the dimerization of the VicK PAS domains, which sheds light on the functional importance of PAS dimerization.
To look for clues of ligands for VicK PAS domains, we compared the VicK PAS domains with representative structures of all ligand-bound PAS domains according to a recent comprehensive review by Henry and Crosson [11]. Despite their limited sequence identities (6%–10%), all PAS domains can be structurally aligned to VicK with low rmsd of 2–2.5 Å (Figures S3 and S4). The relatively large shifts were observed in loops L3 to L5 and helix α5 of the PAS domains of NifL, FixL, DcuS, and PhoQ, which form pockets for a variety of ligands to bind. Therefore, the VicK PAS domain may bind some ligands differently from these PAS domains because it has a unique cavity and tunnel properties.
The C-terminal end of VicK (aa 199–450) contains a histidine-specific ATPase, which is often divided into one DHp domain and one CA domain with a short linker (aa 270–278) (Figure 1A). The dimeric architecture of these domains looks like a butterfly, wherein the DHp domain (aa 199–269) is a four-helix bundle of helices α6 and α7 with a phosphoryl receptor histidine located in the middle (Figure S5A). Surrounding this helical core are two CA domains with a layer of four helices (α8 to α11) on a layer of β-sheets (β7 to β12), wherein α10 is a short helix (Figure S5B). The two loop regions, L7 (aa 303–308) and L11 (aa 391–395) in the CA domain, called the lid, could not be well defined because of a weak electron density and are labeled with light dashed lines. One linker (aa 270–274), which connects the DHp and CA domains, was also disordered (Figure S5A, the monomer in magenta).
Both CA domains (aa 278–450) of VicK adopt a classical histidine kinase fold and are well aligned with an rmsd of 1.3 Å despite the missing loop L11 (Figure S5B). The CA domain of VicK is similar to HK853 (rmsd of 1.7 Å) except for structural differences in loops L7 and L8 in addition to the loop L11 region (Figure S5C). We could not clearly define ATP binding because of the weak electron density at the current resolution.
Unlike the HAMP domain, the DHp domain forms an anti-parallel four-helical coiled-coil. Interestingly, the two monomers of the DHp domain bear an asymmetric fold (Figure 4A). The two monomers are well aligned with an rmsd of 2.2 Å for the bottom part of the coiled-coil (aa 219–255), and the relatively large rmsd of the alignment is mostly caused by flexibility of loop L6. However, the upper parts of helices α6 and α7 are remarkably different, wherein helix α6 bends ∼25° at Pro222 toward the central DHp axis and α7 moves ∼11° away in the opposite direction to avoid a possible clash.
When aligning the VicK DHp domain with T. maritima HK853, which is a symmetric four-helix coiled-coil, we observed similar bending of monomer A (Figure 4B). The bending is coordinated with shifts of the upper helices, whereas the bottom helices remain stationary, and the upper part of helix α6 of monomer A moves toward the center axis, which possibly drives the other helices away.
Helix α6 of the DHp domain has an intrinsic plasticity for bending (Figure 4C). The bending region is absolutely conserved among over 50 VicK homologs in Streptococci, Lactobacilli, Lactococcus, and Enterococci (Figures 4C, the bottom and S6). Consistently, the DHp domain region (aa 211–225) was predicted to have low helical probability (Figure 4C, the top). Therefore, we hypothesized that the low helical propensity of the DHp domain may play a role in His217 phosphorylation. To test this, we mutated Pro222 to glycine and measured its autokinase activity (Table S1). The P222G mutant retained full activity when compared to wild-type (wt) VicK, as did the T221A mutant. We continued to mutate residues in the bending region, including Val212, Val215, Ser213, and Ser216 to alanine (VS2AA), which is statistically favorable in an α-helix [39]. All mutants, including the combined mutants VSP2AAA, VST2AAA, and VSTP2AAAA, bound S. mutans VicR similarly to wt VicK, which suggested that they likely retain the correct conformation (Figure S7).
The autokinase activity of the single and combined mutants described above was analyzed using γ32P-ATP (Figure 4D; Table S1). As the Thr221 and Pro222 mutants (T221A, P222A, P222G, and TP2AA) retained nearly full activity compared to wt VicK, the two combined mutations of VSP2AAA and VSTP2AAAA showed significantly reduced activity. Although we cannot rule out other effects from multiple sites of mutation, these data are consistent with the model that the low helical propensity of the DHp domain in addition to Pro222 are important for VicK autokinase activity.
As VicK is a multifunctional enzyme, we tested whether the helical bending region of the DHp domain is important for phosphatase activity. Here, we used Phos-tag gel mobility shift assay (PMS) to detect the phosphorylated S. pneumoniae VicR. As a control, over 90% of VicR was phosphorylated by acetyl-phosphate (AcP), which resulted in a mobility retardation shift (Figure 5A, lane 2). Further incubation with wt VicK completely removed the phosphate group from VicR (lane 3, top gel). It is interesting to note that in the absence of 5 mM ATP, little dephosphorylation of VicR was observed (lane 3, middle gel), which suggested that ATP is required for VicK phosphatase activity.
We tested the phosphatase activity of the DHp mutants described above. To our surprise, the single proline or threonine mutations (P222A and T221A) abolished VicK phosphatase activity (Figure 5A, lanes 5 and 7, top gel). In contrast, the P222G and VS2AA mutants, similar to wt VicK, dephosphorylated VicR within the time course of this assay (lanes 6 and 9). Consistently, the combined mutants VSP2AAA, VST2AAA, and VSTP2AAAA, which contain Pro222 and Thr221 substitutions, also showed little phosphatase activity (lanes 10–12). It is interesting to note that the VSP2AAA mutant had substantially more phosphatase activity than the single P222A mutant likely because VSP2AAA mutant lost partially kinase activity (Figure 4D).
We also analyzed some of these mutants using high performance liquid chromatography (HPLC). Phosphorylated S. pneumoniae VicR eluted at ∼7.6 min compared with the native VicR at ∼8.8 min (Figure 5B, run 1). After incubation with VicK in the presence of ATP, phosphorylated VicR completely shifted back to 8.8 min (Figure 5B, run 2). A clear conversion to unphosphorylated VicR was also observed when incubated with the VS2AA mutant (Figure 5B, run 3). All other VicK mutants, including T221A, P222A, TP2AA, and VST2AAA, failed to dephosphorylate VicR (Figure 5B, runs 4–7).
Together, our data suggest that Pro222 and its neighbor, Thr221, are the key residues in the bending region required for the phosphatase activity of VicK.
Although the overall conformation of each CA domain in VicK is the same, their positions relative to the DHp domain show dramatic differences, which generates an asymmetrical ATPase dimer (Figure 6A). The CA domain rotates ∼61° and further translates ∼20 Å down along an axis parallel to the DHp domain. When aligning the CA domain with the symmetric dimeric structure of T. maritima HK853, we found that monomer A takes completely different positions (Figure S8, magenta).
The large shift of the CA domain of monomer A moves its active site to the in cis phosphoryl acceptor His217. Recently, a crystal structure of the B. subtilis YycG CA domain bound to ATP was solved, which is 47% identical to the CA domain of S. mutans VicK [40]. These two structures of the CA domains aligned well with an rmsd of 1.47 Å of all backbone atoms, which is where the ATP from the structure of the YycG CA domain was modeled nicely in the active pocket of the VicK CA domain. To our surprise, we found that the γ-phosphate of ATP approached His217 to form two hydrogen bonds with εΝ or δN of His217 (Figure 6B, yellow dashed lines). In addition, Leu265 from monomer B is close to His217 to form van der Waals interactions (Figure 6B, grey dashed lines). Arg220 of monomer A and Arg269 of monomer B are also in close contact with His217.
In addition to the ATP binding pocket, the CA domain positions itself toward the DHp domain to form a large interface (Figure S9). Many residues are involved in the direct interaction between the DHp and CA domains (Figure 6C). Arg294, Asp326, Gln330, Asn334/337, and Arg385 of the CA domain form hydrogen bonds with Glu218, Thr221, Glu253, Arg256, and Arg259 residues around the middle region of the DHp domain. In addition, Arg382 and Asp387 form hydrogen bonds with Arg206 and Glu207 at the upper part of the DHp domain. Phe295, Ile298, Phe383, and Ile403 of the CA domain also form van der Waals contacts with Asn214, Thr224, Ser225, and Tyr229 of the DHp domain. Ile403, which is located on the back of the ATP binding pocket of VicK, is one of the key residues that contribute to the hydrophobic interaction with Leu267 and the aliphatic side chain of Glu218.
To test whether these interactions are important for autokinase activity, we generated a series of mutants and analyzed their autokinase activity (Figure 6D). The Arg294, Gln297, Ile298, Asn334, and Asn337 mutations (R294A, Q297A/I298A, and N334A/N337A) were nearly as active as wt VicK (Figure 6D, lanes 5, 7, and 9). In contrast, mutations of Asp326, Gln330, Arg382, Arg385, and Phe383 (D326A/Q330A, R382A/R385A, and F383A/R385A) dramatically suppressed VicK autokinase activity (Figure 6D, lanes 8, 10, and 12), whereas the dual mutations of D326A/N337A and R382A/R385A completely eliminated VicK autokinase activity (Figure 6D, lane 13). As references, two mutations in the ATP-binding pocket (K341A/Y342A) and a deletion of the first G loop (del392–395, RAQG) negatively affected autokinase activity (Figure 6D, lanes 11 and 14). Two Ile403 mutations I403W and I403S did not significantly affect VicK autokinase activity (Figure 6D, lanes 3 and 4). Interestingly, the I403W mutation was found to dramatically increase autokinase activity in HK853 (I448W) [23].
Together, the CA domain of monomer A can position itself toward the DHp domain to form a large interface, which is further centered by two groups of key residues. One group is D326/Q330, which forms hydrogen bonds with Arg259 of the DHp domain. Another group is R382 and R385, which form hydrogen bonds with Glu207 and Glu218 of the DHp domain. F383 inserts into a hydrophobic pocket in the DHp domain and further stabilizes the interactions between the CA and DHp domains. Therefore, a mutant (D326A/Q330A/R382A/R385A) that disrupts both groups of key residues completely eliminated the autokinase activity (Figure 6D, lane 13).
The CA domain of VicK monomer B stays away from the DHp domain and represents an inactive state (Figures 1C and S5A). A relatively small interface is formed mainly by van der Waals interactions between Phe383, Ile403, and Leu399 from the CA domain and Phe211, Val215, and Leu264 from the DHp domain, which are further stabilized by several hydrogen bonds between Arg382, Arg385, Asp271, and Glu218 (Figure S10). The buried surface of this interface is 623 Å2, which suggests that it may not be sufficiently stable. Phe383 and Leu399 appear to be the most important residues because they insert into helices α6 and α7 for extensive van der Waals interactions with Leu264 and Val215. When compared with T. maritima HK853 alone, the VicK CA domain rotates ∼76° along an axis vertical to the DHp domain (Figure S8). Overall, the CA domain remains inactive by positioning itself away from the DHp domain and this interface may act as a hinge.
The ability to mount a rapid response to various stress signals is essential for the adaptation and survival of prokaryotic cells. Some of the TCSs convert these stress signals ultimately to transcriptional reprogramming for necessary adaptive responses. How SKs are activated in response to these stimuli to initiate subsequent signal transduction cascades has been a long-standing question. In this study, we report the crystal structure of the entire intracellular portion of VicK (aa 31–450), which comprises an SK dimer with one signal transducer, one signal sensor, and one intact kinase domain. This structure allows us to dissect the molecular mechanisms underlying SK-mediated signal transduction in prokaryotes.
Our crystal structure revealed a long rod-shaped VicK holoenzyme (Figure 1). The three modules of the HAMP, PAS, and DHp/CA domains are connected through two groups of straight helices. Such an extended VicK molecule may provide sufficient surface and accessibility for potential ligands or protein partners to interact with.
The long shaped molecules often have the largest radius of gyration and tend to polymerize. VicK and its homologs have one or two integrated TMs rather than being free cytosolic proteins and are localized as clusters within the membrane [26]. In B. subtilis, YycG proteins center in the division ring possibly through interactions with FtsZ, a tubulin-like protein [41]. In contrast, approximately 420 dimeric VicK holoenzymes are randomly distributed as clusters throughout the periphery of one S. pneumoniae cell [42]. This clustering characteristic indicates that its function might require direct interaction between individual molecules, which is consistent with the physical properties of rod-shaped molecules.
A long-standing question is how HAMP domains serve in signal transduction because they often directly connect TMs and extracellular sensors. The HAMP domains may undergo a 26° rotation, which is derived from the unusual knobs-to-knobs packing of the solution structure of Af1503 HAMP, when they receive transmembrane signals [14],[15]. The crystal structure of concatenated HAMP domains from Pseudomonas aeruginosa Aer-2 implicates that the conformational dynamics may also serve a role in signal transduction [20]. Consistently, a series of structure and functional experiments have shown that intrinsic thermodynamic instability is required for HAMP signaling [16],[17],[19],[21],[43].
The VicK HAMP domain appears to be a stable four-helix coiled-coil with classical knobs-into-holes interactions, wherein three pairs of hydrophobic residues from helix α2 form a central hydrophobic core with knobs-to-knobs packing, and it is unlikely that structural alternations of the HAMP domain play a central role in signal transmission (Figures 2 and S1B). The VicK HAMP domain is further stabilized by Phe82, which is conserved in the majority of streptococcal species (Figures 2A and S11). However, variable residues, including isoleucine, valine, and occasionally threonine, are present at position 82 in some VicK homologs, which suggests that the stability of HAMP domains may vary in different species.
It is worth noting that the α2 helices of the HAMP domain connect with downstream PAS domains through continuous helices, part of which (aa 86–103) form a fairly rigid leucine-zipper (Figure 3C). This long helical structure most likely serves two purposes: (1) The HAMP domain might be further stabilized by additional interlock packing of the coiled-coil; and (2) The HAMP domain can easily transfer any conformational change down to the PAS domain. In contrast, the connection of the α6 helices between the PAS and DHp domains, particularly a short linker between α6 and β5, is rather flexible. Thus, the VicK molecule also appears to have a potential thermodynamic property for signal transduction from the HAMP or PAS domains to the catalytic CA domains.
The VicK PAS domains form a stable dimer through the short leucine-zipper (aa 89–103) (Figure 3C). In general, canonical PAS domains are rather flexible and readily subjected to ligand-induced conformational changes [9]. The three loops, L3–L5, form a mobile surface that could be regulated by unknown ligands (Figure 3B). Winkler et al. recently demonstrated that deletion of the PAS domain but not a triple mutation (D133N, N136Y, and L140R) of S. pneumoniae VicK reduced the autokinase activity and, more dramatically, phosphatase activity [33]. A similar deletion experiment also showed that the PAS domain is essential for the phosphatase activity of T. maritima ThkA [44]. Our S. mutans VicK structure shows that Leu135 (equivalent to N136 of S. pneumoniae VicK) stabilizes helix α4 by interactions with Ile116 and Ile139 (equivalent to L140 of S. pneumoniae VicK), which contributes to the hydrophobic cavity of the PAS domain (unpublished data). Therefore, it is likely that an Ile139 to arginine mutation only disrupted potential ligand binding but not dimerization. In contrast, an L100R mutation disrupted dimerization of the PAS domain, and in turn, the kinase activity of S. pneumoniae VicK [38].
In ligand-free VicK, helix α5 adapts an open conformation when compared to flavin adenine dinucleotide (FAD)-bound NifL, heme-bound FixL, and malate-bound DcuS (Figure S4). Thus, loop L3 is able to provide a sufficient cavity and tunnel for potential ligands to bind. Interestingly, when aligned with PhoQ, helix α4 and loop L3 of the VicK PAS domain adapt significantly different conformations and the trajectory of helix α5 is similar. Unfortunately, despite our efforts, no ligand specific to the VicK PAS domain has been experimentally identified. Only when such ligands are identified will it be possible to analyze how induced conformational changes regulate VicK catalytic activities. Therefore, this remains an interesting area for future research.
The local region around the phosphoryl receptor histidine of DHp domains is highly conserved [45]. We found that this region has a low helical propensity and is subject to significant helical bending, which is consistent with its low helical propensity. The helical bending most likely helps place His217 in close proximity to the CA domain to allow hydrogen bonds to form with the γ-phosphate of ATP (Figures 4 and 6B). Similarly, the DHp domain of B. subtilis DesK bends 50–54° when His188 is phosphorylated [25]. The HK853 DHp domain also bends 20° when bound to its cognate RR468 [46]. Therefore, the helical bending appears to be an intrinsic property of the DHp domain and is relevant to its function. Indeed, we showed that the combined mutations, through introducing alanines (VSP2AAA and VSTP2AAAA) into the DHp domain, abolished VicK autokinase activity (Figure 4D).
Residues Pro222 and Thr221 are conserved among VicK homologs (Figure 4C). Interestingly, these residues are essential for phosphatase activity because mutations of either P222A or T221A abolished the phosphatase activity of VicK (Figure 5). Proline, which produces a kink and 18–35° bending that affects the thermodynamic stability of the α-helix [47], plays key roles in the transmembrane signaling of several proteins, including G-protein-coupled receptors and voltage-gated potassium channels [48]. In addition to proline, threonine and serine residues are able to bend a helix 3–4° larger than alanine, which is important for the channel gating of voltage-dependent connexin32 [49],[50]. Our data in this study demonstrate that Pro222 and Thr221 are essential for phosphatase activity in VicK (Figure 4D).
Glycine may also serve a similar role as proline. The glycine in the middle of an α-helix attributes unique flexibility [51]. B. subtilis DesK has a glycine instead of proline present in this region [25]. Consistently, our P222G, as opposed to the P222A mutant, had phosphatase activity similar to wt VicK (Figure 5A, lane 6). It is worth noting that the DHp domain is not significantly bent in the structure of B. subtilis Spo0B and Spo0F complex compared with T. maritima HK853 and RR468 complex, although a glycine residue is localized adjacent to the phosphoryl receptor histidine [52]. However, it is possible that the crystal structure of the Spo0B and Spo0F complex captures only one state of the DHp domain of Spo0B kinase.
Our structure has shown that monomer A positions toward to its own His217 to form an active state, which is consistent with the findings using heterodimeric kinase mutants of HK853 and PhoR [46]. A flexible linker between the CA and DHp domains has been postulated to play a role in CA domain swinging [23]. In addition, the small interface created by Arg382, Phe383, and Arg385 provides a docking site as well as sufficient freedom for the CA domain to rotate (Figure S10). In the HK853 and RR468 complex, the CA domain swings ∼37° along the axis of the DHp domain when compared with free HK853 [46]. Crystal structures of B. subtilis DesK have shown that the CA domains could position themselves differently relative to the DHp domain [25].
The buried surface of the active monomer A is 1140 Å2. However, this interface is mainly composed of hydrophilic and polar residues, which suggests that these contacts may only be transient (Figure 6C). Our mutagenesis experiments showed that these residues are important for VicK autokinase activity (Figure 6D). Interestingly, while Ile403 mediates van der Waals contacts with Asn214 and Glu218 in the active state, it also contributes to the interface in the inactive state (Figures 6C and S10). Thr221 mediates a hydrogen bond with Asn337 and van der Waals interactions with Phe295 (Figure 6C). The T221A mutation eliminated phosphatase activity but did not affect the autokinase activity (Figures 4D and 5). In contrast, Winkler et al. found that a T221R mutant of S. pneumoniae VicK completely abolished its autokinase activity and greatly reduced phosphatase activity [33]. It is possible that the large arginine side chain may block the CA domain from properly accessing the DHp domain for active site formation.
It is important to note that the asymmetrical positioning of the CA domains and the different conformations of each monomer of the DHp domain are captured in a unique crystal-packing environment. The two VicK dimers form an anti-parallel tetrameric packing where the active CA position is stabilized by direct interactions with HAMP and PAS domains from another dimer in the same asymmetric unit (Figure S12). The N terminus of the HAMP domain makes contacts with the inactive CA domain of the VicK dimer from another asymmetric unit through the extended N terminus of monomer B (Nb) (Figure 1C). However, the overall structures of the HAMP and PAS domains remain symmetric.
Together, coordinated helical shifts of DHp and movement of the CA domains can be combined into a model to illustrate the activation steps for VicK (Figure 7). When both CA domains are inactive, they stay relaxed and further away from the phosphoryl acceptor histidine (I). Upon stimulation, one helix bends ∼25° toward the DHp central axis in coordination with the global shifts of DHp to expose the phosphoryl acceptor histidine. Consequently, one CA domain in cis rotates ∼61° to reach this histidine for initial phosphorylation (II). It is likely that this activation does not happen simultaneously for both CA domains because the DHp domain allows only one helix to bend at a time (Figure 4B). To fully activate both histidines (IV), VicK may go through an intermediate state that is similar to the inactive state (III) before the second CA domain rotates. Finally, the CA domains swing back to the initial state (IV to I) through several steps that remain to be determined.
S. mutans VicK (aa31–450) was cloned into vector pET15 (Novagen). S. mutans VicR (aa1–235) and S. pneumoniae VicR (aa1–234) were cloned into pETHis vector. All these constructs were expressed with N terminal 6× Histidine tag in Escherichia coli BL21/DE3 Rosetta (Novagen). VicK protein was purified through nickel affinity agarose (Qiagen), Q sepharose, phenyl sepharose, and Superdex S200 (GE Healthcare). The two VicR homologs were simply purified by nickel affinity agarose and gel filtration. Protein preps were concentrated down to 8 mg/ml in a buffer of 10 mM Tris (pH 8.0), 100 mM NaCl, 300 mM ammonium acetate, 2 mM DTT, and 1 mM β-mecaptoethanol (β-ME). The preliminary crystals were obtained using the screen kits of JCSG Plus (Qiagen) and Index (Hampton Research). Most crystals from these initial screening could only diffract to ∼4.5 Å. The later optimizations of those crystals defined the best condition of 2.3–2.8 M Sodium formate, 50 mM Tris (pH 7.4–8.6), and 4% PEG 4000 at room temperature. Selenium-methionine labeled protein was prepared following a standard protocol [53].
Bar-like crystals were grown for 2–4 d and immediately frozen in liquid nitrogen after quickly soaked with crystal growth buffer and additional 12% isopropanol as a cryo-protectant. Data were collected in BM17U, Shanghai Synchrotron Radiation facility (SSRF), China and processed by HKL2000 [54] and CCP4 suite [55]. Data collection statistics are summarized in Table S2.
The selenium sites were located using SHELXD [56]. Heavy atom positions were refined and phases were calculated with PHASER [57]. The real-space constraints were applied to the electron density map in DM [58]. An initial model was then built manually using COOT [59]. The model was further refined using PHENIX with stereochemistry and secondary structure information as restraints [60]. The structure and refinement statistics are summarized in Table S2.
Model analyses were performed using a variety of programs. The structural alignments were calculated in Coot [59]. Similar folds searches were carried out using Dali server [61]. The helical bending was calculated using a program HELANAL [62]. The buried surface areas were calculated in CNS [63]. The electrostatic potential surfaces was calculated and graphed by CCP4mg [64], while other graphics were made using Pymol (DeLano Scientific LLC).
The VicK protein prep at ∼4 mg/ml was first resolved on a size exclusion column (Shodex KW-802.5) in a buffer of 50 mM HEPES (pH 7.0) and 200 mM Na2SO4 at 25°C. Data were then collected on a DAWN HELEOS II laser photometer with an emission at 658 nm (Wyatt, USA). Molecular mass was calculated using ASTRA V (Wyatt, USA).
All VicK mutants were generated by our modified Quikchange mutagenesis protocol [65]. All these mutants were purified as wt VicK described above. β–ME was excluded from buffer in the final protein preps. The autokinase activity of VicK was measured by using isotope γ32P-ATP (Perkin-Elmer, NEG002Z001MC) according to a recent protocol published by Winkler and his colleague [33]. The working solution of the hot ATP was freshly made by mixing the hot ATP with an equal volume of 3 mM cold ATP. The VicK proteins at 0.5–10 µM were incubated with a concentration gradient (0–80 µM) of the hot ATP working solution in 10 µl reaction buffer of 50 mM Tris (pH 7.5), 200 mM KCl, and 10 mM MgCl2 for 0.25 to 8 min at room temperature. The measurements were setup at four different time points. The protein concentration and time points for each VicK protein were pre-determined to have the signals within linear ranges. The reactions were stopped by adding 3 µl 4× SDS loading buffer. The resulted mixtures were then subjected to 12% SDS-PAGE before the gels were dried and scanned in Typhoon 9410 (GE Healthcare). Quantifications were carried out using a program Totallab Quant. The autokinase parameters were derived from non-linear regression of the Michaelis-Menten equation. For simple comparison of the various VicK mutants, the VicK protein preps of 2 µg were mixed with 0.3 µl hot ATP working solution in 15 µl reaction buffer and further incubated for 30 min at room temperature. The reactions were stopped with 4× SDS loading buffer and separated in 15% SDS-PAGE. The gels were dried before being exposed to X-ray film.
S. pneumoniae VicR at concentration of 4 µM was phosphorylated in 50 mM AcP, 50 mM Tris (pH 7.4), 50 mM KCl, 2 mM MgCl2, and 20% glycerol for 1 h at 37°C [66],[67]. The phosphorylated VicR was then mixed with VicK wt and mutants at a final concentration of 4 µM for another hour at 37°C. AcP was diluted by at least 10-fold in the latter reaction. The resulted mixtures were first analyzed by PMS [68]. Briefly, the regular 8% SDS gels (29∶1) were prepared with additional 50 µM phos-tag acrylamide (Wako) and 100 µM MnCl2. The gels were run at 120 V for 120 min at 4°C for the best mobility shift and stained with coomassie blue.
The reactions of phosphorylated VicR were also analyzed by HPLC [67]. Briefly, additional 20% glycerol was added into reactions of VicR after treated with AcP as described above, which were further mixed with HPLC running buffers (A: 0.1% trifluoric acid; B: 0.1% trifluoric acid and 100% acetonitrile) to reach 40% acetonitrile. HPLC was run using a reverse-phase C8 column (4.6 mm×250 mm) (Agilent 1200). The phosphorylation state of VicR was confirmed by PMS as described above.
Homologs of the full-length VicK with >50% sequence identities were initially pooled from Genebank using a program BLAST [69]. Redundant entries of 96%–100% identity were identified using a multiple alignment program CLUSTAL [70] and subsequently removed, resulting in >50 unique homologs with 50%–96% identities (Figure S11). The DHp regions corresponding to amino acids 197–269 of these VicK homologs were used to generate the conservation logo using Weblogo server (weblogo.berkeley.edu). Redundant sequences (with 100% identity) of the DHp domain were further removed, resulting in >40 unique sequences (Figure S6). The helicity of the DHp was analyzed by a comprehensive secondary structure prediction program Phyre, which scores each amino acid as 0–9 for a helical probability [71]. DHp engineering was carried out through reiterative process of mutations and secondary structure calculation. Those unfavorable amino acids to α-helices within the bending region of DHp domain were determined according to statistical analyses [39].
The coordinates of the structure and relevant information have been deposited into the Protein Data Bank (4I5S).
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.